betaFeedback

WorksComputer Science

Glyph: An Interpretability Engine for AI-Driven Scientific Discovery in Clinical Data

Abstract

Automated scientific discovery has the potential to advance understanding of disease mechanisms and accelerate identification of early disease markers. Manual data annotation is infeasible at population scale, motivating methods that automatically extract pre-defined phenotypes in multimodal and unstructured clinical data. Such pre-defined feature sets determine and limit which associations these approaches can discover. This limitation extends to agentic AI systems that use these tools to query population-scale data. In contrast, foundation models (FMs) trained on large volumes of clinical data can perform tasks beyond their training objectives. This suggests their latent embeddings may represent novel or unexplored features that could yield insights into disease mechanisms, yet such latent features remain difficult to interpret. To address this challenge, we introduce Glyph, an automated multistage framework that extracts features from multimodal and unstructured clinical data using FMs and sparse autoencoders, interprets them as coded clinical concepts, and evaluates their value for clinical scientific discovery. We applied Glyph to abdominal CT volumes, CT radiology reports and laboratory measurements from 41,628 patients across different institutions. We identified 397 predictive concepts supported by 4,463 Bonferroni-significant associations. Moreover, a majority of discovered associations were automatically reproducible in held-out cohorts across different sites. Glyph concepts provided complementary associations to literature-derived markers, leading to gains in predictive power. We envision that interpretability can extend beyond a mechanistic understanding of neural networks to connecting implicitly learned features with clinical knowledge and establishing their significance for patient outcomes. We release Glyph at https://github.com/StanfordMIMI/Glyph.

Introduction

Introduction

Identifying early disease markers has repeatedly transformed clinical practice and patient outcomes[7]. Clinical discovery is inherently challenging as evidence of potentially novel markers is captured across diverse, often unstructured and high-dimensional modalities, including 2D and 3D imaging, laboratory results, and textual reports[6]. Moreover, the growing volume of clinical data, including approximately 300 million CT scans acquired worldwide each year[8], creates opportunities to identify new markers and uncover unknown associations with a broader range of clinical outcomes. Yet this growth also highlights a gap in our ability to process these data at scale, given the time and expertise required to manually define, annotate, and validate candidate disease markers.

Artificial intelligence (AI) is increasingly proposed as a means of accelerating scientific discovery[1]. In protein folding and materials science, targeted models have sharply reduced the time and cost of identifying promising candidates[9, 10, 11]. Attempts to automate clinical science have searched large databases of manually coded phenotypes through phenome-wide association studies (PheWAS)[2, 12]. Efforts to accelerate this approach use deep learning or agents to automate the annotation of phenotypes directly from images, but still restricts investigation to a preselected set of chosen markers[3, 4, 13]. As a result, clinical discovery still lacks an efficient framework for identifying open-ended, reproducible, and independent disease markers within multimodal data.

One potential solution lies in foundation model (FM) embeddings. Trained on large volumes of multimodal medical data, FMs produce embeddings that characterize patient states[14]. Such embeddings have been observed to perform well on tasks they were not explicitly trained for[14, 15, 16, 17], including disease prognosis and risk stratification, suggesting that they contain clinically meaningful associations beyond their original training objectives. However, these markers remain entangled within high-dimensional vectors, making them difficult to isolate and evaluate as candidate markers. Recently, mechanistic interpretability[18, 19, 20] has emerged as a self-supervised approach to uncover interpretable concepts encoded within FM embeddings, without the need for prospective annotation. Yet its broader potential for transforming FMs into engines of clinical discovery remains largely unexplored.

To address this gap, we introduce Glyph, a framework that turns foundation model (FM) embeddings into concepts for clinical discovery. The name derives from “glyph,” or a symbol that conveys meaning, reflecting our goal of giving hidden information a form that can be studied. Drawing on mechanistic interpretability, Glyph discovers coded concepts directly in FM embeddings and detects them across patients without requiring candidate markers to be defined and annotated in advance. These extracted concepts are measurable, interpretable patterns in patient data that together preserve the FM’s predictive capacity. Glyph integrates them with established clinical variables and tests their associations with disease-related outcomes across patient populations. This enables efficient, closed- and open-ended discovery that is infeasible with embeddings alone or with agentic systems operating on a limited set of pre-defined concepts. Our contributions are as follows:

  1. Establishing the concept space. To make FM embeddings useful for discovery, we develop the Glyph framework to first represent such embeddings in sets of sparse features, and then filter and interpret these as clinically meaningful concepts. Specifically, we evaluated features for significant prognostic associations, reproducibility across three sites, persistence after adjustment for measured confounders, and overall interpretability. Glyph identified 397 predictive concepts bearing 4,463 significant associations to 141 future disease outcomes meeting a Bonferroni threshold of p≤8.28×10−8p \le8.28 \times10^{-8}. Identifying concepts with at least one significant association yielded a space of high-risk concepts, capturing disease risk across modalities. Concept interpretations and prognostic associations also reproduced in external cohorts.

  1. Rediscovery and discovery. The validated concept space recovered established prognostic relationships across several disease axes, including hepatic, renal, cardiovascular, and oncologic disease. When tested in held-out cohorts, these associations reproduced across institutions. Glyph then extended these findings by linking established risk markers to additional outcomes and identifying reproducible prognostic associations that remain poorly characterized in the literature. Beyond disease-specific risk profiles, Glyph identified systemic risk markers across modalities, with one

imaging and three laboratory concepts retaining broad associations across disease groups after adjustment for potential confounders. Glyph revealed that nearly half of the high-risk imaging concepts were not consistently documented in the accompanying radiology reports.

  1. Comparison with literature We investigated whether and in which domains Glyph brought additional prognostic value over the clinical literature. To this end, a co-scientist agent processed 342,647 clinical papers covering each discovery axes, and reproduced 412 published markers in our data. We found using just ten Glyph concepts per axis led to significantly stronger models in external data than literature markers alone (3.34% AUC gain, p=2.66×10−28p = 2.66 \times10^{-28}), with predictive improvements of up to 11.7% in specific axes. Comparing the top ten Glyph concepts with the top ten literature markers per outcome, selected at the Discovery Site, Glyph associations reproduced more frequently in the pooled held-out cohort (64.8% vs. 41.2%, p=2.54×10−6p = 2.54 \times10^{-6}) at higher median confounder-adjusted odds ratios (1.31 vs. 1.16, p=5.0×10−5p = 5.0 \times10^{-5}).

  1. Comparison with pre-coded phenotypes Similarly, we also asked whether pre-coded phenotypes could support the same discoveries, comparing Glyph with an automated PheWAS-style analysis of ICD-derived Phecodes. Across the same outcomes, associations from Glyph reproduced more often than those from Phecodes at Internal Test (69.3% vs. 49.1%, p=9.77×10−4p = 9.77 \times10^{-4}) and External Site #1 (74.3% vs. 63.9%, p=2.93×10−3p = 2.93 \times10^{-3}), and had larger median effect sizes. Overall, Glyph showed a complementary benefit to existing disease coding protocols and clinical literature.

  1. Open-source and use. Glyph extracts additional value from existing pretrained FMs, without retraining. To enable researchers to find our concept space in their cohorts, or build new concept spaces from their data, we open-source Glyph at github.com/StanfordMIMI/Glyph, installable via pip install glyph-discovery. A validated concept space also supports other applications including explainable patient-level predictions and retrieval. We also release the Glyph MCP, enabling clinicians, engineers, and AI agents to investigate patient data and test open-ended hypotheses and applications through concepts.

Results

Glyph was deployed in a multimodal Discovery Site cohort of 14,652 patients, comprising 20,391 3D abdominal CT volumes (3,262,560 preprocessed slices), radiology reports (8,083,882 tokens), and 1,588,003 lab entries spanning 350 types (Figure 1a). These three modalities were embedded using the Merlin image encoder[14], Qwen3-Embedding-8B[11], and Labs FM, respectively (Figure 1b). Glyph then decomposes these embeddings into sets of sparse features (Figure 1c). We identified when Glyph concepts rediscovered established prognostic associations, and when they reveal complementary associations absent from the literature or not discoverable by automated PheWAS-style discovery analysis (Figure 1d).

Figure 1. Figure 1. Transforming foundation model embeddings into engines of clinical discovery. a Glyph can be applied to any data expected to contain useful markers for discovery. b Foundation models compress multimodal clinical data into high-dimensional embeddings, typically used for predefined prediction tasks. The structural contents of embeddings driving these predictions is typically not unexplored. c Glyph instead transforms FM embeddings into engines for discovery, mapping a space of concepts hidden within these embeddings. Glyph goes beyond embeddings by decomposing them into sparse concepts that can be studied individually and in selected combinations (Figure 1c). d Glyph concepts recover associations established in clinical research and ICD-based PheWAS analyses, while revealing complementary associations not yet discovered by either approach. e–f Glyph evaluates concepts for e associations with future disease after correction for multiple testing, f associations that persist after adjustment for measured confounders, and g clinical interpretability. h All findings are tested for reproducibility across multiple cohorts. i Validated concepts provide an interface for asking complex questions of FM embeddings and grounding the answers in the underlying data.

To enable this analysis, we developed a rigorous pipeline to process sparse features in embeddings as concepts useful for discovery (Figure 1e–Figure 1g). To maximize discovery yield, we used the Discovery Site validation set to identify 141 disease outcomes whose onset 1–4 years after abdominal CT was most predictable from FM embeddings (see Methods, § 5.5). These outcomes spanned hepatic, renal, cardiovascular, metabolic, and neurological disease. To surface high-risk concepts, Glyph filters features to those with at least one significant prognostic association (Figure 1e), before assessing their independent risk adjusted for specific confounders (Figure 1f), and estimating each feature’s clinical interpretation (Figure 1g). At each step, we assessed the reproducibility of these analyses across three held-out cohorts comprising 26,976 patients and 45,139 CT studies, with 7,222,240 preprocessed CT slices, 14,342,345 radiology-report tokens, and 2,184,424 lab entries (Figure 1h). This pipeline led to the set of 397 high-risk concepts that can be used in downstream discovery workflows (Figure 1i).

Glyph discovered a clinical concept space in FM embeddings

To first identify a pool of candidate features, Glyph fits one SAE architecture per modality (Figure 2a). On the Discovery Site, SAEs identified 3,315 features in embeddings of reports, 343 in images, and 627 in labs, totaling 4,285 features. Once features are learned, they can be re-extracted elsewhere without

Figure 2. The Glyph framework. An interactive version is available at https://stanfordmimi.github.io/Glyph-research/. a At the Discovery Site, Glyph trains sparse autoencoders (SAEs) to decompose foundation model (FM) embeddings into sparse concepts. b Each SAE encoder is then frozen as a concept extractor and its decoder discarded. Concepts are re-extracted zero-shot in held-out internal and external cohorts to assess reproducibility of their interpretations, prognostic associations, and confounder analyses. c The discovered concept space maps the clinical characteristics and disease burden of patients undergoing abdominal CT at the Discovery Site, alongside routine variables such as age and smoking status. We show 397 concepts (9.26% of the 4,285 candidate features) that were associated with at least one of the 141 prognostic outcomes after Bonferroni correction. Edges show correlations between concepts in the top 2.5%. Clinically related concepts were found to cluster into subregions (dashed boxes) corresponding to liver, renal, and cardiovascular disease, and systemic inflammatory and cholestatic illness. Node size scales with the logarithm of the number of patients who share the specific concept. A confounder-adjusted version of this network is shown in Figure 11.

recalibration (Figure 2b). Specifically, FMs were reapplied to extract embeddings of data in the Internal Test, External Site #1 and External Site #2 cohorts. The trained SAEs were then applied zero-shot to each site’s embeddings to re-extract the same set of 4,285 features. Overall, these features preserved the majority of the original embeddings’ variance and predictive performance (see Extended Figure 7 and § 5.3). These features were then filtered by Glyph’s discovery pipeline, explored in detail in § 3.2, § 3.3 and Section A.1, yielding the set of 397 validated, high-risk concepts. Textual interpretations of these concepts were then attempted via a contrastive approach detailed in Section A.1 and illustrated in Figure 12. Throughout, we refer to concepts by a modality-prefixed index: R for radiology report, I for imaging, and L for laboratory concepts.

Figure 12. Glyph concepts are automatically extracted, interpreted, and reproduced zero-shot at external sites. Each concept is interpreted on the Discovery Site by comparing top- versus non-activating examples, then re-extracted at an Internal Test cohort and two external sites where the same interpretation is re-tested on unseen patients; consistent discrimination shows both concepts and interpretations transfer across institutions. a–c A radiology report concept, labeled metastatic disease, re-extracted zero-shot and confirmed at every site. AutoInterp accuracy is highest for reports: as human-written summaries, their concepts are near-fully interpretable but bounded in discovery potential. d–e An imaging concept whose interpretation is not consistently documented in radiology reports, yet was found to rank among the top 10% of imaging predictors of infection, sepsis, wasting, and organ failure. f Remarkably, 52.5% of imaging concepts were interpretable via corresponding radiology reports. g–i A laboratory concept, labelled monocytosis or neutrophilia, confirmed across sites.

High-risk concepts clustered into neighborhoods reflecting liver, kidney, cardiovascular, musculoskeletal, and malignant disease, as well as inflammatory states and abnormal cholesterol (Figure 2c). Within the liver and kidney neighborhoods, granular concepts from different modalities overlapped. Concepts for fluid accumulation, including effusions and ascites, lay nearby and were closely correlated with laboratory concepts reflecting hypoalbuminemia, reflecting known comorbidities to liver cirrhosis[22]. Laboratory concepts also provided information complementary to imaging and reports, such as evidence of abnormal cholesterol, situated adjacent to kidney disease[23], and inflammatory markers. Other regions contained high-risk imaging concepts with no corresponding report concepts, suggesting radiological features that are not routinely documented. The concept space also isolated the presence of medical devices, including tubes and drains, and postsurgical artifacts, that could confound predictions based on whole embeddings. It also integrated established variables not derived from SAEs, such as patient age and smoking status, alongside the extracted concepts, situating these adjacent to cardiovascular and musculoskeletal disease, which are also known correlates[24]. These 397 high-risk concepts are explored for their discovery potential in the remainder of the paper.

Glyph identified reproducible markers of future disease risk

Having constructed the concept space, we tested whether it contained any significant, reproducible high-risk markers for future disease onset. Because the cohort consisted of patients who underwent abdominal CT, associations were expected to be enriched for abdominal pathology. ICD codes provided broad, cross-institutional outcome labels but may be affected by undercoding, miscoding, delayed entry, and institutional differences in coding practices. Associations may also reflect clinical workflows, motivating the confounder analysis in § 3.3.

To identify high-risk concepts, we first computed per-standard-deviation odds ratios (ORs) between all pairings of 4,285 features and 141 outcomes, yielding 604,185 candidate associations at the Discovery Site. These comprised 467,415 radiology report, 48,363 imaging, and 88,407 laboratory associations (Figure 3a; see Methods, § 5.6). To set a stringent threshold for evidence, we applied the Discovery filter, requiring a Bonferroni-corrected p≤8.28×10−8p \leq8.28 \times10^{-8} and OR > 1.0 (Figure 3b). This analysis uncovered 4,463 significant associations (0.739% of all pairs), establishing that the concept space contains signals of future disease across radiology reports (858; 0.184%), imaging (2,251; 4.65%), and laboratory measurements (1,354; 1.53%). This narrowed the space to the 397 high-risk concepts with at least one significant association used in the remainder of the analyses (as shown in Figure 2b).

Figure 3. Glyph concept-wide association study. a Each Glyph concept is tested for association to each of 141 future disease outcomes within 1 to 4 years, yielding 604,185 concept-outcome pairs evaluated using per-standard-deviation odds ratios. b A Discovery filter (p≤0.05p \le0.05 adjusted for multiple testing via Bonferroni correction, equivalent to p≤8.28×10−8p \le8.28 \times10^{-8}, and OR > 1.0) applied on the Discovery Site yields 4,463 (0.739%) promising associations. c Number of promising associations per modality on the Discovery Site, totalling 4,463; these are fixed and tested for reproducibility without calibration on the Internal Test, External Site #1, and External Site #2 cohorts using a 5% FDR threshold (Benjamini–Hochberg), with the percentage replicating shown per modality and cohort. d Volcano plot of all concept-outcome associations and the Discovery filter selection. e Risk profiles and reproducibility statistics for one organ-specific and one systemic-risk concept. f–h Sample counts and observed probabilities of future disease onset, with confidence intervals, across concept-activation bins.

Having identified thousands of significant associations, we then assessed whether any were reproducible in the internal and external sites. This was tested using a less conservative Reproducibility filter, affording a 5% false discovery rate via the Benjamini–Hochberg procedure (see Methods, § 5.7). Of the 4,463 promising associations, 3,095 (69.3%) were reproduced without calibration on the Internal Test cohort of 3,667 patients (see Figure 3c). Remarkably, a greater number, 3,314 (74.3%), were reproduced at the secondary site, External Site #1, reflecting its larger cohort of 17,224 patients and indicating that associations generalized across institutions. Moreover, External Site #2 which was out of state, reproduced 1,701 (54.7%) of the radiology-report and imaging associations, reinforcing the generalizability of these findings across institutions. Overall, this showed that Glyph automatically surfaced concepts that were reproducibly predictive of future disease across patient populations.

This prompted us to investigate the nature of these associations, and whether they reflected insights in the literature. Less than a third of abdominal CT patients at External Site #2 had a single recorded lab

result, so associations from labs concepts on this cohort were not run. These rediscoveries recapitulated well established risk factors across hepatic, renal, and cardiovascular disease, and reveal several systemic risk factors in our data.

Rediscovery of hepatic disease progression: Among the strongest predictors of hepatic disease were L108 (cholestasis), I65 (splenomegaly), I21 (ascites) and R85 (liver cirrhosis and varices). R85 predicted future primary liver cancer with odds ratios and p-values of 2.45 p=1.2×10−39p = 1.2 \times10^{-39}, 2.55 p=1.2×10−7p = 1.2 \times10^{-7}, 1.73 p=1.3×10−9p = 1.3 \times10^{-9}, 1.64 p=5.7×10−9p = 5.7 \times10^{-9} at the Discovery Site, Internal Test, External Site #1 and External Site #2, respectively (Figure 4f). It was also associated with portal hypertension (OR 2.19–3.57), with both associations replicating across all four cohorts. These findings accord with established complications of cirrhosis[25, 26]. R85 was associated with the onset of 39 of 141 outcomes (Figure 3e), tracing the recognised natural history of cirrhosis through jaundice, hepatic necrosis, varices, splenomegaly and coagulopathy. Associations also extended to renal, respiratory and infection diagnoses, consistent with reported evidence of these conditions in patients with cirrhosis[27, 28, 29].

Figure 4. Confounder-adjusted association study. a Glyph concepts act as variables that can be compared with one another and with established clinical variables in a unified feature space, enabling confounder analysis of FM predictions. b For each of the 4,463 promising associations, the ten most strongly associated concepts across other modalities are automatically selected as candidate confounders, alongside six universal clinical covariates. c Number of associations per modality retaining a significant, adjusted OR > 1.0 after confounder adjustment per site. d–g Association profiles show that after adjustment I21 (ascites) loses most associations, L127 (cytopenia) remains broadly associated, R85 (cirrhotic liver morphology and varices) retains hepatic associations, and I26 (aortic atherosclerosis) retains cardiovascular and neurocognitive associations. h An example of a knowledge-guided confounder analysis for primary liver cancer.

L108 also recovered the majority of hepatic outcomes. These included future primary liver cancer (OR 2.08–2.57) and portal hypertension (OR 1.63–2.92). Li and Apte linked bile-acid accumulation in cholestasis to liver injury and cancer[30], but did not evaluate any prognostic relationships. Reproducible associations extended to future liver transplantation and coagulation (see Figure 3g).

Rediscovery of renal disease progression: The strongest marker of renal disease was L84 (azotemia), which reproducibly traced the stages and complications of chronic kidney disease, including chronic renal failure (2.02–2.52), progression to stage IV disease (1.77–2.73), initiation of renal dialysis (2.04–3.21), eventual kidney transplantation (2.53–2.87), and the metabolic consequences of uremia in anaemia of chronic kidney disease (1.87–2.44) and hyperkalaemia (1.80–2.07)[31]. L84 also extended beyond renal disease to cardiovascular, metabolic, infectious, respiratory, hepatic and haematological outcomes. Elevated blood urea nitrogen (BUN), a measure of azotemia, predicted mortality after acute coronary syndromes in prior work, although that study did not examine later diagnoses across disease groups[32].

By comparison, L69 (impaired renal function, defined by increased creatinine and BUN and reduced eGFR) was far more specific, as its seven significant associations were all renal (against 11 for L84 among 76 spanning every disease group). Its effects were smaller on all four shared outcomes (1.09–1.54 versus 1.77–2.87), it did not predict dialysis initiation or acute renal failure. Notably, urinary protein loss was not isolated by any single concept, despite albuminuria being a second axis of chronic kidney disease staging alongside eGFR[31]; nephrotic-range values were distributed among concepts dominated by hypoalbuminaemia (L34, L45, L61, L87).

Unlike the hepatic rediscoveries, no radiological concept captured renal disease progression, consistent with the weak association between visible kidney structure and renal function[33]. Specifically, R398 (kidney atrophy) and R302 (polycystic kidney disease) reached only modest effect sizes (1.16–1.41) and replicated inconsistently. By contrast, I109 (aortic calcification) and L127 (cytopenia) were associated with 10 and 12 of 13 renal outcomes, respectively, with strong reproduction across cohorts. Prior studies established abdominal aortic calcification as a cardiovascular prognostic marker in CKD[34]; here it also marked later renal diagnoses more consistently than kidney-specific imaging findings.

Rediscovery of cardiovascular risk markers: Glyph recovered early cardiovascular markers of calcification and atherosclerosis. Specifically, R21 (calcification of the coronary arteries or cardiac valves) was a prominent report-derived cardiovascular predictor, reproducibly predicting myocardial infarction (OR 1.25–1.70), heart failure with preserved (1.58–1.83) and reduced (1.40–1.92) ejection fraction, and coronary atherosclerosis (1.52–1.97). Prospective studies have linked coronary and mitral annular calcification to subsequent cardiovascular events[35, 36]. Both R21 and R3 (moderate to severe atherosclerotic disease) replicated strongly at External Sites #1 and #2 (R21: 100% and 73.7%; R3: 91.7% and 86.1%), well above the 33.8% cohort average at External Site #2.

Imaging concepts outperformed report-based concepts: the imaging group, comprising I109 (abdominal aortic calcification), I15 (abdominopelvic atherosclerosis) and I26 (aortic atherosclerosis; Figure 3e), covered 23 of 28 cardiovascular outcomes (82.1%), against 19 of 28 (67.9%) for the report-based group, and replicated more consistently across the Internal Test and External Sites #1 and #2 (96.1%, 98.7% and 93.4% versus 89.6%, 97.9% and 70.8%). Imaging concepts also more strongly recovered atherosclerosis’s known systemic vascular effects, covering more outcomes and with larger effect sizes than report-based concepts across renal (11 of 13, 84.6% vs. 7 of 13, 53.8%) and neurocognitive outcomes (8 of 12, 66.7% vs. 6 of 12, 50.0%)[37], including vascular dementia[38]. Atherosclerosis and vascular calcification concepts (I26, I106, R21 and R420) also showed associations beyond cardiovascular disease, including with femoral and vertebral fractures[39] and chondrocalcinosis.

Most inflammatory concepts (L0, L3, L4, L8, L92) did not have substantial associations with cardiovascular outcomes, which was unexpected given the established role of inflammation in atherosclerosis[40]. L68, more indicative of infection (a blood-count pattern with elevated neutrophils, white cells, monocytes and platelets) than auto-immune inflammation, was associated with 5 of the 28 cardiovascular outcomes. It reproducibly predicted ischaemic heart disease (OR 1.17–1.33) and myocardial infarction (1.21–1.34). L121 captured delayed cardiac conduction on ECG and was associated with later arrhythmia and heart failure, consistent with prospective evidence linking prolonged QRS duration to incident heart failure[41]. L1488 similarly captured abnormal RR intervals, but neither concept’s adjusted associations reproduced consistently, and both may reflect ECG testing prompted by suspected cardiac disease. Concepts capturing abnormal cholesterol (L182, L184, L481, L852) did not emerge as major cardiovascular risk markers in this cohort, despite the established causal role of LDL cholesterol in atherosclerotic cardiovascular disease[42], potentially due to widespread use of statins in this cohort.

Concepts encoding systemic risk markers: Beyond organ-specific associations, several concepts behaved as broad markers of systemic risk, each associated with more than half of the outcomes tested and spanning nearly every disease group. L127 (cytopenia) and L108 (cholestasis) had the broadest risk profiles, associated with 89 (63.1%) and 87 of 141 outcomes (61.7%) respectively, each spanning 10 of 11 disease groups. L108 extended beyond liver disease to predict future infection, electrolyte disturbance, renal failure and cardiac failure. Cholemic nephropathy offers one possible explanation for these renal associations in patients with severe liver disease[43, 44]. L127 (cytopenia) likewise reached cardiovascular, infectious, renal and metabolic outcomes. L84 (azotemia) followed with 76 associations (53.9%) and spanned all 11 groups. This breadth generalized across institutions: L108 reproduced 82 of its 87 associations at the Internal Test cohort (94%) and 84 (97%) at External Site #1, L127 79 (89%) and 86 (97%), and L84 69 (91%) and 75 (99%), each well above the cohort-wide rates of 69.3% and 74.3%.

Fluid accumulation behaved similarly: I21 (ascites) was associated with 68 outcomes across 10 groups and carried the largest single effect of any systemic concept (OR up to 3.17), while I25 (combined ascites, anasarca and effusion) reached 62 outcomes across all 11 groups, and R101 (pleural effusion with ascites) 30. I28 (findings of chronic or end-stage disease) contributed an imaging-derived systemic marker, associated with 84 outcomes (59.6%) across 10 groups and reproducing 76 (90%) and 83 (99%) times. Notably, ascites reproduced least consistently of this set (48 of 68, 71%, and 46, 68%), an early indication that its breadth may rest on shared upstream disease rather than independent risk, which we examine in § 3.3. A second family encoded hypoalbuminaemia and anaemia (L34, L59, L77, L82, L87, L91, L124), each associated with 39 to 71 outcomes across 10 disease groups with median odds ratios of 1.42 to 1.62.

Infection and sepsis: L127 (cytopenia)[45] and L108 (cholestasis) were each associated with 14 of 22 subsequent infectious and inflammatory outcomes, although neither was specific to this outcome group. I76 (severely distended bowel pattern) was associated with 5 of 20, including bacterial pneumonia, sepsis and septicemia, although presence of this image phenotype was not consistently documented in the corresponding radiology reports. Visual review identified severely distended, fluid-filled bowel loops (Figure 12d–Figure 12e; see Section A.1). This recapitulates prior evidence linking bowel obstruction to bacterial translocation and postoperative septic complications[46].

Rediscovery of metastatic sequelae R64 (metastatic disease) predicted systemic complications of advanced malignancy, including cachexia and pancytopenia (Discovery ORs 1.63–1.85, both reproducing at Internal Test and External Site #1), capturing the wasting and marrow failure of disseminated disease[47], alongside neutropenia, bacteremia, and septic shock (ORs 1.45–1.58, same replication pattern), reflecting common infectious complications[48]. Overall, while 25 of R64’s 32 significant associations (78%) replicated at Internal Test and 26 (81%) at External Site #1, none reproduced at External Site #2, likely due to differences in coding practices.

Glyph enabled confounder analysis to identify independent disease markers

Having identified significant, reproducible disease markers, we assessed whether they persisted after adjustment by potential confounders. A concept encoding ascites may appear to predict liver cancer because underlying cirrhosis can cause both fluid accumulation and an increased risk of liver cancer (see Figure 4a). The majority of FM outputs are only tested for predictive correlation to an outcome, and rarely test if this prediction is potentially confounded. This extends the scientific reach of foundation models from correlation towards independence analysis. Glyph concepts, due to being both interpretable and sparse, can be tested for independence alongside known clinical covariates and risk markers in confounder analysis frameworks. As in the association study, these results identify associations that persist after adjustment for measured clinical variables and other concepts, rather than establish causality.

We adjusted each significant association for confounders by applying double machine learning (DML) (see Figure 4b and Methods, § 5.8). 16 confounders were drawn per-outcome, including the top 10 significantly associated concepts, and six universal clinical covariates covering age, sex, race, BMI, alcohol use, and smoking. Applying the Confounder filter to the 4,463 promising associations on the Discovery Site, 1,198 (26.8%) retained an adjusted OR > 1.0 (Figure 4c). Fewer associations were retained at the Internal Test Site and External Sites #1 and #2, potentially reflecting lower statistical power, variation in FM reproducibility, and differences in population distributions. At External Site #2, 643 of 3,109 associations (20.7%) survived the Confounder filter. In the following, we study which significant associations withstood adjustment, while also noting groups of concepts that were most heavily attenuated. Unless otherwise specified, results refer to the Discovery Site.

Independence of hepatic disease risk markers: R85 (cirrhotic liver morphology and varices) was associated with 9 of 11 hepatic outcomes at the Discovery Site, and 6 of these 9 remained significant after confounder adjustment (see Figure 4d). Among these were known consequences of cirrhosis, including portal hypertension (adjusted OR 2.46 p=3.0×10−4p = 3.0 \times10^{-4}, 2.35 p=1.4×10−1p = 1.4 \times10^{-1}, 3.10 p=1.2×10−3p = 1.2 \times10^{-3}, 1.77 p=2.3×10−9p = 2.3 \times10^{-9}), liver abscesses (2.98 p=1.5×10−16p = 1.5 \times10^{-16}, 3.15 p=2.7×10−8p = 2.7 \times10^{-8}, 3.19 p=1.1×10−7p = 1.1 \times10^{-7}, 1.87 p=1.4×10−11p = 1.4 \times10^{-11}), primary liver cancer (4.07 p=1.2×10−71p = 1.2 \times10^{-71}, 2.13 p=4.1×10−7p = 4.1 \times10^{-7}, 1.79 p=1.9×10−8p = 1.9 \times10^{-8}, 1.68 p=2.4×10−7p = 2.4 \times10^{-7}) and liver transplantation (2.11 p=1.6×10−8p = 1.6 \times10^{-8}, 2.06 p=1.0×10−7p = 1.0 \times10^{-7}, 2.15 p=1.0×10−10p = 1.0 \times10^{-10}, 0.99 p=9.8×10−1p = 9.8 \times10^{-1}), with reproduction across all three held-out cohorts in the majority of cases. By contrast, I21 (ascites) and I25 (ascites, anasarca or pleural effusion) retained only one of 10 and none of eight Discovery Site associations, respectively, after adjustment that included R85. This pattern is consistent with fluid accumulation being less specific to hepatic disease than cirrhotic morphology and varices[26].

Independence of renal disease risk markers: As in hepatic disease, confounder adjustment concentrated the renal signal on known markers of renal disease. L84 (azotemia) retained 8 of its 11 Discovery Site renal associations, covering chronic renal failure, stage IV disease, kidney transplantation, oliguria and anuria, hypertensive chronic kidney disease, secondary hyperparathyroidism, acute renal failure and the composite chronic kidney disease outcome, with effects broadly unchanged by adjustment (ORs 1.71–2.53 vs. adjusted ORs 1.27–2.95). L69 (impaired renal function) retained 6 of its 7 associations, all renal, at consistently smaller effects (ORs 1.24–1.44 vs. adjusted ORs 1.45–1.85), remaining highly concentrated on the kidney. The broader I109 (aortic calcification) and L127 (cytopenia) markers retained 4 of 10 and 4 of 12 renal associations, respectively, indicating a mixed profile. At Discovery, R398 (kidney atrophy) retained its sole renal association with hypertensive chronic kidney disease after adjustment. R302 (polycystic kidney disease) retained associations with stage III chronic kidney disease and kidney transplantation, but not oliguria and anuria. In contrast, markers of fluid accumulation and general illness lost their renal links almost entirely, with I21 (ascites) falling from 7 significant renal associations to just 3.

Independence of cardiovascular risk markers: Similarly, R21 (calcification of the coronary arteries and cardiac valves) lost 32 of its 38 associations after adjustment, including those with hypertension and atrial flutter across all cohorts. I26 (aortic atherosclerosis) retained 11 of 21 cardiovascular and 4 of 7 neurocognitive associations after adjustment, while most associations with other outcome groups

Glyph finds novel associations beyond the clinical literature

were attenuated (Figure 4e). Some vascular–musculoskeletal associations attenuated after adjustment, while others persisted, including R21 (coronary/valvular calcification) with femoral fracture and chondrocalcinosis, I106 (atherosclerosis with colonic pathology) with femoral fracture, and R420 (coronary calcification/atherosclerosis) with chondrocalcinosis. We then tested R21 in the discovery cohort using a radiologist-specified set of cardiovascular risk factors: age, sex, smoking, alcohol use, BMI, diabetes, hypertension, chronic kidney disease, sleep apnea, and statin use extracted from medication orders. Moreover, we grouped atherosclerosis and calcification concepts within each modality. The imaging-derived group (5 concepts combined) remained associated with future dementia after adjustment (OR 2.28 vs. adjusted OR 1.41, p=5.2×10−6p = 5.2 \times10^{-6}), whereas the report-derived group (7 concepts) did not (1.61 vs. 0.93, p=0.44p = 0.44). This suggests that imaging captured information about future dementia risk beyond that recorded in radiology reports.

Independence of systemic risk markers: Many concepts previously identified as encoding systemic risk were revealed to be heavily confounded. The most heavily discounted associations came from fluid accumulation. Specifically, I21 (ascites) lost 54 of its 68 (79%) associations under adjustment, among the highest attrition of any concept and well above the cohort-wide average of 73% (see Figure 4f). Similarly, R101 (pleural effusion and ascites) lost 90% and I25 (combined ascites, anasarca, and effusion) lost 82%.

I87 (uncharacterized hepatobiliary pattern) had the largest number of associations both before and after adjustment among all 397 concepts, retaining 52 of its 96 associations (54%). These retained associations spanned all 11 outcome groups, including 8 of 8 respiratory, 9 of 18 cardiovascular and 7 of 11 hepatic associations. This mixed imaging pattern, encompassing gallbladder absence and cirrhotic or portal-hypertensive findings, therefore captured adjusted risk extending beyond hepatic disease. I65 (splenomegaly) retained 30 of its 48 associations (63%) after adjustment, spanning eight outcome groups. These included subsequent primary liver cancer, consistent with the association between incidentally detected splenomegaly and later liver cancer diagnosis reported in a recent population-based cohort study[49]. This association persisted across all four cohorts after adjustment for the clinical variables and concepts reflecting liver dysfunction, coagulopathy and ascites, but was attenuated when cirrhosis and other liver-related findings were included in a targeted analysis (as shown in Figure 4h).

Remarkably, systemic risk markers from labs concepts maintained a majority of their associations after adjustment. Specifically, L108 (cholestasis) maintained 49 of its 87 associations (56%), covering 9 of the 11 outcome groups. Abnormal liver tests have been linked to haemodynamic status in patients with established heart failure[50], but did not assess whether cholestasis preceded later cardiac diagnoses, as observed here. L127 (cytopenia) retained 45 of 89 associations (51%) after adjustment (see Figure 4g). A prospective population study linked lymphopenia, a subtype of cytopenia, to later infection, including sepsis[45]. L127 captures a broader cytopenia pattern associated with outcomes across disease groups. L84 (azotemia) retained 28 of 76 associations (37%), with associations attenuating in respiratory, cardiovascular and oncologic disease but persisting for renal, metabolic, and haematologic outcomes (see Figure 4g).

In contrast, concepts involving hypoalbuminaemia (L34, L59, L77, L82, L87, L91, L124) were heavily attenuated, retaining only 0–17.1% of their associations. Hypoalbuminaemia is an established independent predictor of mortality[51]. In our cohort, however, many of its associations with subsequent disease attenuated after adjustment, consistent with their partly reflecting underlying illness[52]. The extent of attenuation may be specific to patients undergoing abdominal CT and warrants evaluation in broader patient populations.

Independence of infection risk markers: At Discovery, L127 (cytopenia) retained 10 of its 14 infection associations after adjustment, compared with 8 of 14 for L108 (cholestasis). L91 (severe hypoalbuminaemia) retained associations with sepsis and bacteremia. I54 (intra-abdominal or pelvic abscess) retained associations with bacteremia and drug-resistant infections, indicating that localized infectious findings retained prognostic information beyond accompanying abnormalities.

Independence of metastatic sequelae: R64 (metastatic disease) maintained adjusted associations with cachexia[47] at Discovery and Internal Test, acute renal failure[53] at External Site #1, and sepsis[48] across these three cohorts, consistent with reported systemic sequelae of advanced cancer.

Glyph finds novel associations beyond the clinical literature

Having established that Glyph concepts identify significant, reproducible markers of disease risk, we next asked whether they represent novel discoveries beyond those reported in the clinical literature. Specifically, we evaluated 21 future outcomes spanning multiple disease groups. To characterize published abdominal imaging and laboratory markers for these outcomes, an agent acting as a co-scientist systematically identified and filtered 342,647 unique publications using the PubMed API[54] and Paperclip[55] (see Methods for additional details). This yielded 412 literature-derived markers that could be tested in our data, averaging 19.6 per axis, which we then systematically re-evaluated. These included both singular measurements, such as low albumin, and composite clinical predictors such as FIB-4[56], which combines age, transaminases and platelet count. Finally, to assess whether Glyph concepts add prognostic value beyond these markers, we selected the top ten Glyph concepts per outcome using evidence from the Discovery Site, excluding excluding concepts that primarily reflect care intensity such as R399 (pacemaker). We then fit one model using the literature-derived markers alone and another combining them with the selected Glyph concepts to test whether Glyph concepts identified prognostic information beyond published markers.

Results of percentage gains in AUC are presented in Figure 5b. Both the literature baseline and Glyph enhanced model were fit at the Discovery Site, and evaluated on the pooled Internal Test and External Site #1 cohorts. We found that, per-outcome, literature baselines achieved 0.749 AUC (95% CI 0.737–0.761) on held-out data, and that ten Glyph concepts significantly improved this to 0.774 (95% CI 0.762–0.785, p=2.66×10−28p = 2.66 \times10^{-28}). Observed benefits were most concentrated in metabolic, haematologic and infectious outcomes, averaging a relative AUROC improvement of 6.64% (95% CI 5.26–8.01%). These groups also had 5.0 fewer papers on average (10.9 vs 15.9) than the other groups evaluated, indicating they may be under-researched.

Figure 5. Glyph identifies associations beyond the clinical literature and adds prognostic value. a Screening 342,647 PubMed papers across 21 outcomes yielded 294 representative studies and 412 literature-derived predictors. b Adding ten Glyph concepts per outcome improved held-out AUC beyond literature markers, with the largest gains in metabolic, haematologic, and infectious outcomes. c On average, Glyph concepts outperform literature markers, even when top performing markers are selected on the Discovery Set. d Glyph concepts provide discoveries that complement the history of clinical research. The additional prognostic benefit varied across axes, as illustrated by the improvement for HFpEF (top) and the larger gain for sepsis (bottom).

We then explored which concepts were driving gains over literature markers. The top ten Glyph concepts per outcome, selected at the Discovery Site, reproduced more frequently in the pooled held-out cohort (64.8%) than the literature-derived marker set (28.2%, p=1.86×10−18p = 1.86 \times10^{-18}), the top ten performing literature markers (41.2%, p=2.54×10−6p = 2.54 \times10^{-6}), and corresponding Phecodes (54.8%, p=0.062p = 0.062). After adjustment for age, sex and race, reproduction required p<0.05p < 0.05 and the same effect direction as at the Discovery Site. Median confounder-adjusted odds ratios in the pooled evaluation cohort were also significantly higher for Glyph (1.31) than for the literature-derived marker set (1.10, p=5.0×10−5p = 5.0 \times10^{-5}), the top ten literature markers (1.16, p=5.0×10−5p = 5.0 \times10^{-5}), and corresponding Phecodes (1.19, p=5.0×10−5p = 5.0 \times10^{-5}; Figure 5c). A direct comparison between all Glyph concepts and all Phecodes is provided in Supplementary Section A.2 and Figure 13. Overall, this suggests that Glyph captures prognostic information in abdominal CT cohorts, accessible through FM embeddings, that is either underrepresented in the established clinical literature, or is difficult to measure with conventional clinical markers.

Discovery from learned concepts versus curated diagnosis codes

Figure 13. Glyph concepts build on pre-coded discovery. Phecodes group ICD codes into manually curated, clinically meaningful concepts for use in research[2]. We test complementary benefits of data-driven Glyph concepts over Phecodes. a–b Binary Phecodes drawn from the full recorded history yield 3.4× more significant associations than Glyph concepts, but reproduce less often at Internal Test (49% versus 69%) and have a smaller median effect (OR 1.33 versus 1.45). Recent Phecodes, restricted to the scan-adjacent window, also yield more associations than Glyph concepts but reproduce less often. c Directly corresponding Phecodes were often less specific than imaging- or laboratory-derived Glyph concepts and yielded weaker or nonsignificant associations. In other cases, the broader clinical history captured by Phecodes provided a stronger proxy for future risk. For example, hypertension served as a proxy for atherosclerosis (bottom) and outperformed the corresponding Glyph concept for some outcomes.

To review specific axes in detail, we plotted markers according to their publication dates and assessed which of the ten Glyph concepts were represented among the literature-derived markers. Adding Glyph concepts did not significantly improve AUROC across every axis. For example, no significant improvement was observed for dementia (0.871 to 0.870, p=0.80p = 0.80) or Alzheimer’s disease (0.840 to 0.842, p=0.71p = 0.71), suggesting little to no additional prognostic information beyond the literature-derived markers for these outcomes. In contrast, for heart failure with preserved ejection fraction (HFpEF; Figure 5d), Glyph identified risk markers not represented among the literature-derived markers, including I65 (splenomegaly), L84 (azotemia), L127 (cytopenia), L108 (cholestasis), and I87 (uncharacterized hepatobiliary pattern). Adding Glyph concepts significantly improved AUROC by 3.6% from 0.824 to 0.854 (p=4.7×10−7p = 4.7 \times10^{-7}). Similarly, along the abdominal health to sepsis axis, adding Glyph concepts significantly increased AUROC by 4.9% from 0.673 to 0.706 (p=3.2×10−7p = 3.2 \times10^{-7}; Figure 5e). These gains demonstrate additional prognostic information beyond the evaluated literature-derived markers, including contributions from imaging findings not routinely documented in reports, such as I76.

Glyph MCP provides an interface for open-ended investigation

Finally, we demonstrate how Glyph concepts support open-ended, agentic discovery via the Glyph MCP, through a set of illustrative workflows for agents and clinicians. The MCP enables more complex, user-guided discovery workflows. Targeted disease modeling: Beyond automatic discovery, Glyph concepts can be manually assembled into composite features to test a user-defined hypothesis about a

disease, for example testing whether an undocumented imaging concept repeatedly linked to septicemia is confounded by body composition or existing disease markers (Figure 6). Concept ranking for a target outcome: A user can query the concept space directly in natural language, for example asking for the top associations for future myocardial infarction; the Glyph MCP resolves this to the underlying outcome and returns the concepts with the largest prognostic association among those surviving Bonferroni correction, predominantly drawn from laboratory panels in this example (Extended Figure 8).

Figure 6. The MCP enables more complex analyses that build on a foundational layer of Glyph concepts. In this example, concepts were combined into nodes that competed against one as independent predictors for inclusion in a disease model of septicemia. I76 (an undocumented imaging concept linked to septicemia) provided a predictive contribution independent of hepatic disease markers, fluid accumulation, BMI, inflammatory markers, malignancy, and evidence of clinical intervention, and contributed to prognostic AUC predictive across three cohorts.

Outside of discovery, Glyph also brings explainability to typical use of foundation models: Individualized risk explanation: Beyond population-level associations, Glyph concepts can explain an individual patient’s risk prediction, attributing it to the specific concepts driving that prediction and contextualizing them against population-level evidence (Extended Figure 9). Patients-like-me: Given a query patient, Glyph retrieves the most similar patient in the cohort and explains their similarity in terms of shared concepts and the rarity of their joint activation, illustrating how the concept space supports case-based reasoning beyond simple embedding distance (Extended Figure 10).

Discussion

In this paper we introduced Glyph, a method for transforming foundation model (FM) embeddings into engines of scientific discovery. At its core, Glyph supposes that FM embeddings, trained on large volumes of clinical data, encode established clinical markers alongside unexplored features that could reveal new insights into disease mechanisms. To extract these features, Glyph builds on mechanistic interpretability and establishes their value as coded, interpretable clinical concepts for scientific investigation. In our dataset, the resulting set of extracted concepts enabled rediscovery of established organ-specific and systemic risk markers associated with the onset of future disease. In addition, we were able to automatically establish that many of these findings were reproducible across three sites without model retraining. Transforming embeddings into separate, interpretable and measurable concepts enabled adjustment for potential confounders identified directly from the data, beyond routinely collected covariates or those selected from prior clinical knowledge. Applying Glyph on large clinical datasets revealed complementary associations and risk factors not reported in established clinical literature. Moreover, complementing literature-derived markers with Glyph concepts significantly improved prognostic prediction across several disease axes. Together, these findings demonstrate that interpretability can extend beyond a mechanistic understanding of neural networks, enabling features learned implicitly by FMs to interact directly with existing clinical knowledge and available clinical outcomes.

Glyph rediscovered known patterns of disease progression across hepatic, renal, and cardiovascular axes. It also recovered known cross-system associations, e.g. linking renal decline to metabolic comorbidities or valvular atherosclerosis to cardiovascular and neurological disease. Our observations also reflected the complementary strengths of different clinical modalities. Specifically, laboratory concepts more consistently captured renal disease progression than kidney-related imaging findings. Similarly, imaging-derived findings often showed more reproducible associations than findings documented in radiology reports, particularly for atherosclerotic and hepatic risk markers. Confounder adjustment further characterized discovered associations by attenuating associations known to involve nonspecific markers of illness severity, such as fluid accumulation and hypoalbuminemia. It also emphasized published prognostic relationships, including associations of imaging-derived vascular findings with neurological outcomes[37], renal findings with downstream cardiovascular outcomes[57], and the renal consequences of cholestatic liver disease[58]. Together, the recovery of known disease relationships in patients undergoing abdominal CT supported the validity of the concept space as a foundation for the exploration of novel hypotheses.

Beyond rediscoveries, Glyph expanded the clinical picture of disease risk, revealing prognostic information beyond what established literature-derived markers capture. In particular, comparisons across modalities highlighted laboratory markers such as cholestasis and cytopenia. After adjustment, these markers retained broad systemic risk profiles, while other laboratory concepts showed more specific associations. Azotemia, for example, retained a predominantly renal, metabolic, and hematological risk profile. Beyond laboratory markers, Glyph uncovered previously uncharacterized imaging phenotypes within CT embeddings, including a mixed hepatobiliary pattern whose adjusted associations remained significant across all outcome groups. Furthermore, complementing literature-derived markers with a selection of top performing Glyph concepts reproducibly improved prognostic prediction, particularly in metabolic, haematological and infectious outcomes, indicating that existing published markers lack prognostic information presently available in abdominal CT and laboratory data. Together, these findings demonstrate a practical approach to exploring a systems-level perspective on disease[59, 60] using routinely acquired clinical data. They also motivate evaluating several laboratory markers and imaging phenotypes as potential targets for opportunistic or incidental screening.

Comparison with coded phenotypes further characterized the clinical information captured by Glyph. Phecodes, which group ICD codes into clinically meaningful categories, provide an established basis for automated PheWAS discovery[2]. In some cases, directly matched Phecodes showed the same disease associations as Glyph concepts, but with overall smaller effects that often did not reach significance. For these clinical findings, the stronger signal captured by Glyph enabled both rediscoveries and candidate discoveries that were not detected using diagnosis codes alone. Where Phecodes yielded stronger associations, they reflected related conditions captured in the broader clinical record, beyond the scope of abdominal CT and laboratory data. Across the selected outcomes, this broader coverage yielded more significant associations overall, whereas Glyph associations had larger median effects and higher reproduction rates. Together, these findings suggest that Glyph complements the coded record by making additional phenotypic information in imaging and laboratory data accessible for discovery.

Several clinical and technical considerations are important in interpreting our findings. Primarily, our selection criteria restricted the cohorts to patients undergoing abdominal CT across acute and outpatient settings. This selection likely enriched the study population for abdominal illness, influencing both the associations detected and their estimated prognostic value. For example, enrichment for abdominal disease in this cohort may have contributed to the prominence of markers of hepatobiliary disease and renal dysfunction, such as cholestasis and azotemia. In addition, enrichment for abdominal disease could partly explain why hepatobiliary abnormalities featured among the strongest adjusted markers of cardiovascular outcomes, alongside established vascular, renal and haematological risk markers. Future work could reproduce these analyses in broader patient populations to assess whether the observed associations and relative prognostic value of these markers are preserved.

Furthermore, outcome data used in our study warrant further consideration. We used Phecodes derived from ICD codes as outcome data to standardize evaluation across institutions[2]. However, such coded data may introduce noise and uncertainty, e.g. regarding the precise timing of observed disease patterns. Consequently, some associations may capture existing but undocumented disease rather than risk of future disease onset, while cases never assigned the corresponding code may be missed entirely. Refining outcome labels through clinician adjudication of longitudinal records against predefined diagnostic criteria could correct diagnosis and timing, but would require extensive chart review, potentially supported by automated search of clinical notes. Ultimately, prospective studies using predefined diagnostic criteria and systematic follow-up would be required for stronger validation of these outcomes.

Glyph enabled systematic confounder adjustment across a broad set of clinical concepts. Associations that persist after adjustment indicate that related concepts provide prognostic information beyond clinical covariates and other identified concepts, but do not establish that they represent independent disease processes or causal drivers of disease on a retrospective dataset. Broader clinical coverage, including additional imaging, clinical notes, categorical laboratory results, flowsheets, medication use, and genetic factors, could help account for omitted factors and reveal further confounders and candidate markers. Furthermore, the fixed correlation threshold used in our analysis was designed to exclude potential mediators, downstream consequences through which the tested concept could affect the outcome, but adjustment results were sensitive to the chosen threshold. Future work should guide confounder selection using knowledge of disease mechanisms, rather than statistics alone[61].

Future work could extend the systematic comparison with published literature presented in § 3.5 beyond the selected outcomes. Although it requires substantial token usage, this assessment provides a framework for distinguishing established relationships recovered by Glyph, published associations not detected in our cohorts, and potentially novel findings. Future work could also investigate how concepts change over time and how their combinations contribute to disease risk. The analyses shown in this work related concepts measured around individual CT examinations to future outcomes. Repeated measurement could reveal when concepts emerge, how they evolve, and how their relationships change through disease progression, providing temporal evidence to guide causal investigation. Combined with more complete treatment histories, these trajectories could support investigation of changes associated with clinical intervention. Changes in concept activation could also be evaluated as quantitative indicators of disease state or treatment response by comparing their sensitivity with existing clinical measures. Beyond these temporal extensions, our PheWAS analysis tested concepts individually and did not assess their combined prognostic information. Future work could examine combinations and interactions between concepts to identify more specific clinical phenotypes and risk associations not apparent from individual markers.

Technical considerations and insights arise both from the source embeddings and from how their features are extracted and interpreted. Text embeddings yielded finer-grained concepts than image embeddings, even though radiology reports provide a compressed description of the images. This observation motivates future work in designing image encoders for fine-grained concept discovery, alongside methods for resolving more specific findings within existing embeddings. We also extracted embeddings separately by modality, which simplified attribution of concepts to their source and examination of relationships across modalities. However, some predictive patterns may be inherently multimodal, motivating joint embeddings that could capture combinations of imaging, laboratory, and report findings not fully represented in any single modality. One insight from our confounder analyses is that prediction models built on FM embeddings can draw on both potentially actionable markers and confounded proxies for illness. Risk scores derived from the full FM embedding may therefore identify patients at risk without reliably guiding how that risk can be reduced.

Textual interpretations describe the clinical findings associated with a feature, but do not necessarily establish its precise clinical meaning. In a subset of extracted concepts, automated interpretation required a combination of a small number of distinct clinical or imaging findings, suggesting that not all concepts were monosemantic. It remains unclear whether this reflects polysemanticity in the FM embeddings, failure of the SAE to separate distinct features, or a property of underlying data, in which certain features frequently co-occur making feature separation difficult. Clinical knowledge could guide when related concepts should be grouped or separated into more specific findings, with Glyph providing an efficient basis for refining and testing these definitions. Future work could also complement our global evaluation with local comparisons against neighboring concepts and clinically similar cases that do not activate the concept in question. These comparisons could help distinguish defining findings from co-occurring data properties, severity, or clinical context.

In conclusion, Glyph provides a framework for extracting knowledge from foundation models, enabling reproducible, population-scale discovery of disease markers and associations. Glyph can be applied to any dataset represented by FM embeddings to rapidly investigate a chosen discovery axis, using an agentic co-scientist framework to assess the novelty and added value of extracted findings against the published literature. This enables researchers to automatically validate data-driven insights, such as those presented here, or discover and share new concept networks and insights from their own datasets. Overall, we envision that Glyph can accelerate clinical science by establishing a standard for sharing validated FM features as data-driven phenotypes, connecting discoveries across datasets with existing clinical knowledge.

Methods

Datasets

All datasets are partitioned into a Discovery Site and a series of Testing sets. The Discovery Site is used to train sparse autoencoders, select disease outcomes, interpret concepts, and rank associations. Testing sets are held out entirely and used solely to assess whether insights replicate in patient cohorts that played no part in discovery, in order of increasing institutional distance from Stanford. All datasets were collected under IRB approval with a waiver of informed consent owing to retrospective use.

Discovery Site – Stanford Main hospital. The Discovery Site cohort comprises consecutive abdominal CT examinations performed at Stanford academic medical center from December 2012 to October 2018, yielding 20,391 CT studies from 14,652 unique patients (3,262,560 CT slices, 1,588,003 lab entries, and 22,781,540 historical diagnoses). No patient contributed studies to more than one split. We divide the cohort into a train set of 15,331 studies from 11,009 patients (2,452,960 slices, 6,083,609 report tokens, 1,194,451 lab entries, 17,017,170 diagnoses) and a validation set of 5,060 studies from 3,643 patients (809,600 slices, 2,000,273 report tokens, 393,552 lab entries, 5,764,373 diagnoses). Embedding normalization statistics are computed on the Discovery Site. Finally, all datasets contain abdominal CT images and radiology reports that capture the current health of the patient; laboratory analyses are restricted to the Discovery Site, Internal Test and External Site #1. All have longitudinal ICD codes with at least 4 years of follow-up for association studies, and the six universal clinical covariates, including race.

CT preprocessing follows[14]. From each study, the series with the greatest number of axial slices was selected to maximize information content, and a contrast-phase detection algorithm confirmed that 97% of selected series were portal venous phase. Volumes were preprocessed by orienting axes (left–right, posterior–anterior, inferior–superior), resampling to 1.5 mm in-plane resolution and 3 mm slice thickness, clipping and normalizing Hounsfield units from [−1000,1000][-1000,1000] to [0,1][0,1], and center-cropping to 224×224×160224 \times224 \times160 voxels.

Report preprocessing follows[14]. Only the findings section of each report was used, as it provides the most granular per-organ-system observations. It covers the lower thorax, liver and biliary tree, gallbladder, spleen, pancreas, adrenal glands, kidneys and ureters, gastrointestinal tract, peritoneal cavity, pelvic organs, vasculature and lymph nodes, and musculoskeletal system.

Longitudinal lab measurements were collected from the EHR for each patient. Only numerical lab types were retained, defined as those where at least 90% of recorded measurements were numeric, resulting in 350 lab types (Supplementary Table 10). To reduce noise in dense measurement periods, a quarterly downsampling strategy was applied, retaining only the last recorded value per patient, lab type, and calendar quarter, using only measurements taken within the year prior to the CT scan date. Each lab value was z-score normalized per lab type and clamped to [−3,3][-3,3] to limit the influence of extreme outliers.

Name (source units)IDMeasurementsNumeric (%)Mean (SD)IQR
A
Anion Gap (mmol/L)24661,870,63999.49.41 (3.61)4.00
Albumin, Ser/Plas (g/dL)17511,374,12799.93.49 (10.2)1.20
ALT (SGPT), Ser/Plas (U/L)11191,322,52998.554.8 (158)27.0
AST (SGOT), Ser/Plas (U/L)13981,320,65399.253.2 (239)21.0
Alk P’TASE, Total, Ser/Plas (U/L)18331,291,861100.0136 (151)66.0
Albumin/Globulin Ratio ((calc))178055,123100.01.59 (0.381)0.400
AST (SGOT), Ser/Plas (IU/L)139747,624100.033.7 (72.1)13.0
ALT (SGPT), Ser/Plas (IU/L)111845,67999.934.5 (64.7)19.0
Alk P’TASE, Total, Ser/Plas (IU/L)183245,241100.0102 (100)43.0
Anion Gap (mEq/L)246529,895100.09.24 (3.37)4.00
Albumin/Globulin Ratio (no units)178229,449100.01.69 (0.495)0.500
Alpha Fetoprotein (ng/mL)209219,60498.2873 (1.12 × 10410^{4})4.80
Amylase, Serum (U/L)238418,64499.893.9 (176)48.0
Alk P’TASE, Total (U/L)183117,145100.0173 (156)97.0
Ammonia, Plasma (umol/L)230116,80695.555.0 (44.8)42.0
B
BUN, Ser/Plas (mg/dL)35201,264,04699.821.7 (16.4)14.0
Basophil % (%)3681870,249100.00.623 (0.529)0.500
Basophil, Absolute (K/uL)3694870,231100.00.0442 (0.0764)0.0400
BASO, % (%)3236339,19698.50.515 (0.502)0.500
BASOS, ABS (K/uL)3248296,83199.90.0358 (0.0561)0.0323
BUN (mg/dL)3512231,53199.616.9 (13.3)10.0
Base Excess, ISTAT (mmol/L)3654205,40499.81.33 (6.45)7.00
Band Neutrophils (%)357478,717100.010.5 (12.0)13.0
Base Excess (a) (mEq/L)363144,101100.0-0.392 (6.24)6.30
Base Excess (vt) (mmol/L)364638,998100.0-0.126 (5.54)6.60
Basophils (%)370235,424100.01.79 (1.90)1.00
Bun/Creatinine Ratio (no units)423031,994100.018.8 (6.54)8.00
BASOS, ABS (cells/uL)325129,441100.042.3 (127)30.0
BASOS, % (%)324028,883100.00.673 (0.560)0.400
Base Deficit, ISTAT (mmol/L)362126,406100.04.91 (3.91)5.00
Basophils, ABS (man diff) (K/uL)371425,787100.00.215 (0.838)0.110
BASOS, ABS (x10E3/uL)325618,773100.00.0330 (0.0569)0.100
Base Deficit (vt) (mEq/L)361718,606100.04.21 (3.16)4.10
Base Excess (vt) (mEq/L)364518,46999.81.97 (5.12)4.20
Base Excess (a) (mmol/L)363218,42199.80.102 (6.06)6.50
Bilirubin, Total (mg/dL)393617,80593.71.42 (3.75)0.500
BASO, ABS (K/uL)323817,65499.80.0385 (0.0370)0.0300
Band neutrophils (%)358116,405100.013.9 (13.6)17.0
BUN/Creatinine (ratio)352813,386100.016.9 (8.53)9.00
Name (source units)IDMeasurementsNumeric (%)Mean (SD)IQR
C
Creatinine, Ser/Plas (mg/dL)75252,017,864100.01.17 (1.21)0.490
Calcium, Ser/Plas (mg/dL)60401,990,281100.08.82 (0.746)0.900
CO2, Ser/Plas (mmol/L)56611,772,336100.025.7 (3.85)5.00
Chloride, Ser/Plas (mmol/L)65891,771,741100.0102 (4.95)5.00
Chloride (mmol/L)6557198,900100.0103 (4.70)5.00
CO2 (mmol/L)5622198,56299.825.4 (3.80)5.00
Corrected Calcium (mg/dL)7295167,940100.09.27 (0.552)0.600
Cholesterol/HDL Ratio (ratio)6677128,317100.03.40 (1.67)1.50
Cholesterol, Total, Ser/Plas (mg/dL)666887,391100.0170 (47.4)60.0
Calcium Ionized (mmol/L)599383,07499.91.14 (0.152)0.150
Cholesterol, Total (mg/dL)666380,75799.9179 (46.7)59.0
Calcium Ionized, Whole Blood (mmol/L)599877,643100.01.15 (0.149)0.150
Chloride, Whole Bld (mmol/L)660973,94999.0107 (6.72)8.00
Calcium, Ionized (mmol/L)603269,87499.91.12 (0.160)0.140
ctO2 (a) (mL/dL)2448660,074100.013.6 (2.85)3.90
ctO2 (v) (mL/dL)2448858,171100.09.46 (3.03)3.60
Creatinine, Urine (mg/dL)757348,62399.595.1 (70.1)81.8
Creatinine,ISTAT (mg/dL)759443,17899.81.18 (1.28)0.500
Chloride, Whole Blood (mmol/L)661142,93199.1106 (6.35)7.00
CK, Total, Ser/Plas (U/L)544539,72899.3565 (2.82×1032.82 \times10^{3})152
Chloride, Ser/Plas (mEq/L)658733,356100.0104 (4.76)5.00
CO2, Ser/Plas (mEq/L)565933,02899.925.6 (3.45)4.00
Cancer Antigen 125 (U/mL)614627,67898.9234 (1.17×1031.17 \times10^{3})56.0
CK, Total (U/L)544224,66098.6366 (1.28×1031.28 \times10^{3})138
Carcinoembryonic Antigen (CEA) (ng/mL)630224,00196.668.7 (493)8.00
Calcium,Ion, ISTAT (mmol/L)608520,69999.81.18 (0.147)0.140
Cortisol, Serum (ug/dL)736817,32997.913.8 (13.9)10.5
Calcium (mg/dL)597517,238100.09.29 (0.869)0.600
Cancer Antigen 19-9 (U/mL)615716,98191.61.16×1041.16 \times10^{4} (1.07×1051.07 \times10^{5})650
Course Elapsed Days (no units)740616,728100.014.7 (12.4)19.0
Creatinine, Urine (mg/dl)757615,73999.9102 (65.7)75.1
Creatinine (mg/dL)746415,595100.04.46 (23.5)0.370
Cholesterol/HDL Ratio ((calc))667211,371100.03.41 (1.15)1.30
CKD-EPI 2021CR,EGFR (mL/min/1.73)547010,986100.080.1 (23.3)32.0
CKMB % Relative Index (%)547610,889100.03.59 (2.94)2.70
D
Direct LDL Chol (mg/dL)834169,54099.999.1 (37.9)50.0
E
Eosinophil % (%)9034870,251100.02.30 (2.62)2.50
Eosinophil, Absolute (K/uL)9038870,235100.00.157 (0.271)0.160
eGFR for African American (mL/min/1.73 m2)24596638,60590.092.1 (37.7)55.0
EOS, % (%)8803368,24498.62.33 (2.68)2.50
EOS, ABS (K/uL)8811314,48499.90.156 (0.269)0.160
eGFR (mL/min/1.73m2)2458186,580100.081.7 (26.2)36.0
Eosinophils (%)904955,904100.03.50 (4.89)3.00
Name (source units)IDMeasurementsNumeric (%)Mean (SD)IQR
Eosinophils, ABS (man diff) (K/uL)907341,259100.00.327 (0.905)0.240
eGFR (mL/min/1.73)2458041,11299.882.6 (24.9)34.0
ESR, Automated (mm/hr)888036,67499.729.0 (28.8)31.0
EOS, ABS (cells/uL)881429,441100.0166 (200)137
EOS, ABS (x10E3/uL)881918,773100.00.189 (0.197)0.100
E/E’ Ratio (no units)856316,793100.012.3 (10.7)6.20
EGFR (mL/min/1.73m2)876214,910100.079.4 (25.4)37.0
ESR, (automated) (mm/hr)887410,90999.529.3 (29.7)30.0
ESR, (manual) (mm/hr)887810,68398.629.0 (29.3)32.0
F
FIO2 (%)949084,158100.042.3 (25.7)20.0
Ferritin (ng/mL)973980,48799.9555 (2.33 × 10³)313
Fibrinogen (mg/dL)976653,18199.8341 (180)247
Free Thyroxine (ng/dL)997348,98299.71.23 (0.375)0.390
Free T4 (ng/dL)995424,80499.71.22 (0.357)0.400
FEF 25-75 (no units)944321,145100.01.68 (1.14)1.61
FEV1 (no units)945621,145100.01.92 (0.810)1.09
FEV1/FVC (no units)946321,145100.071.9 (13.8)18.9
FVC (no units)956121,145100.02.67 (1.03)1.36
Fast Vent Tachycardia Detect (no units)965714,271100.06.52 (159)0.00
G
Glucose, Ser/Plas (mg/dL)106591,997,727100.0128 (58.9)43.0
Glucose by Meter (mg/dL)105521,701,28299.8165 (68.6)78.0
Globulin (g/dL)104511,164,733100.03.39 (0.847)1.10
Glucose, Whole Blood (mg/dL)10680120,02099.9153 (60.3)60.0
Globulin (g/dL (calc))1045055,123100.02.66 (0.539)0.600
Gamma-Glutamyl Trans (GGT) (U/L)1031635,80999.1217 (386)178
GRAN, % (%)1024730,647100.060.4 (16.1)17.0
Glucose,ISTAT (mg/dL)1069818,09299.4145 (87.4)61.0
Glucose, Non-fasting (mg/dL)1062613,225100.0227 (139)231
Glucose, Fasting (mg/dL)1060211,72499.9107 (39.2)21.0
H
Hematocrit (%)117891,779,663100.033.8 (7.00)10.9
Hemoglobin (g/dL)118681,763,945100.011.1 (2.38)3.70
Heart Rate (bpm)11744324,57699.982.9 (21.5)28.0
Hematocrit (HCT) (%)11796201,191100.032.8 (6.73)10.8
Hemoglobin (HGB) (g/dL)11833200,229100.010.7 (2.28)3.70
HDL Cholesterol (mg/dL)11085158,14599.855.4 (18.8)24.0
Hemoglobin A1c (%)11886145,05099.66.54 (1.65)1.60
HCO3 (a), ISTAT (mmol/L)10934136,455100.025.8 (6.42)7.30
HCO3 (v), ISTAT (mmol/L)1095991,209100.025.5 (5.64)5.80
Hct (Est) (%)1171275,806100.029.6 (6.30)8.00
HgB (g/dL)1219747,676100.09.41 (2.01)2.40
HCO3 (a) (mEq/L)1093044,101100.024.7 (6.39)6.50
HCO3 (mEq/L)1094240,202100.023.7 (5.25)5.60
Hct (Est) (a) (%)1171339,201100.029.8 (6.43)8.00
HCO3 (mmol/L)1094639,000100.024.6 (5.60)6.50
25-Hydroxy D, Total (ng/mL)33336,86398.733.6 (15.2)17.0
Name (source units)IDMeasurementsNumeric (%)Mean (SD)IQR
25-Hydroxy D3 (ng/mL)33735,97898.329.8 (15.9)21.0
Hgb(Calc), ISTAT (g/dL)1224134,392100.011.1 (2.69)4.10
Hct, ISTAT (%)1173323,47899.432.5 (8.00)12.0
HCO3 (a) (mmol/L)1093118,424100.025.0 (6.43)6.50
Hemoglobin A1c (% of total Hgb)1188517,49999.76.50 (1.50)1.60
High Sensitivity CRP (mg/L)1225212,39797.79.22 (24.6)6.40
Hct(Calc), ISTAT (%)1172511,10599.832.7 (7.70)11.0
I
Imm. Granulocyte, % (%)12912451,677100.00.721 (1.21)0.400
Imm. Granulocyte, Abs (K/uL)12915451,661100.00.0693 (0.202)0.0400
INR (no units)12661272,87299.91.62 (0.868)0.800
INR (INR)12644171,55999.81.60 (0.942)0.700
Iron, Total (ug/dL)1332579,10599.963.1 (47.5)52.0
INR, ISTAT (INR)1267526,85597.51.68 (0.922)1.20
Immature Granulocytes (%)1293020,288100.00.239 (0.603)0.200
IgG, Serum (mg/dL)1286017,44299.9915 (595)582
Immature Grans (Abs) (x10E3/uL)1292916,729100.00.0135 (0.0898)0.00
IgA, Serum (mg/dL)1276711,68694.8219 (298)165
L
Lymphocyte % (%)14418870,283100.021.1 (12.2)16.6
Lymphocyte, Absolute (K/uL)14431870,270100.01.48 (2.85)1.05
LYM, % (%)13860358,18498.521.7 (13.1)17.7
LYM, ABS (K/uL)13865302,59499.91.47 (2.58)1.07
Lipase (U/L)14200148,06696.7151 (264)133
Lactate, Whole Bld (mmol/L)13940129,12199.42.16 (1.84)1.40
Lymphocytes (%)14448110,483100.020.5 (19.1)21.0
LDH, Total, Ser/Plas (U/L)1363889,81899.1338 (327)173
Lymphocytes, ABS (man diff) (K/uL)1446880,719100.02.07 (13.3)1.15
Lactate, ISTAT (mmol/L)1393570,00198.31.74 (1.78)1.16
Lactic Acid (mmol/L)1394862,84599.42.54 (2.42)1.40
LDL/HDL Ratio (ratio)1373260,81899.91.83 (0.910)1.10
LEFT VENTRICULAR EJECTION FRACTION (%)1374154,79998.157.4 (11.3)10.3
Lidocaine (ug/mL)1417137,93094.82.94 (2.97)1.70
LYMPH, ABS (K/uL)1388629,715100.01.93 (4.23)0.950
LYMPHS, % (%)1389128,881100.026.5 (11.5)14.0
LYMPH, ABS (cells/uL)1388828,451100.01.85×1031.85 \times10^3 (6.41×1036.41 \times10^3)973
LDL Calculated (NIH) (mg/dL)1366525,849100.090.2 (36.7)50.0
LDL (Calculated-Friedewald) (mg/dL)1364625,449100.091.9 (36.3)49.0
LVEDD (cm)1385524,297100.04.53 (0.786)1.00
LV Wall Thickness (LVPWD) (cm)1385324,239100.01.05 (0.230)0.300
LV Wall Thickness (IVSD) (cm)1385124,238100.01.09 (0.252)0.300
LV EDV (ml)1381924,079100.089.6 (44.2)47.3
LDL (Calculated) (mg/dL)1364423,30397.9106 (35.9)49.0
Lym, ABS (man diff) (K/uL)1439919,494100.02.12 (7.67)1.50
LYM, ABS (x10E3/uL)1387118,364100.02.00 (5.52)1.00
Lactate, ISTAT (mmole/L)1393616,29398.81.63 (1.40)1.10
LV Lead Channel Impedance (ohm)1382211,519100.0638 (531)304
LDL (Calculated) (mg/dL (calc))1364311,09299.998.1 (34.7)49.0
Name (source units)IDMeasurementsNumeric (%)Mean (SD)IQR
Lymphocytes % (%)1443410,509100.025.8 (12.4)15.8
M
MCHC (g/dL)146531,755,052100.032.9 (1.40)1.70
MCV (fL)146781,754,100100.090.7 (7.72)8.70
MCH (pg)146471,738,293100.029.8 (2.96)3.30
Monocyte % (%)15483870,249100.08.70 (4.00)4.10
Monocyte, Absolute (K/uL)15487870,235100.00.633 (0.431)0.380
Magnesium, Ser/Plas (mg/dL)14991694,311100.01.96 (0.388)0.300
MONO, % (%)14829344,08398.58.29 (4.04)4.10
MONO, ABS (K/uL)14836314,49499.90.606 (0.404)0.380
Mean Corpuscular Volume (MCV) (fL)15107199,193100.088.3 (8.02)9.20
Mean Corpuscular Hemoglobin (MCH) (pg)15095199,106100.028.9 (3.12)3.60
Mean Corpuscular Hemoglobin Concentration (MCHC) (g/dL)15102199,103100.032.7 (1.40)1.70
Monocytes (%)15500136,552100.08.64 (7.53)6.50
MPV (fL)14865119,96699.98.64 (2.07)3.30
Monocytes, ABS (man diff) (K/uL)1551578,392100.00.771 (2.26)0.640
Metamyelocytes (%)1519642,84599.32.87 (2.97)2.50
Myelocytes (%)1569742,31399.33.32 (3.56)3.00
Metamyelo, ABS (man diff) (K/uL)1519034,552100.00.434 (1.32)0.310
Myelo, ABS (man diff) (K/uL)1569334,520100.00.545 (2.15)0.360
MONOS, ABS (cells/uL)1485028,663100.0552 (772)261
MONO, ABS (x10E3/uL)1484418,779100.00.554 (0.341)0.300
MID, % (%)1479718,447100.010.1 (4.12)3.60
MID, ABS (K/uL)1479818,447100.00.719 (1.17)0.400
Mono, ABS (man diff) (K/uL)1546218,341100.00.794 (1.03)0.800
N
Neutrophil, Absolute (K/uL)16074832,263100.05.57 (4.06)3.87
Neutrophil % (%)16059830,438100.067.2 (14.0)19.0
nRBC, % (%)24724541,869100.00.0711 (1.33)0.00
nRBC, Abs (K/uL)24728541,868100.06.85×10−36.85 \times10^{-3} (0.114)0.00
NEUT, ABS (K/uL)15830332,99199.95.69 (4.38)4.10
NEUT, % (%)15823331,17998.267.8 (14.8)20.3
Non-HDL Chol, Calc (mg/dL)16177128,341100.0117 (43.3)55.0
Neut, ABS (Seg+Band) (man diff) (K/uL)1604787,232100.07.11 (9.07)7.20
NT-proBNP (pg/mL)1595062,54795.33.20×1033.20 \times10^3 ( 6.96×1036.96 \times10^3 )2.56×1032.56 \times10^3
Neutrophils, % (%)1610241,337100.062.8 (15.3)16.9
NEUTS, ABS (K/uL)1584239,808100.04.34 (4.03)2.67
Neutrophils (%)1608136,749100.058.3 (16.1)17.6
NEUTS, ABS (cells/uL)1584428,662100.03.92×1033.92 \times10^3 ( 2.82×1032.82 \times10^3 )2.16×1032.16 \times10^3
Neut, ABS (man diff) (K/uL)1604919,496100.07.75 (10.8)8.30
NEUT, ABS (x10E3/uL)1583818,773100.04.05 (2.40)2.20
nRBC/100WBC (/100 WBC)2473314,319100.09.11 (30.0)4.00
Nucleated Count (/uL)1632911,44399.81.04×1041.04 \times10^4 ( 8.63×1048.63 \times10^4 )756
Neutrophils %, Body Fluid (%)1607811,297100.031.7 (32.1)53.0
O
O2 Saturation, ISTAT (%)16402109,015100.095.6 (6.14)5.00
Name (source units)IDMeasurementsNumeric (%)Mean (SD)IQR
O2 Saturation (v) (%)1639573,048100.073.4 (15.8)22.1
O2 Saturation, ISTAT (Ven) (%)1640757,896100.061.6 (21.0)31.0
O2 Saturation (a) (%)1638752,691100.095.6 (3.94)2.00
O2 Saturation (V), ISTAT (%)1638533,115100.060.8 (21.3)33.0
O2 Saturation (A), ISTAT (%)1637432,443100.095.8 (6.31)4.00
O2-Hemoglobin (a) (%)1641117,401100.096.2 (3.56)1.80
25-OH Vitamin D, Total (ng/mL)34313,51097.834.7 (16.4)20.0
Osmolality, Urine (mOsm/kg)1662611,32299.9408 (177)232
P
Potassium, Ser/Plas (mmol/L)184012,016,302100.04.09 (0.563)0.700
Platelet count (K/uL)181041,684,105100.0228 (127)135
Protein, Total, Ser/Plas (g/dL)187231,324,762100.06.82 (0.965)1.20
Phosphorus, Ser/Plas (mg/dL)17893443,627100.03.52 (1.18)1.30
P-R Interval (ms)16799293,87999.7165 (37.2)37.0
pH, urine (no units)24875258,05299.96.18 (0.857)1.50
Prothrombin Time (seconds)18805257,975100.018.3 (6.85)6.70
Platelet Count (PLT) (K/uL)18071200,972100.0239 (134)154
pH (no units)24842197,81199.46.74 (0.882)1.42
Part. Thromboplastin Time (seconds)17498171,55799.246.9 (27.7)21.7
Prothrombin Time (sec)18803148,08899.717.8 (7.35)5.90
P Axis (deg)16793145,94597.845.9 (36.9)35.0
P Axis (degrees)16794144,58599.742.6 (31.9)35.0
PO2 (a), ISTAT (mmHg)17091141,736100.0113 (68.9)54.0
PH (a), ISTAT (no units)16943141,614100.07.39 (0.106)0.100
pCO2 (a), ISTAT (mmHg)24786141,60199.942.2 (11.3)11.4
PO2 (v), ISTAT (mmHg)1709491,568100.036.7 (18.8)17.0
PH (v), ISTAT (no units)1694691,449100.07.38 (0.0796)0.0800
PCO2 (v), ISTAT (mmHg)1688891,43199.942.8 (11.0)11.2
Part. Thromboplastin Time (sec)1749680,38199.143.8 (24.5)16.5
Potassium, Whole Bld (mmol/L)1841674,24799.94.04 (0.740)0.700
pH (v) (no units)2484672,61699.97.36 (0.0763)0.0940
pH (a) (no units)2483571,85499.97.38 (0.0801)0.0900
pCO2 (v) (mm Hg)2480144,44699.947.0 (13.5)14.5
pO2 (v) (mm Hg)2491344,44696.255.2 (30.9)23.1
Potassium, Whole Blood (mmol/L)1841843,140100.04.10 (0.749)0.700
Potassium, Ser/Plas (mEq/L)1839839,663100.04.09 (0.629)0.700
PSA (ng/mL)1724738,28490.921.3 (158)3.30
Potassium, ISTAT (mmol/L)1839036,84099.54.06 (0.784)0.800
Platelet count (Thousand/uL)1810634,98699.6222 (98.7)116
Prothrombin Time (Sec.)1879833,943100.018.5 (7.06)6.85
pCO2 (v) (mmHg)2480229,16399.547.4 (11.8)12.2
pO2 (v) (mmHg)2491429,15791.648.2 (24.1)14.0
pCO2 (a) (mmHg)2478328,264100.040.7 (9.09)8.00
pO2 (a) (mmHg)2489528,26198.6193 (102)132
pCO2 (a) (mm Hg)2478228,125100.040.2 (10.4)9.20
pO2 (a) (mm Hg)2489428,08899.3178 (109)140
Platelet Count (K/uL)1801027,173100.0232 (143)155
Protein, Total, Urine (mg/dL)1873223,61091.7103 (210)79.0
Platelet count (x10E3/uL)1811521,75899.8239 (101)118
Protein/Creat ratio (mg/mg)1875219,73892.61.41 (3.02)1.10
Name (source units)IDMeasurementsNumeric (%)Mean (SD)IQR
Part. Thromboplastin Time (Sec.)1749419,61599.547.0 (25.9)22.7
Plan Total Prescribed Dose (cGy)1796018,130100.04.44×1034.44 \times10^3 ( 1.61×1031.61 \times10^3)2.50×1032.50 \times10^3
Plan Total Fractions Prescribed (no units)1795918,106100.018.4 (9.54)16.0
Plan Fractions Treated to Date (no units)1795418,095100.09.69 (7.64)12.0
Plan Prescribed Dose Per Fraction (Gy)1795718,063100.03.25 (2.93)1.00
Parathyroid Hormone (pg/mL)1747017,74998.4110 (166)74.5
Protein, Total (g/dL)1869717,387100.06.92 (1.11)1.50
PT, ISTAT (sec)1729816,871100.017.5 (9.03)6.50
pCO2 (a) (mm HG)2478115,84099.939.2 (8.44)7.10
pO2 (a) (mm HG)2489315,832100.0182 (95.5)136
Prealbumin (mg/dL)1843914,31699.317.1 (7.72)8.00
p50 (A) (mm Hg)2475612,48399.927.2 (5.39)2.70
p50 (v) (mm Hg)2476512,417100.029.1 (2.81)3.60
Promyelocytes (%)1858411,20197.32.16 (2.92)1.00
Promyelo, ABS (man diff) (K/uL)1857710,882100.00.363 (1.14)0.250
Q
QRSD Interval (ms)18914324,513100.097.8 (24.0)21.0
QTC Interval (ms)18918324,14499.7432 (39.9)44.0
QT Interval (ms)18916323,75299.9392 (52.2)65.0
QRS Axis (deg)18911165,45699.519.8 (55.6)69.0
QRS Axis (degrees)18912158,666100.022.5 (51.4)62.0
R
RDW (%)191921,758,819100.015.7 (3.00)3.50
RBC (MIL/uL)191121,626,330100.03.73 (0.821)1.25
RR (ms)19362324,108100.0771 (191)256
RDW SD (fL)19211313,731100.049.6 (8.69)9.70
Red Blood Cells (RBC) (MIL/uL)19509199,260100.03.74 (0.810)1.25
Red Cell Distribution Width (RDW) (%)19512198,954100.016.0 (3.20)3.90
RBC (M/uL)1905570,542100.03.89 (0.720)0.960
RBC (Million/uL)1911434,985100.04.20 (0.724)0.970
Retic % (Automated) (%)1973322,62698.52.84 (2.59)1.97
RBC (x10E6/uL)1912521,802100.04.33 (0.669)0.790
RVSP (mmHg)1943119,462100.034.3 (13.7)15.7
Retic, Absolute (Automated) (K/uL)1975919,043100.089.7 (62.9)61.0
RA Brady Sensing Blanking Period (ms)1899718,877100.0150 (46.6)100
Reference Point Dosage Given to Date (Gy)1953618,212100.024.7 (15.9)25.0
Reference Point Session Dosage Given (Gy)1953817,367100.03.13 (2.68)1.00
RV Brady Sensing Blanking Period (ms)1938215,88399.8149 (57.2)80.0
RV Lead Channel Impedance (ohm)1939814,357100.0472 (295)169
RA Lead Channel Impedance (ohm)1900411,252100.0455 (805)110
RV Lead Channel Sensing Intrinsic Amplitude (mV)1940111,186100.011.3 (6.52)8.25
Name (source units)IDMeasurementsNumeric (%)Mean (SD)IQR
S
Sodium, Ser/Plas (mmol/L)211371,794,503100.0137 (4.42)5.00
Specific Gravity, urine (no units)21230257,38997.31.02 (9.61×10−39.61 \times10^{-3})0.0120
Sodium (mmol/L)21082201,110100.0137 (4.06)4.00
Specific Gravity (no units)21225127,64297.01.24 (15.0)0.0110
Sodium, Whole Blood (mmol/L)21153116,861100.0139 (5.60)6.00
Segmented Neutrophils (%)2068281,296100.054.8 (22.9)32.8
Sodium, Ser/Plas (mEq/L)2113536,252100.0138 (3.70)4.00
Sodium, ISTAT (mmol/L)2112736,150100.0139 (5.72)6.00
Seg neutrophils (%)2067819,210100.047.8 (23.1)36.0
T
Total Bilirubin, Ser/Plas (mg/dL)22508825,30797.51.08 (2.70)0.500
Triglyceride, Ser/Plas (mg/dL)22843199,73999.8137 (138)87.0
T Wave Axis (deg)21585162,39299.749.1 (54.9)51.0
T Wave Axis (degrees)21586157,076100.045.4 (54.1)49.0
TSH (uIU/mL)21963155,16799.43.29 (10.1)1.93
TCO2 (a), ISTAT (mmol/L)21728102,17799.126.9 (6.36)7.00
TCO2 (v), ISTAT (mmol/L)2173391,20999.526.6 (5.55)6.00
Tacrolimus (FK506) (ng/mL)2199185,86797.57.90 (4.16)5.00
Tacrolimus (FK506) by IA (ng/mL)2199559,56398.37.66 (4.12)4.80
Transferrin Sat (%)2273159,35199.822.8 (17.7)18.0
Total Cells Counted (no units)2252054,879100.099.2 (12.2)0.00
TIBC (ug/dL)2180149,820100.0298 (90.6)118
TCO2 (a) (mEq/L)2172546,279100.026.0 (6.60)6.50
TCO2, ISTAT (mmol/L)2173839,65599.726.5 (5.74)7.00
tCO2 (mmol/L)2495039,007100.026.0 (5.85)6.70
tHB (a) (g/dL)2496338,880100.09.63 (2.13)2.70
tCO2 (mEq/L)2494937,102100.024.7 (5.61)5.90
TSH (mIU/L)2195731,65499.33.11 (12.9)1.68
tHB (g/dL)2496531,613100.09.90 (2.23)3.10
Thrombin Time (seconds)2228023,15097.418.9 (9.49)3.40
Troponin I, POCT (ng/mL)2290621,218100.00.0330 (0.505)0.0100
TCO2 (a) (mmol/L)2172618,424100.026.4 (6.76)6.70
Troponin I, POCT (ng/ML)2290512,400100.00.0369 (0.685)0.0100
Transferrin (ug/dL)2272610,82299.7301 (93.9)117
U
Urea Nitrogen,Ser/Plas (mg/dL)23203489,12297.518.8 (13.3)11.0
Unconjugated Bilirubin (mg/dL)2314182,01098.40.852 (1.29)0.600
Urobilinogen, Urine (mg/dL)2335969,43299.40.472 (0.813)0.00
Urea Nitrogen, Ser/Plas (mg/dL)2319226,79393.316.3 (12.2)10.0
Urine PH (no units)2333226,41699.65.98 (1.04)2.00
UIBC (ug/dL)2308615,97798.6208 (83.6)106
V
Vancomycin, Trough (ug/mL)2370530,68398.116.5 (8.17)8.50
Variant Lym, ABS (man diff) (K/uL)2371718,269100.01.17 (6.04)0.420
Vitamin D, 25-Hydroxy Total (ng/mL)2391318,12799.640.0 (17.6)21.0
Name (source units)IDMeasurementsNumeric (%)Mean (SD)IQR
W
WBC (K/uL)240451,798,93999.88.47 (7.93)5.00
White Blood Cells (WBC) (K/uL)2426699,53699.68.41 (7.51)5.60
WBC (Thousand/uL)2410235,01399.76.64 (7.48)3.00
WBC (x10E3/uL)2411721,93299.46.86 (6.17)3.00

Table 10. Retained laboratory input types and available measurement-level summary statistics.

For each patient encounter associated with a CT scan, all ICD-9 and ICD-10 codes assigned during that encounter were collected and mapped to hierarchical phenotypes using PheWAS Phecode mapping[2], collapsing 16,553 distinct ICD codes to 1,692 phenotype categories.

Internal Test cohort – Stanford Main hospital. The Internal Test cohort contains 5,137 CT studies from 3,667 unique patients drawn from the same institution as the Discovery Site (821,920 CT slices, 405,802 lab entries, 6,134,090 diagnoses). These patients were held out from all stages of discovery and represent the first test of reproducibility under matched distributional conditions.

External Site #1. The External Site #1 cohort contains 25,803 CT studies from 17,224 patients at a community hospital outside Stanford (4,128,480 CT slices, 1,764,079 lab entries, 19,610,547 diagnoses). No data from this cohort was used at any stage of discovery or model training, and the shift in patient population and clinical practice makes it a more demanding test of generalization.

External Site #2 – Mayo Clinic. The Mayo Clinic cohort consists of patients across four Mayo Clinic hospitals, comprising 14,199 CT studies from 6,085 patients. No data from this cohort was used at any stage of discovery or model training, and as an out-of-state health system it represents the greatest institutional distance from the Discovery Site.

Dataset CharacteristicsDiscovery (Train + Val)Internal TestExternal Site #1External Site #2
n%n%n%n%
Accessions (Abdominal CTs)20,391–5,137–25,803–14,199–
Unique Patients14,652–3,667–17,224–6,085–
Age (mean)54.1 ± 18.9–54.5 ± 19.1–60.6 ± 18.9–62.1 ± 16.9–
Sex
Female11,40755.92,92757.014,03954.46,99449.3
Male8,98244.02,21043.011,76445.67,20550.7
Other20.0100.0000.0000.00
Smoking Status
Current Smoker1,4607.23536.91,3805.31,63311.5
Heavy Alcohol Use2,41911.966713.02,87511.12,32716.4
BMI (mean)26.9 ± 6.9–26.9 ± 6.9–27.4 ± 6.4–27.1 ± 2.7–
Laboratory values (6mo prior)
Lab values1,934,671–494,575–1,772,983–145,868–
Accessions with lab values19,84597.35,01197.524,71095.84,15529.3

Table 1. Table 1. Summary of dataset characteristics across the Discovery, Internal Test, External Site #1, and External Site #2 (Mayo Clinic) cohorts, including the number of unique patients and abdominal CTs, age, sex distribution, smoking status, heavy alcohol use, and BMI. Values are reported as counts (n) and percentages where applicable.

Foundation Models

Each data modality is encoded by a the most performant foundation model chosen at the time of the study. FM weights are frozen throughout all downstream analyses; the SAEs and association pipelines operate entirely on static embeddings produced by these models.

CT imaging For CT images, we use Merlin[14], a 3D inflated ResNet-152 pretrained on abdominal CT volumes paired with radiology reports and EHR diagnosis codes. Merlin processes each CT volume and produces a 2,048-dimensional embedding that serves as the input to all downstream concept extraction and association analyses.

CT radiology report For radiology report findings, we use Qwen3-Embedding-8B[21], a general-domain text embedding model selected as the top-performing model on the MTEB leaderboard at the time of this study[62]. Each findings section is encoded into a single 4,096-dimensional vector.

Laboratory values We use EHR Labs FM, an in-house masked autoencoder pretrained on laboratory histories from the Discovery Site. The encoder is a three-layer Feature Tokenizer Transformer with four attention heads and a hidden dimension of 768 that jointly encodes laboratory values, test types and measurement times. For each history, values are randomly masked at a rate sampled uniformly between 0 and 90%, with at least one value masked. Laboratory-specific prediction heads reconstruct the masked values using a mean squared error objective. Training uses AdamW with a learning rate of 10−410^{-4}, a batch size of 256 and 1,500 optimisation steps, with checkpoint selection based on validation reconstruction loss. For downstream analyses, the reconstruction heads are discarded and the encoder remains frozen. Its output embeddings are averaged over non-padding tokens to produce a single 768-dimensional vector.

Sparse Autoencoders

FM embeddings are dense and polysemantic: individual dimensions typically encode mixtures of unrelated clinical concepts simultaneously, making direct interpretation infeasible. Sparse autoencoders address this by learning a larger, overcomplete dictionary of directions in the embedding space, where each sample activates only a small subset of dictionary entries. This sparsity encourages each entry to correspond to a single, distinct clinical concept.

Architecture. For each modality, a separate SAE operates on the z-normalized FM embedding x~=(x−μ)/σ\tilde{x} = (x-\mu)/\sigma, where statistics are computed on the Discovery train set. Z-normalization is critical so that the SAE learns relative structure in the embedding space rather than the distance of all points from the origin. A linear encoder maps x~\tilde{x} to pre-activation scores over a dictionary of n=8,192n = 8{,}192 features,

z=x~Wenc+benc,Wenc∈Rd×n(1)z = \tilde{x}W_{\mathrm{enc}} + b_{\mathrm{enc}}, \qquad W_{\mathrm{enc}} \in\mathbb{R}^{d \times n} \tag*{(1)}

and Top-KK sparsity retains only the KK highest-scoring features, zeroing the rest to produce the sparse representation zsparse=TopK⁡(z,K)z_{\mathrm{sparse}} = \operatorname{TopK}(z,K). A linear decoder then reconstructs the embedding,

x^=zsparseWdec+bdec,Wdec∈Rn×d.(2)\hat{x} = z_{\mathrm{sparse}}W_{\mathrm{dec}} + b_{\mathrm{dec}}, \qquad W_{\mathrm{dec}} \in\mathbb{R}^{n \times d}. \tag*{(2)}

The decoder is initialized as the transpose of the encoder, and its columns are constrained to unit norm throughout training, with the gradient component parallel to each column removed after each step to prevent shrinkage. We ablate five sparsity levels, K∈{5,10,20,40,80}K \in\{5,10,20,40,80\}.

Matryoshka SAEs. Standard Top-KK SAEs tend to produce only fine-grained features, a phenomenon known as feature splitting, neglecting broader concepts higher in the clinical hierarchy. To address this, we additionally train Matryoshka SAEs[63], which impose nested reconstruction losses across an ordered set of prefix sizes M\mathcal{M},

LMatryoshka=1∣M∣∑m∈M∥x~−zsparse[:m]Wdec[:m,:]−bdec∥22.(3)\mathcal{L}_{\mathrm{Matryoshka}} = \frac{1}{|\mathcal{M}|} \sum_{m \in\mathcal{M}} \left\lVert\tilde{x} - z_{\mathrm{sparse}}[:m]W_{\mathrm{dec}}[:m,:] - b_{\mathrm{dec}} \right\rVert_{2}^{2}. \tag*{(3)}

Smaller prefixes are forced to encode coarser concepts while larger prefixes introduce progressively finer distinctions, producing a hierarchical concept space. We test a standard Top-KK SAE and two Matryoshka SAE configurations, with nested dictionary sizes M1={128,512,2048,8192}\mathcal{M}_1 = \{128,512,2048,8192\} and M2={256,1024,4096,16384}\mathcal{M}_2 = \{256,1024,4096,16384\}.

Training. SAEs are trained on the Discovery train set by minimizing MSE reconstruction loss, L=∥x~−x^∥22\mathcal{L} = \lVert\tilde{x} - \hat{x} \rVert_{2}^{2}, using the Adam optimizer with a learning rate of 5×10−45 \times10^{-4}, a batch size of 2,000, and a maximum of 1,000 steps with early stopping based on validation loss. One SAE is trained per modality.

Feature extraction and configuration selection. Once trained, SAE weights are frozen and sparse features are inferred on all datasets by passing each FM embedding through the pretrained SAE. Each sparse feature is then z-normalized by subtracting its mean and dividing by its standard deviation, both computed on the Discovery Site, providing a consistent basis for comparison across modalities and datasets. Based on the ablation results described in Supplementary Section A.3 and Figure 14, we select K=40K = 40 with the M={128,512,2048,8192}\mathcal{M} = \{128,512,2048,8192\} Matryoshka architecture as a single consistent configuration across all modalities. The resulting sparse feature vectors from each modality are concatenated to form a unified multimodal concept space, in which each sample is represented by a single sparse vector spanning imaging, report, and laboratory concepts.

Architectural ablation of the sparse autoencoder across 141 disease prediction tasks

Figure 14. Architectural ablation of the sparse autoencoder across 141 disease prediction tasks. Results are organized by modality, with panels a–c corresponding to CT imaging (Merlin), radiology reports (Qwen3-Embedding-8B), and laboratory values (Labs FM), respectively. Within each modality, results are organized by metric, showing sparse autoencoder reconstruction loss (MSE), R2R^2 between original and reconstructed foundation-model embeddings, and mean downstream AUC across the 141 disease prediction tasks. Reconstruction loss and R2R^2 are evaluated on the discovery training set and unseen external data, while downstream AUC assesses whether the Glyph concept space preserves the predictive performance of the underlying foundation models with minimal degradation. Across modalities, diminishing returns in reconstruction are observed beyond K=40K = 40, with R2R^2 improving and subsequently plateauing at this sparsity level.

Automated Interpretability

By design, sparse features activate for only a small subset of samples. This sparsity creates the pre-condition for interpretability: if a feature fires for a restricted, coherent cohort, it may encode a single identifiable clinical phenomenon. To convert features into named concepts at scale, we automate the interpretation process using a large language model (LLM), eliminating the need for manual inspection of thousands of features.

Generative interpretation. For each SAE feature, we identify the samples in the Discovery Site whose activation exceeds the 75th percentile of all positive activations, forming a high-activating cohort. We then present nhigh=20n_{\mathrm{high}} = 20 randomly sampled high-activating samples to an LLM, alongside nnon=20n_{\mathrm{non}} = 20 randomly sampled samples with zero or near-zero activation as a contrast reference. The contrast is essential: it grounds the interpretation in what is specific to the concept rather than what is common across the patient population. The LLM is instructed to reason via explicit chain-of-thought before producing a single-sentence summary of the form “This cohort is characterized by [interpretation].” We use Gemini 2.5 Pro via a PHI-safe GCP Vertex instance. Full prompts are provided in Supplementary Section A.4.

Discriminative validation. Validation serves two purposes: assessing whether the interpretation faithfully describes the feature, and testing whether this description generalises to held-out sets. For each concept, we sample nhigh=20n_{\mathrm{high}} = 20 and nnon=20n_{\mathrm{non}} = 20 from the test set, then crucially we shuffle the combined pool. The LLM receives the textual interpretation together with the shuffled sample data, is informed that exactly half the samples belong to the cohort, and must classify each sample as either belonging to the concept cohort (1) or not (0). Discrimination accuracy, calculated by the fraction of samples correctly classified, is computed on the Internal Test set as a measure of monosemanticity, and separately on each external cohort to assess cross-site and -institutional generalization.

Outcome selection

To maximize the potential for rediscoveries and new discoveries, we use an automated approach to select the highest signal outcome targets for exploration using the Discovery Site. We first map 16,553 unique ICD-9 and ICD-10 codes to 1,692 hierarchical phenotypes termed Phecodes[2]. From these we define a candidate pool of 1,278 outcomes with support in our data. Studies were labeled as progressors if the first recorded diagnosis of an outcome occurred 1–4 years after the study date. Controls were selected as having no recorded diagnosis of that outcome, and at least four years of follow-up. All other studies, including those with a recorded diagnosis of the outcome before the CT date, were excluded from that outcome’s analysis. Following[64], we then compute on the Discovery Site the aggregate rank (by reciprocal rank fusion) of the ability of each modality and model combination to predict each outcome. The 150 highest-ranked outcomes were then manually reviewed, and nine excluded as not interpretable as clinical disease endpoints. These comprised non-specific catch-all Phecodes (for example, other disorders of the liver), abnormal-finding and administrative codes (for example, other abnormal blood chemistry), and device or procedure status codes (for example, ileostomy status). This left the 141 outcomes used in this study, covering conditions across oncological, cardiovascular, metabolic, and acute disease categories. To see the performance of each FM on this set, see Figure 7.

Four-panel plots showing embedding AUC, sparse autoencoder reconstruction quality, sparse concept-space outcome AUC, and effect sizes across cohorts and modalities

Figure 7. a, Outcomes were selected by ranking Phecode prediction AUCs on the Discovery Site validation set. After excluding non-disease codes, 141 of the top 150 outcomes were retained. Without refitting, imaging probes transferred with minimal loss across cohorts, whereas report performance declined at External Site #2. b, SAEs fitted at the Discovery Site were applied unchanged across cohorts. Reconstruction remained high for imaging and labs but declined steadily for reports. c, Replacing dense embeddings with sparse concepts preserved most predictive information, with less loss for imaging than reports. d, Median odds ratios of significant associations followed a similar per-site trend to the predictive performance, maintaining high OR for imaging- and labs-derived concepts and steadily dropping for radiology report-derived concepts.

Association study

Taking inspiration from the multi-stage design of genome-wide association studies[2], we estimate the association between each Glyph concept and each future disease outcome as a per-standard-deviation odds ratio (OR), enabling direct comparison across modalities whose activations operate on different scales.

Standardization. Concept activations are z-normalised using a per-site mean and standard deviation, so that each OR reflects the multiplicative change in outcome odds per one standard deviation increase in concept activation.

OR computation. For each concept-outcome pair, we fit an unadjusted logistic regression model on the Discovery Site,

logit⁡(P(Y=1))=α+βc~,(4)\operatorname{logit}(P(Y = 1)) = \alpha+ \beta\tilde{c}, \tag*{(4)}

where c~\tilde{c} is the standardised concept activation. The per-standard-deviation OR is defined as exp⁡(β)\exp(\beta), and 95% confidence intervals are reported as exp⁡(β±1.96⋅SE)\exp(\beta\pm1.96 \cdot\mathrm{SE}).

Statistical significance. We assess significance using the Wald pp-value derived from the asymptotic covariance matrix, computed as the Moore-Penrose pseudo-inverse of the Fisher information matrix,

I(β)=XTWX,(5)I(\beta) = X^{T} W X, \tag*{(5)}

where WW is a diagonal matrix with entries pi(1−pi)p_i(1 - p_i). The z-statistic z=β/SEz = \beta/ \mathrm{SE} yields a two-tailed pp-value. Each CT study was analyzed as a separate observation to capture the patient’s clinical state at the time of imaging, with independence assumed across studies.

Discovery filter. To concentrate downstream analysis on statistically supported, directionally consistent associations, we apply a Bonferroni correction across all concept-outcome pairs, setting the threshold at pBonferroni=0.05/Npairs=8.28×10−8p_{\mathrm{Bonferroni}} = 0.05/N_{\mathrm{pairs}} = 8.28 \times10^{-8}. We additionally require OR>1.0\mathrm{OR} > 1.0 to restrict attention to predictive rather than protective associations. Associations satisfying both criteria on the Discovery Site form the set of promising associations carried forward into all subsequent analyses.

Reproducibility filter

Having identified significant associations on the Discovery Site, we assess whether they represent real, generalizable phenomena by re-evaluating them on held-out cohorts. Concept activations on held-out data are computed by applying the same frozen SAE weights without modification, requiring no manual labeling or recalibration of new samples.

For each promising association, we re-compute the same OR on each held-out cohort and apply the Benjamini-Hochberg (BH) procedure at a 5% false discovery rate across the set of re-evaluated associations. An association is considered reproducible if the BH-adjusted pp-value is significant and directionality is maintained (OR > 1.0).

Confounder adjustment

A predictive association between a concept and an outcome may be driven by a shared confounder rather than by a direct relationship. To assess whether associations persist after adjustment for measured confounders, we apply double machine learning (DML) with cross-fitting[65]. This allows us to account for nonlinear relationships and interactions among measured confounders that logistic regression with linear covariate terms may miss. The estimator uses a Neyman-orthogonal estimating equation, whose expectation is locally insensitive to first-order errors in the nuisance functions at their true values. Orthogonality therefore reduces sensitivity to nuisance estimation errors, while cross-fitting limits overfitting bias.

Because our outcomes are binary, we use a logistic partially linear model,

logit⁡P(Y=1∣T,W)=θT+r(W),(6)\operatorname{logit} P(Y = 1 \mid T,W) = \theta T + r(W), \tag*{(6)}

for concept activation TT, binary outcome YY, and confounder set WW, where θ\theta is the adjusted log-odds ratio of interest and r(W)r(W) is the baseline log-odds function. We estimate the nuisance functions using XGBoost with five-fold outer and five-fold inner cross-fitting, so that each observation is evaluated using nuisance models fitted without that observation. All nuisance construction, including inner cross-fitting, is confined to the corresponding outer training fold. Models use 100 trees with maximum depth four. Each CT study is treated as an independent observation, and activations and continuous covariates are standardized using Discovery training statistics.

Baseline log-odds (rr). We estimate r(W)r(W) using the Full Model Refitting procedure of Liu et al.[66]. Within each outer training fold, a joint model of YY on (T,W)(T,W) produces inner out-of-fold predictions of the full conditional probability. Their logits serve as an intermediate target for estimating a preliminary exposure coefficient and the baseline function r(W)r(W), accounting for the concept’s contribution to the full predictor.

Residualized activation (T~\widetilde{T}). For the sigmoid-based score used here, orthogonality requires a variance-weighted conditional mean of the concept activation[67],

m(W)=E[v(T,W)T∣W]E[v(T,W)∣W],v(T,W)=μ(T,W){1−μ(T,W)},(7)m(W) = \frac{\mathbb{E}[v(T,W)T \mid W]}{\mathbb{E}[v(T,W) \mid W]}, \qquad v(T,W) = \mu(T,W)\{1-\mu(T,W)\}, \tag*{(7)}

where μ(T,W)=P(Y=1∣T,W)\mu(T,W) = P(Y = 1 \mid T,W). We estimate m(W)m(W) by regressing TT on WW, with variance weights calculated from the inner out-of-fold predictions of the full outcome model. The fitted weighted regressor is applied to the outer held-out fold to obtain residuals T~i=Ti−m^(Wi)\widetilde{T}_i = T_i-\widehat{m}(W_i).

The adjusted effect θ^\widehat{\theta} is then obtained by solving the orthogonal score equation

∑i(Yi−σ(θTi+r^(Wi)))T~i=0,(8)\sum_i (Y_i-\sigma(\theta T_i+\widehat{r}(W_i)))\widetilde{T}_i=0, \tag*{(8)}

by root-finding, where σ\sigma is the logistic sigmoid function. Standard errors are calculated using the corresponding sandwich estimator. We report the adjusted OR as exp⁡(θ^)\exp(\widehat{\theta}), with a 95% Wald confidence interval and a two-sided Wald pp-value testing θ=0\theta= 0.

Confounder selection. For each association, confounders are drawn from two sources. The six universal clinical covariates are patient age, sex, BMI, alcohol use, smoking status and race, with race represented by indicator variables. In addition, up to ten conceptual confounders are automatically selected per outcome from the significant Glyph associations ranked by unadjusted OR. These are selected independently within each dataset to account for differences between institutions and populations. This also provides the opportunity to test confounders from other modalities, though we disallow confounders from the radiology report to interfere with those from the image, and vice versa, as they were derived from the same data source. To limit overadjustment for overlapping representations of the tested concept and potential mediators, we applied a correlation threshold. We selected an absolute Pearson correlation threshold of 0.35 through qualitative inspection to balance excluding likely downstream consequences against retaining potential confounders, and excluded candidates at or above this threshold. For example, when testing ascites, a broader concept combining ascites, anasarca, and pleural effusion was excluded from the adjustment set (r=0.61). Similarly, when testing aortic atherosclerosis, a broader concept representing abdominopelvic atherosclerosis was excluded (r=0.41). However, related atherosclerosis and aortic calcification concepts remained eligible for adjustment when analyzing aortic atherosclerosis (r=0.31–0.34), whereas upstream liver pathology was excluded when analyzing splenomegaly (r=0.46). The threshold therefore served as a heuristic, since correlation alone cannot distinguish overlapping representations, mediators, and confounders[61].

Confounder filter. Similar to the Reproducibility filter, an association is considered to survive confounder adjustment if the BH-adjusted pp-value (at 5% FDR, applied to the full set of Bonferroni-significant associations) remains significant and the adjusted OR > 1.0. Adjustment was evaluated for the 4,463 significant Discovery associations in each cohort, except at External Site #2, where only the 3,109 imaging and radiology report associations were evaluated.

CohortCorrectionThreshold
Discovery (Train)Benjamini–Hochberg (5% FDR)2.24×10−22.24 \times10^{-2}
Internal TestBenjamini–Hochberg (5% FDR)1.06×10−21.06 \times10^{-2}
External Site #1Benjamini–Hochberg (5% FDR)9.50×10−39.50 \times10^{-3}
External Site #2Benjamini–Hochberg (5% FDR)1.46×10−21.46 \times10^{-2}

Table 3. Significance thresholds for the confounder filter. Unlike the association filters (Table 2), every cohort, including Discovery, uses a Benjamini–Hochberg threshold (5% FDR) computed over the confounder-adjusted pp-values of the promising associations tested on that cohort.

CohortCorrectionThreshold
Discovery (Train)Bonferroni8.28×10−88.28 \times10^{-8}
Internal TestBenjamini–Hochberg (5% FDR)3.46×10−23.46 \times10^{-2}
External Site #1Benjamini–Hochberg (5% FDR)3.72×10−23.72 \times10^{-2}
External Site #2Benjamini–Hochberg (5% FDR)2.44×10−22.44 \times10^{-2}

Table 2. Significance thresholds for the association filters. The Discovery Site uses a Bonferroni-corrected threshold across all tested concept-outcome pairs; each held-out cohort uses its own Benjamini–Hochberg threshold (5% FDR) computed over the promising associations re-evaluated on that cohort.

Inter-concept causal graph. To recover directed relationships among concepts, we learned a single sparse directed acyclic graph (DAG) over all documented and undocumented concepts together with a fixed set of universal confounders (age, sex, smoking status, heavy alcohol use, and BMI) as additional nodes, allowing the structure learner to attribute shared demographic and lifestyle variation to edges from the confounder nodes rather than to spurious concept–concept edges (see Figure 11). We standardized all node activations and fit a linear structural equation model X=XW+εX = XW + \varepsilon with an ℓ1\ell_1 sparsity penalty on the weighted adjacency WW, subject to a continuous acyclicity constraint. We used DAGMA[68], which enforces acyclicity through a log-determinant characterization and a continuation schedule that progressively tightens the constraint toward an exact DAG, building on the continuous-optimization formulation of NOTEARS[69]. Entries of WW below a magnitude threshold were set to zero, and the surviving weights define the directed edges. Whereas the primary concept-outcome analyses use double machine learning (DML), which is confounder-robust but requires fitting separate nuisance models per target[65], the interconcept graph involves on the order of d2d^2 candidate relationships across all dd concepts, making per-pair DML computationally prohibitive; we therefore learn the full structure, confounders included, in a single continuous optimization and reserve DML for the smaller, hypothesis-driven outcome analyses in the main text.

Figure 11. Confounder-adjusted inter-concept associations. a The full adjusted network. Edges denote directed dependencies from the fitted sparse SEM, retained where the standardized coefficient exceeds ∣w∣≥0.1|w| \ge0.1; edge width scales with ∣w∣|w| and colour denotes sign (red positive, blue negative). Dashed edges involve at least one undocumented concept. b Concepts within two degrees of separation from liver cirrhosis (R85), a disease hub known to link to splenomegaly (I65) and portal hypertension (R309). c Patient age shows adjusted associations with many concepts, particularly cardiovascular disease and spinal degeneration (R1). d In this abdominal CT cohort, labs indicating cholestasis (L108) form the major bridge between modalities, connecting abnormal lab states, such as elevated alkaline phosphatase (L33) and cytopenia (L127), to structural imaging phenotypes such as splenomegaly (I65) and intra-abdominal fluid (I105).

This procedure is an approximate analog to confounder adjustment rather than a guarantee of it: a concept-concept edge is adjusted for a confounder only if that confounder’s weight survives thresholding, and edge directions reflect the acyclicity constraint and model fit rather than verified causal mechanisms. Unexpected edges may thus reflect genuine population-specific structure, but could also stem from residual confounding or artifacts of thresholding and optimization.

Comparison with the clinical literature

To assess whether Glyph identifies prognostic information beyond established clinical markers, we compared its concepts with risk markers reported in the literature. Here, we describe the literature search and screening, reconstruction of published markers, and methods used to compare their associations and predictive value with Glyph. For each outcome, PubMed and Paperclip queries combined endpoint names and synonyms with an input scope defining eligible abdominal imaging findings, laboratory measurements and clinical risk factors. Publication in a leading relevant journal or at least 50 citations was required, retaining 83,649 outcome-specific records. Title screening advanced 14,801 records to abstract review, of which 4,748 underwent further assessment of their measurements, endpoints and local feasibility. Original methods and tables were consulted where marker definitions were unclear. At least one reported marker could be reconstructed for 1,822 records, before deduplicating repeated measurement families and selecting 293 representative studies. A later report-based COPD marker was included outside the initial impact criterion, bringing the total to 294. Markers were reconstructed using published thresholds, transformations and composite-score definitions, with approximate measurements labelled as proxies. Counting markers separately for each outcome, the final 412 entries comprised 36 explicit multi-input scores, ratios or biomarker rules (8.7%), 31 BMI or estimated kidney-function measures (7.5%), and 345 single measurements, findings or recorded conditions (83.7%). Thus, 67 markers (16.3%) combined multiple inputs under the broader definition including BMI and estimated kidney function. Queries, source papers, screening decisions and fitted definitions were retained for each outcome.

All prediction models used the same outcome-eligible CT studies and included age, sex and recorded race. Markers required at least 100 observed Discovery measurements and, for binary rules, at least 25 observations in each category. Missing measurements were replaced with the Discovery training mean and accompanied by a missingness indicator. Ridge logistic regression coefficients and preprocessing were learned on Discovery training data, with regularization selected by validation AUROC and the selected model applied without refitting. Models were fitted to convergence before validation AUROC was used to select regularization. Patients did not cross fitting, tuning or evaluation splits. We repeated fitting on five bootstrap samples of independent Discovery training studies, keeping the concept sets, validation set and held-out cohorts fixed. Reported AUROCs are the arithmetic means of the five fitted-model AUROCs, rather than AUROCs of averaged predictions. If an augmented model did not improve validation AUROC, the corresponding baseline was retained. Evaluation used the pooled Internal Test and External Site #1 cohorts. External Site #2 was not included because direct access to the clinical data needed to reconstruct the literature markers was unavailable at the time of the study. AUROC confidence intervals and two-sided comparisons used paired study-level influence functions averaged across the five fits. Across-outcome means gave outcomes equal weight, retaining shared-study covariance across models and outcomes; relative-gain uncertainty used the delta method. Inference conditions on the fitted models and selected markers and does not include training or selection uncertainty.

Individual-marker associations were estimated using the orthogonal logistic DML estimator described above, adjusted for age, sex and recorded race, with five-fold outer and inner cross-fitting and independent-study sandwich uncertainty. Exposures, including binary rules, were standardized using Discovery training statistics. Ten Glyph concepts per outcome were curated using Discovery clinically adjusted rankings and clinical interpretations, excluding concepts primarily reflecting intervention or workflow. For the size-matched OR and reproduction comparisons, up to ten literature markers per outcome were selected using the adjusted Discovery-site OR ranking and fixed across held-out cohorts. This selection did not restrict the literature baseline used in the prediction models. There were 194 selected literature markers across outcomes. Reproduction required BH-adjusted q<0.05q < 0.05 in the pooled held-out cohort and the same effect direction as Discovery, with BH applied across the 768 Glyph, literature and matched-Phecode marker–outcome associations. Differences in reproduction rates were tested using two-sided Fisher exact tests, treating reproduction of each association as independent. Median OR comparisons gave outcomes equal weight and used 20,000 shared-study multiplier bootstrap samples to preserve dependence across markers and outcomes, with marker sets and nuisance fits fixed. The minimum attainable bootstrap pp-value was 1/20,0011/20{,}001. Benjamini–Hochberg correction was applied to the six displayed held-out OR comparisons, with separate families for the three Discovery and three pooled comparisons. KDEs used all supported outcomes within each marker family; pairwise median comparisons used the outcomes supported by both families.

Code and Data Availability

Analysis and reproduction code is available at https://github.com/StanfordMIMI/Glyph-research. Released concept activations and interpretations are hosted at https://huggingface.co/datasets/RobbieHolland/Glyph-research, subject to access restrictions. Pretrained SAE checkpoints and normalization statistics are available at https://huggingface.co/RobbieHolland/Glyph-research. External Site #2 (Mayo Clinic) study-level data and clinical covariates required for confounder-adjusted refitting are not distributed.

Open source

The pipeline described above, spanning embedding extraction, SAE training, automated interpretation, hypothesis search, and visualization, is available through the open-source library Glyph (https://pypi.org/project/glyph-discovery/). Users can learn new sparse concepts from their own data or apply the pretrained dictionaries released with this paper using the corresponding foundation models. Foundation models can be selected through built-in presets or supported Hugging Face model identifiers; custom models can be connected by specifying their preprocessing and embedding output. Listing 1 illustrates the Python API for learning new concepts, Listing 2 shows how to measure pretrained concepts, and Listing 3 demonstrates running the workflow through a coding agent.

When training new dictionaries, fit evaluates SAE configurations and favors sparse representations that achieve a validation reconstruction R2R^2 of at least 0.85, allowing modest increases in the number of active concepts for further improvements in reconstruction. A warning is issued if this threshold is not reached. hypothesis_search runs the association study described in § 5.6, with optional confounder adjustment. visualize produces an interactive concept-space explorer with association statistics and Bonferroni-significant concepts highlighted. Additional agent skills support literature searches, comparisons against published markers, and concise visual reports. Automated interpretation is optional and requires configured access to a supported language model.

from glyph import Glyph

glyph = Glyph(config_path='config', config_overrides=[
    'data.metadata=/path/to/metadata.csv',
    'data.input_cols=[image_path, report_path]',
    'data.outcome_cols=[pneumonia]',
    'data.embedding_map={image_path: dinov3, report_path: qwen3-embedding-0.6b}',
])

glyph.fit()                 # selects a sparse SAE using validation R^2
glyph.hypothesis_search()   # association testing over sparse concepts
glyph.visualize(outcome='pneumonia', split='test')

Listing 1. Learning new concepts with Glyph. The example uses DINOv3 for images and Qwen3-Embedding-0.6B for report text. The embedding_map connects each metadata input column to its encoder. fit selects a sparse SAE configuration using validation reconstruction performance, warning if R2<0.85R^2 < 0.85. hypothesis_search tests sparse concepts against the specified outcomes, and visualize displays the interactive concept space and association results.

from glyph import Glyph

# 1. EXTRACT A SINGLE PUBLISHED CONCEPT ON YOUR DATA
ct = Glyph.from_pretrained('ct-merlin')
i87 = ct.measure('I87', metadata='ct_metadata.csv')

# Optionally restrict extraction to one study.
i87 = ct.measure(
    'I87', metadata='ct_metadata.csv', sample_id='scan_001'
)

# 2. EXTRACT THE FULL PRETRAINED CONCEPT SPACE ON YOUR DATA
# Example: apply the paper's entire laboratory dictionary.
labs = Glyph.from_pretrained('labs-loinc')
concepts = labs.measure(metadata='labs_metadata.csv')
labs.visualize()

# Select a concept from the already measured space.
l127 = concepts['L127']

Listing 2. Extracting published concepts on new data. Single-concept extraction. Specify a concept code to measure one published concept across a cohort or for a selected study, illustrated with imaging concept I87. Full-space extraction. Omit the concept code to apply the paper’s entire pretrained dictionary for a modality to the supplied cohort, illustrated with the laboratory dictionary. The resulting study-by-concept table can be visualized, and individual concepts such as L127 can be selected directly. The metadata supplies study identifiers and input paths; Glyph computes the matching foundation-model embeddings and concept activations on the user’s data while preserving the paper’s concept indices. Neither outcome labels nor SAE retraining is required for concept extraction and exploration; labels are needed for association testing and significance highlighting.

$ claude

> Install Glyph from github.com/RobbieHolland/Glyph and its bundled skills
  for Claude Code, then run the pipeline on my cohort.
  metadata:  <path/to/metadata.csv>
  inputs:    <input_col_1, input_col_2, ...>
  outcomes:  <outcome_1, outcome_2, ...>
  encoders:  <input_col_1: model_name_or_HF_ID, input_col_2: ...>
  Report the reconstruction R^2, draw the concept map, and list the concepts
  most associated with each outcome, including odds ratios and which survive
  Bonferroni correction. Use the literature skill to identify established
  markers with supporting sources, and produce a brief visual report.

> Now apply the paper's pretrained concepts instead of fitting a new SAE.
  Use the appropriate bundle: <ct-merlin | findings-qwen3 | labs-loinc>.
  Tell me which associations reproduce in my cohort.

Listing 3. Running Glyph from a coding agent. Bundled skills guide Claude Code or Codex through configuration, analysis, literature comparisons, visualization, and reporting. The agent installs the package, configures the inputs and outcomes, runs the requested analyses, and summarizes the results. The second prompt applies the paper’s pretrained dictionaries and corresponding foundation models to a new cohort without retraining the SAE.

The Glyph MCP returns the top 10 predictors of myocardial infarction, including laboratory concepts and imaging findings.

Figure 8. The Glyph MCP brings a new capability; the ability to immediately perform, user-specified requests about high-dimensional unstructured data, such as imaging. Here, Claude Code returns the most highly associated concepts for future risk of myocardial infarction, the majority of which are labs concepts. Responses are grounded by the Glyph layer.

Glyph generates individualised risk explanations for a given outcome

Figure 9. Glyph generates individualised risk explanations for a given outcome. For a query patient (AC42403ee, 74.9% predicted CKD risk, 93.6th percentile), the system surfaces the concepts most responsible for that prediction alongside their attribution scores, activation levels, and population rarity. Cross-referencing against the top population-level risk features reveals which epidemiologically validated signals are active in this patient, grounding the individual explanation in cohort-level evidence.

Glyph concept layer showing a patients-like-me scenario in which the most similar patient to a query is found and the similarity is explained in terms of shared concepts and the rarity of their co-occurrence

Figure 10. The Glyph concept layer is flexible to other forms of discovery beyond prognosis. Here, Glyph performs a patients-like-me scenario where the most similar patient to a query is found. Their similarity is then explained in terms of shared concepts, and the rarity of such a co-occurrence at their respective levels of activation.

Supplementary material

Interpretation of Glyph concepts and validation across sites

We next assess the interpretability of Glyph’s discovered features, using a contrastive variant of AutoInterp[20] and interpretation testing outlined in § 5.4. To make analysis more token-efficient, we only interpret high-risk concepts, defined by those with at least one significant association to future disease (see § 3.2). Prior work highlighted the difficulty of interpreting radiological SAE features at an F1 threshold of 0.75[70], motivating our choice of a discrimination accuracy threshold of ≥75%\ge75\% for considering a concept interpretable.

Using this method and threshold, 88.6% of radiology report concepts were interpretable internally, and 76.6% and 86.7% of interpretations successfully generalized to External Sites #1 and #2, respectively (see Figure 12c). This high rate was expected, as these concepts are derived from embeddings of text that explicitly describe clinical findings. Interpreting concepts derived from images represents a greater challenge, as Merlin likely relies on its own internal set of concepts that may not correspond to anything radiologists routinely or consistently document in reports. Remarkably, 52.5% of Merlin image concepts were broadly interpretable by corresponding radiology reports (see Figure 12d–Figure 12f). Labs concepts were more automatically interpretable than imaging (see Figure 12g–Figure 12i), and captured physiological states such as monocytosis.

Investigating image concepts not interpretable by AutoInterp, we find many still carry identifiable meaning, when viewed directly in images by radiologists. For example, we noticed I76 was not interpretable via AutoInterp, but was among the top 1% closest imaging-derived markers to a broad swath of laboratory concepts, spanning inflammation, cholesterol, and protein levels (annotated in 2b). Strikingly, despite not being routinely documented, I76 was significantly associated with the onset of infection, sepsis, septicemia, and wasting. Prompted by these results, radiologists visually inspected images retrieved by this concept, finding it activates in patients with severely distended, fluid-filled bowel loops (see Figure 12d–Figure 12e).

Glyph provided complementary associations to pre-coded phenotypes

Having established that Glyph concepts yield significant and reproducible markers of future disease risk, we asked whether pre-coded PheWAS for automated clinical discovery would have identified the same discoveries[2]. We treated Phecodes as exposures and applied the same discovery pipeline, including multiple-test corrected Discovery and Reproducibility filters, searching among the same 141 outcomes where Glyph has the maximum expected yield.

Aggregate binary-exposure results are presented in Figure 13a–Figure 13b. Recent Phecodes recorded from −180 to +30 days relative to the scan yielded 7,843 promising associations, with significantly lower replication than Glyph concepts at Internal Test (47.2% vs. 69.3%, p=9.77×10−4p = 9.77 \times10^{-4}) and External Site #1 (59.9% vs. 74.3%, p=1.95×10−3p = 1.95 \times10^{-3}) Using each patient’s entire recorded coding history increased the Phecode yield to 15,274 promising associations, approximately 3.4× the number identified by Glyph, of which 49.1% and 63.9% reproduced in the two held-out cohorts. The median Discovery–site effect was smaller for both Phecode arms than for Glyph concepts (OR 1.29 for Recent Phecodes and 1.33 for all-history Phecodes, versus 1.45 for Glyph). The greater yield reflects the broader scope of ICD codes, which span the full medical record, whereas Glyph concepts were limited to phenomena observable in abdominal CT images, and laboratory panels. Comparable yield may therefore be possible as Glyph expands to additional modalities and examinations.

We next assessed which Glyph associations could have been identified from codes alone (see Figure 13c). For previously reported associations, we identified if there was a matching Phecode and compared their associations directly. Directly matched associations were often recoverable from Phecodes. For example, Phecode 571.51 (cirrhosis of the liver) reproducibly predicted primary liver cancer (OR 2.03–2.25), independently corroborating the association identified for R85 (cirrhotic liver morphology and varices; OR 1.64–2.55). Across the majority of comparisons with a matching Phecode, the Glyph concept had a larger odds ratio. Some matched Phecodes did not pass the Discovery filter, and some additional Glyph associations had no corresponding Phecode. Overall, Phecodes identified more associations across the full medical record, while Glyph concepts yielded larger median effects, reproduced more frequently,

Architectural ablations

ApproachExposuresPairsP cutoffSignificantRateMedian OR (IQR)Internal TestExternal #1
Recent phecodes [−180, +30 d]1,226172,8662.89e−077,8434.54%1.29 (1.23–1.38)3,701 (47.2%)4,700 (59.9%)
All-history phecodes [−∞, +30 d]1,226172,8662.89e−0715,2748.84%1.33 (1.25–1.44)7,499 (49.1%)9,765 (63.9%)
Glyph concepts [−180, 0 d]4,285601,0508.32e−084,4660.74%1.45 (1.34–1.60)3,094 (69.3%)3,316 (74.2%)
Discovery SiteInternal TestExternal Site #1External Site #2
Effect-size distribution for Glyph and the phecode baselines
Effect size of discovery-significant associations, by discovery approach
Glyph conceptsRecent phecodesAll-history phecodes
Per-SD odds ratio
1.291.331.45
1.201.171.38
1.181.151.29
1.19
Fraction of associations
dashed line = median OR
phecode arms not run
labs concepts not evaluated
OutcomeGlyph concept [−180, 0 d]ORrepAny prior phecode [−∞, +30 d]ORrepRatioAdvantage
Primary liver cancerR85 Cirrhotic liver morphology and varices2.45***T1571.51 Cirrhosis of liver without mention of alcohol2.25***T11.09×✓
Portal hypertensionI21 Ascites3.17***T1572.0 Ascites (non malignant)1.78***T11.78×✓
Liver transplantL108 Cholestasis — elevated ALP or bilirubin2.88***T1277.4 Disorders of bilirubin excretion1.23 n.s.—2.35×✓✓
Chronic renal failureL84 Azotemia — elevated creatinine or urea2.02***T1790.6 Other abnormal blood chemistry1.48***−11.37×✓
Chronic renal failureI28 Chronic or end-stage disease findings1.92***T1— no coded equivalent— n.s.——✓✓✓
Myocardial infarctionR21 Coronary artery / cardiac valve calcification1.33***T1411.4 Coronary atherosclerosis1.34***T10.99×=
Congestive heart failureI26 Atherosclerotic disease of the aorta1.88***T1440.9 Atherosclerosis of aorta1.34***T11.41×✓
Alzheimer’s diseaseI15 Moderate–severe abdominopelvic atherosclerosis1.87***T1440.9 Atherosclerosis of aorta1.19 n.s.—1.57×✓✓
Ischaemic heart diseaseL68 Elevated neutrophils or monocytes1.33***T1288.2 Elevated white blood cell count1.30***−11.02×=
Gram-positive septicaemiaI76 Severely distended bowel with fluid (un documented)1.93***T1560.1 Paralytic ileus1.77***−11.09×✓
Septic shockR64 Metastatic disease1.45*T1198.0 Secondary malignant neoplasm1.23 n.s.—1.18×✓✓
Heart failure, preserved EFI26 Atherosclerotic disease of the aorta1.87***T1401.1 Essential hypertension (alternative)2.71***T10.69×✗

concept OR higher · ✗ code OR higher · ✓✓ code did not pass the Discovery filter · ✓✓✓ no coded equivalent · = within 5%. rep: T Internal Test, 1 External Site #1.

and produced stronger associations within directly matched phenotypes. Phecodes therefore capture information beyond the modalities available to Glyph, while Glyph concepts extract more predictive signal from the modalities it does observe.

Architectural ablations

The quality of Glyph’s concept space, and hence the sensitivity of all downstream discovery, depended directly on the architecture of the sparse autoencoder. Two parameters governed standard Top-KK SAEs: sparsity KK, which set the number of features active per sample, and dictionary size NN, the maximum number of distinct features available. Standard SAEs, however, have a known tendency known as feature splitting to recover only fine-grained concepts while neglecting broader phenomena higher in the conceptual hierarchy. The Matryoshka SAE addressed this by organising features into nested groups of increasing specificity, simultaneously encoding concepts at multiple scales. To identify the architecture best suited to Glyph’s multimodal clinical data, we systematically evaluated a standard Top-KK SAE alongside two Matryoshka variants with nested dictionary sizes [128, 512, 2048, 8192] and [256, 1024, 4096, 16384], across five sparsity levels (K∈{5,10,20,40,80}K \in\{5, 10, 20, 40, 80\}) and all three modalities.

Across all modalities, increasing KK was the dominant factor for improving reconstruction quality, reducing MSE and raising R2R^2 between original and reconstructed FM embeddings (Figure 14a–Figure 14b). All modalities exhibited diminishing returns beyond K=40K = 40, beyond which additional active features yielded marginal improvements while increasing computational cost. The number of alive features in the dictionary also grew with KK, reflecting the richer encoding capacity afforded by denser activations. Modality differences in the required number of features broadly tracked embedding dimensionality: radiology reports, embedded with the 4096-dimensional Qwen3-Embedding-8B, required more active features than CT images via Merlin (2048-dimensional), which in turn required more than laboratory values via the Labs FM (768-dimensional).

Matryoshka SAEs improved generalisation to external data relative to the standard Top-KK baseline, an advantage more clearly visible on the External Site #1 cohort than on the internal discovery set. The SAE trained on Discovery data maintained strong performance across all test sets, though with a modest drop on external cohorts that mirrored the reduced replication rates in § 3.2 and § 3.3, and may partly reflect this distributional shift. No meaningful benefit was observed between the two Matryoshka dictionary sizes, and we accordingly selected the Top-K=40K = 40 Matryoshka SAE with nested sizes [128, 512, 2048, 8192] for all analyses, balancing concept quality, generalisation, and computational cost, and achieved over 90% reconstruction of the original embeddings (R2>0.90R^2 > 0.90) across all modalities and FMs.

Beyond reconstruction fidelity, a key requirement was that the sparsified concept space preserved the predictive power of the underlying foundation models, confirming that the interpretability step did not discard clinically relevant signal (Figure 14c). Across modalities, Glyph concepts closely matched FM predictive performance, with particularly tight agreement for CT images (AUC 0.73 vs. 0.73 on the internal test cohort, 0.71 vs. 0.70 on External Site #1) and laboratory values (0.73 vs. 0.71; 0.66 vs. 0.65). Radiology report concepts showed a modestly larger gap (0.72 vs. 0.63; 0.68 vs. 0.62), consistent with the greater information density of the high-dimensional Qwen3-Embedding-8B text embedding space. Overall, predictive performance was largely preserved for imaging and laboratory concepts, with larger reductions for radiology report concepts.

Lastly, we assessed whether the derived concepts were objectively interpretable to an independent automated judge (Figure 14d). A self-reported interpretation could trivially appear coherent, so we instead posed interpretability as a discrimination task: a second LLM, given only the text interpretation, must classify unseen patients as high- or non-activating for that concept. Using the AutoInterp pipeline (see Methods, § 5.4), interpretations generated on the discovery set were evaluated on unseen patients without modification, testing whether each description accurately characterised new activations of the same concept. Radiology report concepts were highly interpretable, with 86.6% achieving an interpretability score above 75% on the internal test cohort and 76.6% on External Site #1. Laboratory concepts were competitive, with 77.6% achieving above 75% interpretability. CT image concepts were the least interpretable (52.5% on test, 45.5% on External Site #1), plausibly because sparse image features can capture visual structure not routinely documented in paired radiology reports. Overall, we concluded that Glyph produced a concept space that is broadly intelligible across all three clinical modalities.

AutoInterp prompts

The following prompt is populated with radiology reports findings and lab histories, before being passed to the LLM.

=== Task ===:
You are an expert radiologist specializing in abdominal CT scans and lab values. A cohort of patients were selected according to some shared characteristic specific to this cohort. You are provided radiology reports, lab histories, or both, and your task is to analyze them in order determine what are the shared feature(s). Radiology reports and lab histories may be filtered to the words, phrases or recordings most relevant to the shared characteristic to faciliate its identification. The importance level of those words, phrases or recordings should be considered in the analysis.

While the patient data will contain many different pathologies, there should be some undercurrent or one or two features which run through >80% of them. These characteristics should be positive, and somewhat unusual to be among so many reporst from different patients at once. Things like 'no evidence of bowel obstruction' is so common it is not useful. The explanation should be general enough to apply to the highest number of reports in this cohort, but not so general that it is no longer specific to this cohort, and would apply to many of the reference patients from the larger dataset.

To arrive at the most accurate and precise explanation of what this cohort is characterized by, you must engage in explicit chain of thought reasoning. Begin by thoroughly examining all provided data, noting any patterns or commonalities. Pay close attention to recurring terminology, described anatomical structures, and especially consistent pathological findings. Consider how these elements might interrelate to form a singular, distinctive feature that the cohort could be characterized by.

As you progress through your analysis, verbalize your thought process. Explain each step of your reasoning, from initial observations to intermediate conclusions, and finally to your overall assessment. This chain of thought approach will help ensure a comprehensive and well - reasoned final explanation.

After this detailed analytical process, formulate a single, specific explanation of what the cohort is characterized by. Focus on specific features or combinations of features, using 'and' to connect multiple elements if necessary and 'or' if there are multiple possibilities. Refrain from explaining the pathology itself (e.g., avoid statements like "This feature represents X, which is defined as ...").

Below are the patient data, listed in order of how strongly they belong to the cohort (keeping in mind that all of them belong strongly compared to the average patient in the larger dataset):

For reference, here are examples of patients randomly sampled from the much larger dataset that do not belong to this cohort:

{reference_patients}

Finally, the data from the cohort characterized by the feature in question:

{patient_data}

=== Format ===:
It is crucial that you present your final explanation in the following format, started with an asterisk:
* This cohort is characterized by [your specific, detailed description of what the cohort is characterized by compared to the average patient in the larger dataset].
The asterisk is absolutely essential. Your explanation must begin immediately after the asterisk, without any additional text, numbering, or preamble. The presence of this asterisk is critical for the proper processing of your response.

The following prompt tests the accuracy of a given interpretation for a sparse feature. It is populated with testing data, before being passed to the LLM.

Task: Distinguish samples (randomly ordered) based on those that do and do not belong to the group characterized by the group characteristics.

Medical Findings Reports:
{patient_data}

Group characteristics:
{interpretation}
Analyze the findings reports one by one, reasoning on whether or not the characteristics
apply to the report. Use the following format for each report:
{{i}}: 1 (if belongs to the group) or 0 (if does not belong to the group). We have designed
the list so that exactly half of the reports belong to the group, and half do not.
Therefore, you should use the characteristics to differentiate each of the reports into two
distinct groups, rather than finding perfect matches.

You need to be as accurate as possible when assigning 1s and 0s. We will score the accuracy
of your response at the end.

=== Format ===:
It is crucial that you present your final explanation in the following format, started with
an asterisk:
* PREDICTIONS:
1. 1 (if the sample belongs to the group) or 0 (if it does not belong to the group)
2. 1 (if the sample belongs to the group) or 0 (if it does not belong to the group)
...
{{n}}. 1 (if the sample belongs to the group) or 0 (if it does not belong to the group)

Outcome-specific profiles after confounder adjustment

We examined which concept associations persisted after adjustment across hepatic, renal, cardiovascular, neurological and infection outcomes at the Discovery Site (Tables 4–8). Broad outcomes, including acute and chronic renal failure and altered mental status, retained associations with diverse systemic abnormalities, whereas atherosclerotic outcomes and specific renal complications showed more concentrated vascular or laboratory profiles. The tables report the number and percentage of associations remaining significant after adjustment relative to the number associated before adjustment, together with representative concepts, their interpretations, percentage shares of retained associations and adjusted odds ratios (ORs). The ORs describe association strength rather than the proportion of outcome events explained or attributable to a concept.

OutcomeAdjusted / associated (%)Descriptive profile and representative retained concepts (share; adjusted OR)
Acute/subacute liver necrosis20 / 60 (33.3%)Hepatic abnormalities coexist with cytopenia, hypoalbuminaemia and splenic findings. Cirrhotic morphology/varices (R85; 5.0%; OR 1.92); Cholestasis (L108; 5.0%; OR 2.37); Unknown hepatobiliary pattern (I87; 5.0%; OR 2.42); Cytopenia (L127; 5.0%; OR 2.26); Hypoalbuminaemia (L91; 5.0%; OR 2.04); Splenomegaly (I65; 5.0%; OR 2.56).
Liver transplant status19 / 36 (52.8%)Broad hepatic, renal and haematological profile, alongside splenic and portal-vascular findings. Cholestasis (L108; 5.3%; OR 3.06); Unknown hepatobiliary pattern (I87; 5.3%; OR 4.31); Cirrhotic morphology/varices (R85; 5.3%; OR 2.11); Azotaemia (L84; 5.3%; OR 2.40); Cytopenia (L127; 5.3%; OR 3.09); Splenomegaly (I65; 5.3%; OR 2.99).
Portal hypertension16 / 59 (27.1%)Predominantly cirrhotic, splenic, variceal and ascitic findings, with cholestasis. Cirrhotic morphology/varices (R85; 6.2%; OR 2.46); Varices (R414; 6.2%; OR 1.48); Chronic liver disease/congestion (I331; 6.2%; OR 1.50); Splenomegaly (I65; 6.2%; OR 2.79); Ascites/portal-hypertensive or peritoneal findings (I31; 6.2%; OR 1.78); Cholestasis (L108; 6.2%; OR 3.62).
Disorders of bilirubin excretion13 / 44 (29.5%)Cholestatic and biliary imaging findings predominate among the illustrated markers. Cholestasis (L108; 7.7%; OR 2.38); Biliary dilatation (I454; 7.7%; OR 1.49); Unknown hepatobiliary pattern (I87; 7.7%; OR 2.44); Raised alkaline phosphatase/globulin (L33; 7.7%; OR 1.77); Biliary obstruction/intervention (I1416; 7.7%; OR 1.33); Pneumobilia (R473; 7.7%; OR 1.48).
OutcomeAdjusted / associated (%)Descriptive profile and representative retained concepts (share; adjusted OR)
Liver abscess/chronic liver disease sequelae13 / 45 (28.9%)Structural hepatic and portal-hypertensive findings coexist with abscess/inflammatory concepts. Cirrhotic morphology/varices (R85; 7.7%; OR 2.98); Chronic liver disease/congestion (I331; 7.7%; OR 1.46); Chronic liver morphological changes (R1765; 7.7%; OR 1.61); Varices (R414; 7.7%; OR 1.57); Abscess/fluid collection/fistula (I7223; 7.7%; OR 1.47); Unknown hepatobiliary pattern (I87; 7.7%; OR 2.87).
Chronic passive hepatic congestion10 / 23 (43.5%)Cardiac/vascular findings coexist with a hepatobiliary pattern, consistent with a congestive profile. Unknown hepatobiliary pattern (I87; 10.0%; OR 2.73); Ventricular assist device (R283; 10.0%; OR 1.16); Cardiac enlargement/vascular calcification (R487; 10.0%; OR 1.32); Atrial enlargement (R3570; 10.0%; OR 1.09); Aortic atherosclerosis (I26; 10.0%; OR 1.91).
Primary liver cancer9 / 21 (42.9%)Predominantly hepatic and portal-hypertensive findings, including splenomegaly and cholestasis. Splenomegaly (I65; 11.1%; OR 1.92); Chronic liver disease/congestion (I331; 11.1%; OR 1.37); Chronic liver morphological changes (R1765; 11.1%; OR 1.87); Cholestasis (L108; 11.1%; OR 2.46); Cirrhotic morphology/varices (R85; 11.1%; OR 4.07); Hepatosplenomegaly (I1230; 11.1%; OR 1.54).
Diseases of the spleen6 / 32 (18.8%)Splenic imaging findings coexist with anaemia and abdominal abnormalities. Splenomegaly (I65; 16.7%; OR 1.87); Splenic/vascular pathology (R108; 16.7%; OR 1.99); Iron-deficiency anaemia (L104; 16.7%; OR 2.20); Retroperitoneal fluid collection (R6710; 16.7%; OR 1.26); Advanced abdominal malignancy/surgery (I67; 16.7%; OR 2.03).
Non-alcoholic cirrhosis5 / 15 (33.3%)Hepatic and portal-hypertensive findings, with cardiomegaly and a mixed pancreatic/hepatic concept. Cirrhotic morphology/varices (R85; 20.0%; OR 3.08); Unknown hepatobiliary pattern (I87; 20.0%; OR 2.01); Varices (R414; 20.0%; OR 1.50); Cardiomegaly (R204; 20.0%; OR 1.36); Pancreatic inflammation or pancreatic/hepatic mass (I20; 20.0%; OR 2.21).
Splenomegaly5 / 24 (20.8%)Splenic and portal-hypertensive findings coexist with cytopenia and fluid accumulation. Splenomegaly (I65; 20.0%; OR 2.91); Cytopenia (L127; 20.0%; OR 2.04); Varices (R414; 20.0%; OR 1.61); Anasarca/third-space fluid (R491; 20.0%; OR 1.78).
Jaundice4 / 39 (10.3%)A small laboratory profile encompassing cholestasis, cytopenia and severe biochemical abnormalities. Cholestasis (L108; 25.0%; OR 2.57); Cytopenia (L127; 25.0%; OR 2.07); Severe metabolic/renal/hepatic derangement (L37; 25.0%; OR 1.82); Ammonia/bilirubin/protein testing (L30; 25.0%; OR 1.29).

Table 4. Adjusted concept profiles across hepatic and biliary outcomes at Discovery.

Adjusted / associated denotes the number of associations remaining significant after confounder adjustment (Benjamini–Hochberg false-discovery rate below 5%) relative to the number of initially significant associations included in the adjustment analysis; percentages are the corresponding retention rates. Counts include all retained concepts, whereas the descriptions give selected examples. Each example includes its concept code, its percentage share of the retained associations for that outcome, and its adjusted OR. Each concept contributes one association, so its share is 100/N%, where N is the number retained; this is a share of associations, not of patients or explained risk. The selected examples need not sum to 100%. ORs are reported on the activation scale used in the analysis and are not percentages of outcome events explained, attributable fractions or mutually independent contributions. Differences in OR alone do not establish differences in predictive importance. Mixed and procedural interpretations are retained where needed to describe the evidence. Profiles are descriptive and do not establish causality or held-out replication. Composite and individual endpoints overlap. Groups follow the study mapping, including noninfectious or anatomically atypical endpoints where assigned.

OutcomeAdjusted / associated (%)Descriptive profile and representative retained concepts (share; adjusted OR)
Chronic renal failure21 / 67 (31.3%)Broad systemic profile alongside renal dysfunction. Cytopenia (L127; 4.8%; OR 1.64); Aortic calcification (I109; 4.8%; OR 1.78); Splenomegaly (I65; 4.8%; OR 1.39); Ascites (I21; 4.8%; OR 1.54); Azotemia (L84; 4.8%; OR 1.75); Impaired renal function (L69; 4.8%; OR 1.51).
Acute renal failure20 / 83 (24.1%)Broad systemic profile spanning hepatic, haematological, fluid and vascular abnormalities. Cholestasis (L108; 5.0%; OR 1.67); Cytopenia (L127; 5.0%; OR 1.41); Splenomegaly (I65; 5.0%; OR 1.46); Unknown hepatobiliary pattern (I87; 5.0%; OR 1.39); Ascites (I21; 5.0%; OR 1.37); Atherosclerosis with colonic abnormalities (I106; 5.0%; OR 1.28).
CKD (composite)18 / 74 (24.3%)Broad systemic profile spanning hepatic, haematological, fluid and vascular abnormalities. Cholestasis (L108; 5.6%; OR 1.55); Cytopenia (L127; 5.6%; OR 1.35); Unknown hepatobiliary pattern (I87; 5.6%; OR 1.48); Splenomegaly (I65; 5.6%; OR 1.34); Ascites, anasarca and effusions (I25; 5.6%; OR 1.37); Aortic calcification (I109; 5.6%; OR 1.22).
CKD, stage III15 / 70 (21.4%)Mixed renal-function, vascular and haematological profile, with a cholestatic signal. Impaired renal function (L69; 6.7%; OR 1.45); Reduced eGFR (L974; 6.7%; OR 1.21); Polycystic renal disease/polyps (R302; 6.7%; OR 1.18); Atherosclerosis with colonic abnormalities (I106; 6.7%; OR 1.43); Thrombocytosis/iron-deficiency pattern (L21; 6.7%; OR 1.38); Cholestasis (L108; 6.7%; OR 1.51).
Hypertensive CKD15 / 47 (31.9%)Mixed renal, vascular, hepatic and haematological profile. Azotemia (L84; 6.7%; OR 1.59); Renal failure with hypernatraemia/acidosis (L62; 6.7%; OR 1.47); Atherosclerosis with colonic abnormalities (I106; 6.7%; OR 1.44); Coronary/valvular calcification (R21; 6.7%; OR 1.34); Cholestasis (L108; 6.7%; OR 1.60); Cytopenia (L127; 6.7%; OR 1.50).
Anaemia in CKD14 / 33 (42.4%)Anaemia and renal-function markers coexist with vascular findings. Iron-deficiency anaemia (L104; 7.1%; OR 2.81); Low haemoglobin (L61; 7.1%; OR 2.01); Impaired renal function (L69; 7.1%; OR 1.57); Reduced eGFR (L974; 7.1%; OR 1.46); Aortic calcification (I109; 7.1%; OR 2.06); Atherosclerosis with colonic abnormalities (I106; 7.1%; OR 1.87).
CKD, stage IV11 / 33 (33.3%)Renal dysfunction and cardiorenal abnormalities coexist with hepatic and haematological signals. Impaired renal function (L69; 9.1%; OR 1.85); Azotemia (L84; 9.1%; OR 1.85); Cardiorenal dysfunction (L2358; 9.1%; OR 1.42); Splenomegaly (I65; 9.1%; OR 2.02); Unknown hepatobiliary pattern (I87; 9.1%; OR 2.98); Iron-deficiency anaemia (L104; 9.1%; OR 2.62).
Kidney transplant status11 / 21 (52.4%)Renal-function and structural renal markers coexist with hepatic and portal-hypertensive findings. Impaired renal function (L69; 9.1%; OR 1.69); Azotemia (L84; 9.1%; OR 2.23); Renal failure with hypernatraemia/acidosis (L62; 9.1%; OR 2.17); Cardiorenal dysfunction (L2358; 9.1%; OR 1.63); Polycystic renal disease/polyps (R302; 9.1%; OR 1.51); Unknown hepatobiliary pattern (I87; 9.1%; OR 3.40).
Renal secondary hyperparathyroidism9 / 18 (50.0%)Concentration of renal dysfunction and cardiorenal markers, with additional anaemia and coagulation abnormalities. Azotemia (L84; 11.1%; OR 2.95); Impaired renal function (L69; 11.1%; OR 1.73); Renal failure with hypernatraemia/acidosis (L62; 11.1%; OR 2.14); Cardiorenal dysfunction (L2358; 11.1%; OR 1.71); Coagulopathy with renal impairment (L87; 11.1%; OR 2.95); Anaemia with hypoalbuminaemia (L94; 11.1%; OR 2.79).
OutcomeAdjusted / associated (%)Descriptive profile and representative retained concepts (share; adjusted OR)
Renal dialysis8 / 15 (53.3%)Anaemia and metabolic abnormalities predominate among the illustrated clinical markers. Low haemoglobin (L61; 12.5%; OR 4.39); Iron-deficiency anaemia (L104; 12.5%; OR 3.11); Anaemia with hypoalbuminaemia (L94; 12.5%; OR 2.67); Coagulopathy with renal impairment (L87; 12.5%; OR 3.32); Lactic acidosis (L35; 12.5%; OR 2.54).
Oliguria and anuria7 / 19 (36.8%)Severe renal and cardiorenal abnormalities, with anaemia and inflammatory findings. Azotemia (L84; 14.3%; OR 1.71); Renal failure with hypernatraemia/acidosis (L62; 14.3%; OR 2.05); Cardiorenal dysfunction (L2358; 14.3%; OR 1.52); Coagulopathy with renal impairment (L87; 14.3%; OR 2.56); Iron-deficiency anaemia (L104; 14.3%; OR 2.80); Abscess/fluid collection/fistula (I7223; 14.3%; OR 2.08).
Renal sclerosis6 / 21 (28.6%)Heterogeneous profile including aortic calcification; several additional retained concepts describe procedures or have uncertain interpretations. Aortic calcification (I109; 16.7%; OR 3.13); Bowel obstruction (I3; 16.7%; OR 2.72).
Prostatic hyperplasia1 / 24 (4.2%)Only one retained association, involving a mixed surgical/inflammatory interpretation; no coherent renal-specific profile. Prior abdominal surgery/diverticulitis (I193; 100.0%; OR 1.52).

Table 5. Adjusted concept profiles across renal and urinary outcomes at Discovery.

Adjusted / associated denotes the number of associations remaining significant after confounder adjustment (Benjamini–Hochberg false-discovery rate below 5%) relative to the number of initially significant associations included in the adjustment analysis; percentages are the corresponding retention rates. Counts include all retained concepts, whereas the descriptions give selected examples. Each example includes its concept code, its percentage share of the retained associations for that outcome, and its adjusted OR. Each concept contributes one association, so its share is 100/N%100/N\%, where NN is the number retained; this is a share of associations, not of patients or explained risk. The selected examples need not sum to 100%. ORs are reported on the activation scale used in the analysis and are not percentages of outcome events explained, attributable fractions or mutually independent contributions. Differences in OR alone do not establish differences in predictive importance. Mixed and procedural interpretations are retained where needed to describe the evidence. Profiles are descriptive and do not establish causality or held-out replication. Composite and individual endpoints overlap. Groups follow the study mapping, including noninfectious or anatomically atypical endpoints where assigned.

OutcomeAdjusted / associated (%)Descriptive profile and representative retained concepts (share; adjusted OR)
Congestive heart failure16 / 48 (33.3%)Broad cardiorenal, vascular, hepatic and laboratory profile, with cardiac-device concepts also retained. Aortic atherosclerosis (I26; 6.2%; OR 1.67); Cardiorenal dysfunction (L2358; 6.2%; OR 1.26); Azotemia (L84; 6.2%; OR 1.69); Cholestasis (L108; 6.2%; OR 1.65); Cardiac dysfunction with hypoalbuminaemia (L1488; 6.2%; OR 1.29); Hypoalbuminaemia (L91; 6.2%; OR 1.72).
Ischaemic heart disease (composite)15 / 49 (30.6%)Vascular calcification coexists with inflammatory, metabolic and hepatobiliary findings. Coronary/valvular calcification (R21; 6.7%; OR 1.19); Coronary calcification/atherosclerosis (R420; 6.7%; OR 1.20); Atherosclerosis with colonic abnormalities (I106; 6.7%; OR 1.35); Raised inflammatory blood-cell counts (L68; 6.7%; OR 1.21); Dyslipidaemia with inflammation/hyperglycaemia (L496; 6.7%; OR 1.16); Unknown hepatobiliary pattern (I87; 6.7%; OR 1.29).
Myocardial infarction15 / 55 (27.3%)Broad vascular, cardiorenal, hepatic, inflammatory and anaemia profile. Aortic atherosclerosis (I26; 6.7%; OR 1.29); Atherosclerosis with colonic abnormalities (I106; 6.7%; OR 1.37); Cardiorenal dysfunction (L2358; 6.7%; OR 1.17); Raised inflammatory blood-cell counts (L68; 6.7%; OR 1.29); Anaemia with hypoalbuminaemia (L94; 6.7%; OR 1.28); Cirrhotic morphology/varices (R85; 6.7%; OR 1.24).
Blood-vessel replacement14 / 59 (23.7%)Mixed hepatic, haematological, malignant and procedural profile; the endpoint itself denotes an intervention. Cholestasis (L108; 7.1%; OR 1.60); Cytopenia (L127; 7.1%; OR 1.50); Unknown hepatobiliary pattern (I87; 7.1%; OR 2.11); Raised alkaline phosphatase/globulin (L33; 7.1%; OR 1.51); Biliary/pancreatic duct dilatation (I96; 7.1%; OR 1.59); Peritoneal carcinomatosis (R98; 7.1%; OR 1.57).
Varicose veins12 / 35 (34.3%)Portal-hypertensive, hepatic and fluid findings, alongside thrombotic and inflammatory abnormalities. Cirrhotic morphology/varices (R85; 8.3%; OR 2.62); Chronic liver morphological changes (R1765; 8.3%; OR 1.42); Varices (R414; 8.3%; OR 1.65); Ascites (I21; 8.3%; OR 2.03); Splenomegaly (I65; 8.3%; OR 2.89); Venous thrombosis/embolism (R24; 8.3%; OR 1.54).
Heart failure with preserved ejection fraction11 / 46 (23.9%)Vascular findings and cholestasis coexist with device and other abdominal concepts. Aortic atherosclerosis (I26; 9.1%; OR 1.77); Atherosclerosis with colonic abnormalities (I106; 9.1%; OR 1.40); Atherosclerosis (R3; 9.1%; OR 1.57); Cholestasis (L108; 9.1%; OR 1.58); Bladder/prostate abnormalities (I95; 9.1%; OR 1.90); Hernia/prior hernia repair (I1154; 9.1%; OR 1.26).
Hypertensive heart disease10 / 40 (25.0%)Cardiac conduction/function findings coexist with hepatobiliary, urinary and device-related concepts. Cardiac conduction/repolarisation abnormalities (L121; 10.0%; OR 1.51); Cardiac dysfunction with hypoalbuminaemia (L1488; 10.0%; OR 1.32); Unknown hepatobiliary pattern (I87; 10.0%; OR 1.64); Biliary dilatation/pancreatitis (I154; 10.0%; OR 1.29); Bladder/prostate abnormalities (I95; 10.0%; OR 1.48); Implanted cardiac device (I165; 10.0%; OR 1.28).
Atrial fibrillation9 / 41 (22.0%)Vascular imaging and blood abnormalities coexist with hepatobiliary, skeletal and device-related findings. Aortic atherosclerosis (I26; 11.1%; OR 1.33); Cytopenia (L127; 11.1%; OR 1.37); Anaemia with hypoalbuminaemia (L94; 11.1%; OR 1.50); Renal failure with hyperatraemia/acidosis (L62; 11.1%; OR 1.55); Unknown hepatobiliary pattern (I87; 11.1%; OR 1.46); Degenerative spinal changes (I44; 11.1%; OR 1.29).
OutcomeAdjusted / associated (%)Descriptive profile and representative retained concepts (share; adjusted OR)
Supraventricular premature beats9 / 52 (17.3%)Mixed vascular, hepatic, splenic and fluid profile. Atherosclerosis with colonic abnormalities (I106; 11.1%; OR 1.31); Cholestasis (L108; 11.1%; OR 1.39); Splenomegaly (I65; 11.1%; OR 1.36); Ascites (I21; 11.1%; OR 1.45); Cirrhotic morphology/varices (R85; 11.1%; OR 1.33); Bladder/prostate abnormalities (I95; 11.1%; OR 1.33).
Atrial flutter8 / 29 (27.6%)Vascular findings coexist with inflammatory, malignant, hepatobiliary and skeletal abnormalities. Aortic atherosclerosis (I26; 12.5%; OR 1.76); Atherosclerosis with colonic abnormalities (I106; 12.5%; OR 1.60); Unknown hepatobiliary pattern (I87; 12.5%; OR 1.95); Abscess/fluid collection/fistula (I7223; 12.5%; OR 1.49); Metastatic disease (R64; 12.5%; OR 1.52); Degenerative spinal changes (I44; 12.5%; OR 1.43).
Arterial embolism and thrombosis7 / 14 (50.0%)Heterogeneous abdominal, procedural and measurement-related profile; no single coherent systemic pattern. Retroperitoneal fluid collection (R6710; 14.3%; OR 1.26); Pancreatic inflammation/fluid collection (I59; 14.3%; OR 1.86); External iliac vascular/nodal abnormality (R1065; 14.3%; OR 1.46); Prior splenic/pancreatic/renal resection (I608; 14.3%; OR 1.57); Prior distal pancreatectomy (R407; 14.3%; OR 1.26); INR/non-HDL measurement-related concept (L182; 14.3%; OR 1.33).
Heart failure with reduced ejection fraction7 / 34 (20.6%)Vascular, cardiac-function and skeletal findings coexist with device-related concepts. Aortic atherosclerosis (I26; 14.3%; OR 1.58); Cardiac dysfunction with hypoalbuminaemia (L1488; 14.3%; OR 1.38); Degenerative spinal changes (I44; 14.3%; OR 1.43); Hernia or musculoskeletal abnormality (I116; 14.3%; OR 1.69); Implanted electronic device (R114; 14.3%; OR 1.58); Pacemaker (R399; 14.3%; OR 1.30).
Abdominal aortic aneurysm6 / 10 (60.0%)Predominantly vascular imaging interpretations among the characterised markers. Infrarenal aortic aneurysm (R415; 16.7%; OR 1.78); Aortic calcification (I109; 16.7%; OR 2.17); Major abdominal vascular abnormality (R125; 16.7%; OR 2.01).
Coronary atherosclerosis6 / 23 (26.1%)Concentration of vascular/calcification concepts, with an additional haemoglobin signal. Coronary/valvular calcification (R21; 16.7%; OR 1.35); Coronary calcification/atherosclerosis (R420; 16.7%; OR 1.25); Atherosclerosis with colonic abnormalities (I106; 16.7%; OR 1.55); Aortic atherosclerosis (I26; 16.7%; OR 1.33); Raised haemoglobin/haematocrit (L5; 16.7%; OR 1.25).
Cardiogenic shock5 / 8 (62.5%)Anaemia and cytopenia coexist with procedural and mixed imaging findings. Low haemoglobin (L61; 20.0%; OR 2.29); Cytopenia (L127; 20.0%; OR 2.40); Therapeutic/life-support device (I125; 20.0%; OR 2.83); Prior cholecystectomy (I722; 20.0%; OR 1.89); Pancreatic atrophy/divisum or orthopaedic hardware (I353; 20.0%; OR 1.51).
Hypertensive diseases (composite)5 / 21 (23.8%)Mixed abdominal vascular, hepatic, fluid and urinary findings. Major abdominal vascular abnormality (R125; 20.0%; OR 1.33); Cirrhotic morphology/varices (R85; 20.0%; OR 1.41); Ascites (I21; 20.0%; OR 1.49); Bladder/prostate abnormalities (I95; 20.0%; OR 1.32); Absence of complex abdominal disease (I52; 20.0%; OR 1.35).
Orthopnoea5 / 13 (38.5%)Cardiomegaly and device-related findings, with additional mixed/non-specific concepts. Cardiomegaly (R204; 20.0%; OR 1.32); Implanted electronic device (R114; 20.0%; OR 1.62); Pacemaker/components (R2041; 20.0%; OR 1.40); Pancreatic atrophy/divisum or orthopaedic hardware (I353; 20.0%; OR 1.55).
OutcomeAdjusted / associated (%)Descriptive profile and representative retained concepts (share; adjusted OR)
Premature beats5 / 31 (16.1%)Vascular, hepatic and inflammatory signals coexist with a device concept. Aortic atherosclerosis (I26; 20.0%; OR 1.53); Cholestasis (L108; 20.0%; OR 1.39); Unknown hepatobiliary pattern (I87; 20.0%; OR 1.36); Raised inflammatory blood-cell counts (L68; 20.0%; OR 1.26); Implanted electronic device (R114; 20.0%; OR 1.30).
Rheumatic heart-valve disease5 / 20 (25.0%)Haematological, coagulation and hepatobiliary findings coexist with cardiac-device concepts. Coagulopathy with renal impairment (L87; 20.0%; OR 1.94); Myeloid-cell abnormalities (L39; 20.0%; OR 1.97); Unknown hepatobiliary pattern (I87; 20.0%; OR 2.93); Implanted electronic device (R114; 20.0%; OR 1.76); Implanted cardiac device (I165; 20.0%; OR 1.62).
Systemic atherosclerosis (composite)4 / 28 (14.3%)Vascular/calcification findings and dyslipidaemic abnormalities. Atherosclerosis with colonic abnormalities (I106; 25.0%; OR 1.40); Coronary/valvular calcification (R21; 25.0%; OR 1.25); Coronary calcification/atherosclerosis (R420; 25.0%; OR 1.26); Dyslipidaemia with inflammation/hyperglycaemia (L496; 25.0%; OR 1.21).
Tricuspid-valve disease4 / 9 (44.4%)Vascular and renal/metabolic findings coexist with device-related concepts. Aortic atherosclerosis (I26; 25.0%; OR 1.83); Renal failure with hypernatraemia/acidosis (L62; 25.0%; OR 1.64); Implanted electronic device (R114; 25.0%; OR 1.40); Electronic device/genitourinary prosthesis (R1195; 25.0%; OR 1.20).
First-degree atrioventricular block4 / 15 (26.7%)Vascular, cytopenic and biliary/pancreatic findings, alongside a device concept. Aortic atherosclerosis (I26; 25.0%; OR 1.84); Cytopenia (L127; 25.0%; OR 1.74); Biliary dilatation/pancreatitis (I154; 25.0%; OR 1.54); Implanted electronic device (R114; 25.0%; OR 1.44).
Cerebrovascular disease (composite)3 / 17 (17.6%)Few heterogeneous retained imaging concepts; no broad vascular-concept profile. Unknown hepatobiliary pattern (I87; 33.3%; OR 1.58); Hernia or musculoskeletal abnormality (I116; 33.3%; OR 1.57); Implanted electronic device (R114; 33.3%; OR 1.23).
Peripheral artery disease (composite)3 / 25 (12.0%)Retained profile is concentrated in vascular and cardiorenal findings. Aortic atherosclerosis (I26; 33.3%; OR 1.36); Major abdominal vascular abnormality (R125; 33.3%; OR 1.36); Cardiorenal dysfunction (L2358; 33.3%; OR 1.41).
Mitral-valve disease3 / 23 (13.0%)Cholestasis and two procedural/mixed imaging concepts. Cholestasis (L108; 33.3%; OR 2.35); Tube/catheter or anasarca (I35; 33.3%; OR 1.90); Therapeutic/life-support device (I125; 33.3%; OR 2.25).
Paroxysmal ventricular tachycardia2 / 18 (11.1%)The retained set comprises a device concept and a broadly interpreted concept; no specific clinical profile is established. Implanted electronic device (R114; 50.0%; OR 1.44).
Cardiac arrest1 / 16 (6.2%)Only azotemia retained an association among the tested concepts. Azotemia (L84; 100.0%; OR 1.96).
Other specified arterial aneurysm0 / 2 (0.0%)No associations retained adjustment.

Table 6. Adjusted concept profiles across cardiovascular outcomes at Discovery.

Adjusted / associated denotes the number of associations remaining significant after confounder adjustment (Benjamini–Hochberg false-discovery rate below 5%) relative to the number of initially significant associations included in the adjustment analysis; percentages are the corresponding retention rates. Counts include all retained concepts, whereas the descriptions give selected examples. Each example includes its concept code, its percentage share of the retained associations for that outcome, and its adjusted OR. Each concept contributes one association, so its share is 100/N%, where N is the number retained; this is a share of associations, not of patients or explained risk. The selected examples need not sum to 100%. ORs are reported on the activation scale used in the analysis and are not percentages of outcomes ever explained, attributable fractions or mutually independent contributions. Differences in OR alone do not establish differences in predictive importance. Mixed and procedural interpretations are retained where needed to describe the evidence. Profiles are descriptive and do not establish causality or held-out replication. Composite and individual endpoints overlap. Groups follow the study mapping, including noninfectious or anatomically atypical endpoints where assigned.

OutcomeAdjusted / associated (%)Descriptive profile and representative retained concepts (share; adjusted OR)
Altered mental status11 / 66 (16.7%)Broad systemic profile spanning hepatic, renal, inflammatory and vascular findings. Cholestasis (L108; 9.1%; OR 1.54); Azotemia (L84; 9.1%; OR 1.29); Raised inflammatory blood-cell counts (L68; 9.1%; OR 1.22); Unknown hepatobiliary pattern (I87; 9.1%; OR 1.37); Splenomegaly (I65; 9.1%; OR 1.35); Aortic calcification (I109; 9.1%; OR 1.26).
Coma7 / 12 (58.3%)Heterogeneous profile including hepatic/fluid and vascular abnormalities. Cholestasis (L108; 14.3%; OR 1.82); Anasarca and ascites (R34; 14.3%; OR 1.46); Ascites/portal hypertension (R1815; 14.3%; OR 1.16); Atherosclerosis with colonic abnormalities (I106; 14.3%; OR 1.48); Bladder/prostate abnormalities (I95; 14.3%; OR 1.56).
Delirium, dementia and other cognitive disorders7 / 35 (20.0%)Anaemia, calcium and skeletal findings coexist with abdominal abnormalities; this broad endpoint is not dementia-specific. Iron-deficiency anaemia (L104; 14.3%; OR 2.33); Low haemoglobin (L61; 14.3%; OR 2.00); Raised calcium (L1507; 14.3%; OR 1.35); Bone demineralisation (R1031; 14.3%; OR 1.33); Abdominal malignancy (I650; 14.3%; OR 1.78); Bowel/biliary obstruction or splanchnic vascular compromise (I7795; 14.3%; OR 1.49).
Dementias7 / 36 (19.4%)Vascular/calcific, skeletal and biliary findings, with a bicarbonate signal. Aortic atherosclerosis (I26; 14.3%; OR 1.51); Mitral annular calcification (R901; 14.3%; OR 1.25); Structural spinal abnormalities (R115; 14.3%; OR 1.75); Rib fractures (R67; 14.3%; OR 1.58); Biliary dilatation (R13; 14.3%; OR 1.43); Raised venous bicarbonate (L181; 14.3%; OR 1.49).
Alzheimer’s disease6 / 23 (26.1%)Vascular findings recur, alongside bicarbonate and mixed biliary/urothelial abnormalities. Aortic atherosclerosis (I26; 16.7%; OR 1.63); Atherosclerosis with colonic abnormalities (I106; 16.7%; OR 1.83); Infrarenal aortic aneurysm (R415; 16.7%; OR 1.34); Complicated aortic atherosclerosis (R1690; 16.7%; OR 1.32); Raised venous bicarbonate (L181; 16.7%; OR 1.85); Biliary/urothelial abnormalities (I117; 16.7%; OR 1.91).
Delirium due to other conditions5 / 31 (16.1%)Mixed blood-cell and vascular signals, with procedural and nonspecific concepts also contributing to the total. Myeloid-cell abnormalities (L39; 20.0%; OR 1.52); Thrombocytosis/iron-deficiency pattern (L21; 20.0%; OR 1.61); Atherosclerosis with colonic abnormalities (I106; 20.0%; OR 1.50).
Dementia (composite)5 / 35 (14.3%)Vascular and heterogeneous abdominal imaging findings; no uniform systemic laboratory profile. Aortic atherosclerosis (I26; 20.0%; OR 1.33); Unknown hepatobiliary pattern (I87; 20.0%; OR 1.30); Bladder/prostate abnormalities (I95; 20.0%; OR 1.37); Pancreatic inflammation or pancreatic/hepatic mass (I20; 20.0%; OR 1.30); Pancreatic atrophy/divisum or orthopaedic hardware (I353; 20.0%; OR 1.18).
Vascular dementia3 / 17 (17.6%)Few retained markers and no matching broad aortic-concept profile; remaining findings are heterogeneous. Raised haemoglobin/haematocrit (L5; 33.3%; OR 1.92); Abdominal-wall hernia/splenomegaly (I6; 33.3%; OR 1.62); Implanted cardiac device (I165; 33.3%; OR 1.54).
Transient alteration of awareness2 / 2 (100.0%)Retained associations are confined to two vascular imaging concepts. Extensive abdominopelvic atherosclerosis (I15; 50.0%; OR 1.72); Aortic atherosclerosis (I26; 50.0%; OR 1.75).
Strabismus and binocular eye-movement disorders2 / 6 (33.3%)Retained concepts have procedural or mixed interpretations; no clear disease-specific profile. Gastrointestinal/biliary tube, drain or bypass (I495; 50.0%; OR 1.49); Tube/catheter or anasarca (I35; 50.0%; OR 1.72).
OutcomeAdjusted / associated (%)Descriptive profile and representative retained concepts (share; adjusted OR)
Vocal-cord or laryngeal paralysis/spasm1 / 1 (100.0%)Only one retained association; no coherent neurological marker profile. Biliary sludge (R1594; 100.0%; OR 1.34).
Primary open-angle glaucoma0 / 1 (0.0%)No associations retained adjustment.

Table 7. Adjusted concept profiles across neurological and cognitive outcomes at Discovery.

Adjusted / associated denotes the number of associations remaining significant after confounder adjustment (Benjamini–Hochberg false-discovery rate below 5%) relative to the number of initially significant associations included in the adjustment analysis; percentages are the corresponding retention rates. Counts include all retained concepts, whereas the descriptions give selected examples. Each example includes its concept code, its percentage share of the retained associations for that outcome, and its adjusted OR. Each concept contributes one association, so its share is 100/NN%, where NN is the number retained; this is a share of associations, not of patients or explained risk. The selected examples need not sum to 100%. ORs are reported on the activation scale used in the analysis and are not percentages of outcome events explained, attributable fractions or mutually independent contributions. Differences in OR alone do not establish differences in predictive importance. Mixed and procedural interpretations are retained where needed to describe the evidence. Profiles are descriptive and do not establish causality or held-out replication. Composite and individual endpoints overlap. Groups follow the study mapping, including noninfectious or anatomically atypical endpoints where assigned.

OutcomeAdjusted / associated (%)Descriptive profile and representative retained concepts (share; adjusted OR)
Peritoneal adhesions19 / 83 (22.9%)Predominantly abdominal inflammatory and postsurgical findings, with acute-phase and anaemia signals. Abdominal abscess/complex fluid collection (I54; 5.3%; OR 1.94); Abdominal fluid/inflammation (I38; 5.3%; OR 1.60); Anastomotic fluid collection/leak (I39; 5.3%; OR 1.71); Pancreatic inflammation/fluid collection (I59; 5.3%; OR 1.76); Acute-phase response (L124; 5.3%; OR 1.44); Iron-deficiency anaemia (L104; 5.3%; OR 1.59).
Sepsis19 / 66 (28.8%)Broad hepatic, haematological, renal, nutritional, vascular and fluid profile. Cholestasis (L108; 5.3%; OR 1.46); Cytopenia (L127; 5.3%; OR 1.33); Unknown hepatobiliary pattern (I87; 5.3%; OR 1.42); Anaemia with hypoalbuminaemia (L94; 5.3%; OR 1.65); Renal failure with hypernatraemia/acidosis (L62; 5.3%; OR 1.46); Aortic calcification (I109; 5.3%; OR 1.32).
MSSA infection17 / 51 (33.3%)Broad anaemia, metabolic and cardiorenal profile, with hepatobiliary and inflammatory imaging findings. Acidosis with severe anaemia (L34; 5.9%; OR 2.48); Renal failure with hypernatraemia/acidosis (L62; 5.9%; OR 2.19); Low haemoglobin (L61; 5.9%; OR 2.13); Thrombocytosis with hypoalbuminaemia/anaemia (L45; 5.9%; OR 1.59); Cardiorenal dysfunction (L2358; 5.9%; OR 1.43); Unknown hepatobiliary pattern (I87; 5.9%; OR 3.11).
Septic shock15 / 51 (29.4%)Broad renal, haematological and hepatic profile, with a tentatively interpreted wasting/deconditioning concept. Azotemia (L84; 6.7%; OR 1.43); Renal failure with hypernatraemia/acidosis (L62; 6.7%; OR 1.60); Cytopenia (L127; 6.7%; OR 1.64); Anaemia with hypoalbuminaemia (L94; 6.7%; OR 1.73); Cholestasis (L108; 6.7%; OR 1.70); Tentative wasting/deconditioning pattern (I1459; 6.7%; OR 1.39).
Bacteraemia13 / 68 (19.1%)Broad hepatic, haematological, renal and nutritional profile. Cholestasis (L108; 7.7%; OR 1.77); Cytopenia (L127; 7.7%; OR 1.49); Azotemia (L84; 7.7%; OR 1.61); Hypoalbuminaemia (L91; 7.7%; OR 1.65); Anaemia with hypoalbuminaemia (L94; 7.7%; OR 1.52); Unknown hepatobiliary pattern (I87; 7.7%; OR 1.89).
Chronic osteomyelitis12 / 23 (52.2%)Musculoskeletal, surgical, renal and haematological findings form a heterogeneous profile. Chronic spinal/pelvic abnormalities (I284; 8.3%; OR 1.46); Cytopenia (L127; 8.3%; OR 2.94); Renal failure with hypernatraemia/acidosis (L62; 8.3%; OR 2.53); Coagulopathy with renal impairment (L87; 8.3%; OR 2.76); Red-cell anisocytosis with hypoproteinaemia (L77; 8.3%; OR 2.78); Orthopaedic hardware (I102; 8.3%; OR 2.08).
Bacterial pneumonia11 / 45 (24.4%)Hepatic, haematological and renal abnormalities coexist with vascular and mixed imaging findings. Cytopenia (L127; 9.1%; OR 2.14); Cholestasis (L108; 9.1%; OR 1.95); Azotemia (L84; 9.1%; OR 1.51); Red-cell anisocytosis with hypoproteinaemia (L77; 9.1%; OR 2.15); Atherosclerotic vascular disease (I276; 9.1%; OR 1.47); Biliary/urothelial abnormalities (I117; 9.1%; OR 1.82).
Drug-resistant infection10 / 45 (22.2%)Cytopenic/nutritional, musculoskeletal and inflammatory findings coexist with vascular and abdominal abnormalities. Cytopenia (L127; 10.0%; OR 1.62); Red-cell anisocytosis with hypoproteinaemia (L77; 10.0%; OR 1.80); Chronic spinal/pelvic abnormalities (I284; 10.0%; OR 1.39); Abdominal abscess/complex fluid collection (I54; 10.0%; OR 1.50); Splenomegaly (I65; 10.0%; OR 1.68); Aortic atherosclerosis (I26; 10.0%; OR 1.63).
OutcomeAdjusted / associated (%)Descriptive profile and representative retained concepts (share; adjusted OR)
Viral pneumonia9 / 20 (45.0%)Hepatic, haematological and nutritional findings coexist with vascular and other imaging abnormalities. Cholestasis (L108; 11.1%; OR 2.51); Cytopenia (L127; 11.1%; OR 2.12); Hypoalbuminaemia (L91; 11.1%; OR 1.91); Aortic atherosclerosis (I26; 11.1%; OR 1.72); Atherosclerosis with colonic abnormalities (I106; 11.1%; OR 1.98); Hiatal/paraesophageal hernia (R1878; 11.1%; OR 1.17).
Gram-positive septicemia7 / 48 (14.6%)Hepatobiliary and urinary findings coexist with anaemia, procedural and other imaging concepts. Unknown hepatobiliary pattern (I87; 14.3%; OR 4.56); Urinary obstruction/intervention (I398; 14.3%; OR 1.60); Low point-of-care haemoglobin (L1182; 14.3%; OR 1.57); Cystic fibrosis imaging pattern (R391; 14.3%; OR 1.26); Indwelling catheter (R75; 14.3%; OR 1.69).
Pneumococcal pneumonia7 / 24 (29.2%)Heterogeneous urinary, vascular and musculoskeletal findings; several concepts include hardware. Bladder/prostate abnormalities (I95; 14.3%; OR 1.77); Atherosclerosis with colonic abnormalities (I106; 14.3%; OR 1.63); Rectal impaction/inflammation or bladder-wall thickening (I452; 14.3%; OR 1.38); Hip pathology (R412; 14.3%; OR 1.33); Orthopaedic hardware (I102; 14.3%; OR 1.45).
Empyema / pneumothorax6 / 22 (27.3%)Malignant, nutritional and postsurgical findings; the endpoint combines infectious and noninfectious conditions. Metastatic disease (R64; 16.7%; OR 1.82); Hypoalbuminaemia (L91; 16.7%; OR 1.63); Pancreatic surgery/severe pancreatitis (R42; 16.7%; OR 1.46); Prior pancreatic/hepatic surgery (I70; 16.7%; OR 1.53); Prior distal pancreatectomy (R407; 16.7%; OR 1.22).
MRSA infection5 / 26 (19.2%)Renal/metabolic and anaemia findings coexist with urinary-obstruction and procedural concepts. Coagulopathy with renal impairment (L87; 20.0%; OR 2.27); Low haemoglobin (L61; 20.0%; OR 2.03); Renal failure with hypernatraemia/acidosis (L62; 20.0%; OR 1.71); Urinary obstruction/intervention (I398; 20.0%; OR 1.44); Inferior vena cava filter (R456; 20.0%; OR 1.35).
Pneumonia5 / 25 (20.0%)Hepatic and haematological abnormalities coexist with vascular disease. Cytopenia (L127; 20.0%; OR 1.37); Cholestasis (L108; 20.0%; OR 1.93); Anaemia with hypoalbuminaemia (L94; 20.0%; OR 1.36); Unknown hepatobiliary pattern (I87; 20.0%; OR 1.36); Extensive abdominopelvic atherosclerosis (I15; 20.0%; OR 1.40).
Bronchopneumonia / lung abscess4 / 10 (40.0%)Renal/metabolic and procedural or mixed concepts; no coherent specific imaging profile. Renal failure with hypernatraemia/acidosis (L62; 25.0%; OR 1.93); Therapeutic/life-support device (I125; 25.0%; OR 2.51); Tube/catheter or anasarca (I35; 25.0%; OR 2.09).
E. coli infection4 / 23 (17.4%)A small, heterogeneous imaging profile spanning vascular, hepatobiliary, urinary and musculoskeletal findings. Aortic calcification (I109; 25.0%; OR 1.61); Biliary/urothelial abnormalities (I117; 25.0%; OR 1.59); Unknown hepatobiliary pattern (I87; 25.0%; OR 1.55); Chronic spinal/pelvic abnormalities (I284; 25.0%; OR 1.38).
Gram-negative septicemia4 / 8 (50.0%)Cholestasis and cytopenia coexist with pancreatic inflammation/fluid and ascites. Cholestasis (L108; 25.0%; OR 2.12); Cytopenia (L127; 25.0%; OR 2.03); Pancreatic inflammation/fluid collection (I59; 25.0%; OR 1.67); Ascites (I21; 25.0%; OR 1.81).
Mycoses3 / 6 (50.0%)Few retained concepts, including catheter-related findings; no coherent physiological profile is established. Adrenal adenoma (R3993; 33.3%; OR 1.14); Central venous catheter (R428; 33.3%; OR 1.36); Indwelling catheter (R75; 33.3%; OR 1.56).
OutcomeAdjusted / associated (%)Descriptive profile and representative retained concepts (share; adjusted OR)
Aspiration pneumonitis3 / 40 (7.5%)Vascular/colonic and lung-base imaging findings; a small retained profile. Atherosclerosis with colonic abnormalities (I106; 33.3%; OR 1.57); Lung-base ground-glass opacity/consolidation (R99; 33.3%; OR 1.41); Lung-base abnormalities (R776; 33.3%; OR 1.40).
Viral enteritis3 / 8 (37.5%)Cholestasis and cytopenia coexist with an orthopaedic-hardware concept. Cholestasis (L108; 33.3%; OR 2.25); Cytopenia (L127; 33.3%; OR 2.05); Orthopaedic hardware (I102; 33.3%; OR 1.45).
Urethral stricture2 / 13 (15.4%)Only two heterogeneous imaging concepts; neither provides a clear infection-specific interpretation. Orthopaedic hardware (I102; 50.0%; OR 2.22); Pancreatic-tail lesion (R136; 50.0%; OR 1.44).
Endocarditis1 / 9 (11.1%)Only an abdominal abscess/complex-fluid concept retained an association. Abdominal abscess/complex fluid collection (I54; 100.0%; OR 1.56).

Table 8. Adjusted concept profiles across infection-related outcomes at Discovery.

Adjusted / associated denotes the number of associations remaining significant after confounder adjustment (Benjamini–Hochberg false-discovery rate below 5%) relative to the number of initially significant associations included in the adjustment analysis; percentages are the corresponding retention rates. Counts include all retained concepts, whereas the descriptions give selected examples. Each example includes its concept code, its percentage share of the retained associations for that outcome, and its adjusted OR. Each concept contributes one association, so its share is 100/NN%, where NN is the number retained; this is a share of associations, not of patients or explained risk. The selected examples need not sum to 100%. ORs are reported on the activation scale used in the analysis and are not percentages of outcome events explained, attributable fractions or mutually independent contributions. Differences in OR alone do not establish differences in predictive importance. Mixed and procedural interpretations are retained where needed to describe the evidence. Profiles are descriptive and do not establish causality or held-out replication. Composite and individual endpoints overlap. Groups follow the study mapping, including noninfectious or anatomically atypical endpoints where assigned.

Outcome frequencies across cohorts

Table 9 lists the 141 outcomes and their positive, negative and excluded counts, followed by the observed case percentage in each cohort. Percentages use the outcome-specific eligible population, comprising studies labelled as progressors or controls under the study’s outcome definition. They are not baseline disease prevalence estimates.

OutcomeDiscoveryInternal TestExternal Site #1External Site #2
Hepatic / biliary
Liver Abscess And Sequelae Of Chronic Liver Disease (571.8)143 /7,109 /8,079 (1.97%)42 /2,392 /2,703 (1.73%)126 /2,558 /23,119 (4.69%)105 /6,308 /7,786 (1.64%)
Acute And Subacute Necrosis Of Liver (573.4)127 /7,387 /7,817 (1.69%)34 /2,472 /2,631 (1.36%)93 /2,598 /23,112 (3.46%)22 /6,405 /7,772 (0.34%)
Disorders Of Bilirubin Excretion (277.4)122 /7,306 /7,903 (1.64%)27 /2,426 /2,684 (1.10%)85 /2,596 /23,122 (3.17%)105 /6,322 /7,772 (1.63%)
Splenomegaly (579.2)111 /7,008 /8,212 (1.56%)50 /2,331 /2,756 (2.10%)120 /2,549 /23,134 (4.50%)30 /6,307 /7,862 (0.47%)
Cirrhosis Of Liver Without Mention Of Alcohol (571.51)110 /7,083 /8,138 (1.53%)39 /2,346 /2,752 (1.64%)147 /2,487 /23,169 (5.58%)137 /6,224 /7,838 (2.15%)
Diseases Of Spleen (289.5)102 /7,227 /8,002 (1.39%)25 /2,428 /2,684 (1.02%)84 /2,567 /23,152 (3.17%)38 /6,367 /7,794 (0.59%)
Jaundice (Not Of Newborn) (573.5)101 /7,223 /8,007 (1.38%)39 /2,414 /2,684 (1.59%)138 /2,575 /23,090 (5.09%)47 /6,370 /7,782 (0.73%)
Portal Hypertension (571.81)97 /7,289 /7,945 (1.31%)25 /2,436 /2,676 (1.02%)126 /2,560 /23,117 (4.69%)141 /6,196 /7,862 (2.23%)
Chronic Passive Congestion Of Liver (573.1)66 /7,561 /7,704 (0.87%)13 /2,553 /2,571 (0.51%)51 /2,628 /23,124 (1.90%)5 /6,415 /7,779 (0.08%)
Liver Replaced By Transplant (573.2)36 /7,496 /7,799 (0.48%)15 /2,486 /2,636 (0.60%)14 /2,624 /23,165 (0.53%)154 /6,171 /7,874 (2.43%)
Liver and intrahepatic bile duct cancer (grouped; 155.0, 155.1)29 /7,333 /7,969 (0.39%)4 /2,450 /2,683 (0.16%)6 /2,579 /23,218 (0.23%)27 /6,236 /7,936 (0.43%)
Renal / urinary
Renal disease (grouped; 250.12, 250.22, 269.0, 401.2, 401.22, 580.11, 580.12, 580.13, 580.14, 580.2, 580.3, 580.31, 580.32, 585.1, 585.2, 585.3, 585.31, 585.32, 585.33, 585.34, 585.4, 588.0, 751.21)579 /4,680 /10,072 (11.01%)187 /1,588 /3,362 (10.54%)868 /1,411 /23,524 (38.09%)764 /4,700 /8,735 (13.98%)
Acute Renal Failure (585.1)525 /5,437 /9,369 (8.81%)163 /1,836 /3,138 (8.15%)789 /1,978 /23,036 (28.51%)405 /5,891 /7,903 (6.43%)
Chronic Kidney Disease, Stage iii (585.33)341 /6,785 /8,205 (4.79%)105 /2,280 /2,752 (4.40%)386 /2,216 /23,201 (14.83%)4 /6,368 /7,827 (0.06%)
Hypertensive Chronic Kidney Disease (401.22)284 /6,526 /8,521 (4.17%)77 /2,226 /2,834 (3.34%)629 /1,943 /23,231 (24.46%)373 /5,608 /8,218 (6.24%)
OutcomeDiscoveryInternal TestExternal Site #1External Site #2
Chronic Renal Failure [CKD] (585.3)262 /6,499 /8,570 (3.88%)90 /2,198 /2,849 (3.93%)509 /2,241 /23,053 (18.51%)185 /5,926 /8,088 (3.03%)
Hyperplasia Of Prostate (600.0)187 /6,674 /8,470 (2.73%)60 /2,261 /2,816 (2.59%)377 /2,083 /23,343 (15.33%)290 /5,822 /8,087 (4.74%)
Anemia In Chronic Kidney Disease (285.21)120 /7,189 /8,022 (1.64%)49 /2,383 /2,705 (2.01%)487 /2,335 /22,981 (17.26%)227 /6,128 /7,844 (3.57%)
Chronic Kidney Disease, Stage Iv (585.34)93 /7,380 /7,858 (1.24%)47 /2,473 /2,617 (1.87%)191 /2,536 /23,076 (7.00%)120 /6,225 /7,854 (1.89%)
Oliguria And Anuria (599.6)72 /7,434 /7,825 (0.96%)32 /2,470 /2,635 (1.28%)99 /2,611 /23,093 (3.65%)22 /6,388 /7,789 (0.34%)
Secondary Hyperparathyroidism (Of Renal Origin) (588.2)64 /7,471 /7,796 (0.85%)27 /2,509 /2,601 (1.06%)154 /2,511 /23,138 (5.78%)63 /6,306 /7,830 (0.99%)
Renal Sclerosis, Nos (580.4)45 /7,493 /7,793 (0.60%)27 /2,505 /2,605 (1.07%)60 /2,620 /23,123 (2.24%)68 /6,345 /7,786 (1.06%)
Renal Dialysis (585.31)39 /7,435 /7,857 (0.52%)13 /2,494 /2,630 (0.52%)112 /2,588 /23,103 (4.15%)96 /6,299 /7,804 (1.50%)
Kidney Replaced By Transplant (587.0)30 /7,517 /7,784 (0.40%)9 /2,523 /2,605 (0.36%)8 /2,623 /23,172 (0.30%)158 /6,259 /7,782 (2.46%)
Cardiovascular
Hypertensive diseases (grouped; 401.0, 401.1, 401.2, 401.21, 401.22, 401.3)498 /2,949 /11,884 (14.45%)162 /989 /3,986 (14.07%)676 /755 /24,372 (47.24%)1,154 /3,190 /9,855 (26.57%)
Ischaemic heart disease (grouped; 411.1, 411.2, 411.3, 411.4, 411.8, 411.9)671 /4,944 /9,716 (11.95%)192 /1,672 /3,273 (10.30%)1,001 /1,590 /23,212 (38.63%)533 /5,114 /8,552 (9.44%)
Myocardial Infarction (411.2)512 /6,012 /8,807 (7.85%)142 /2,029 /2,966 (6.54%)788 /2,042 /22,973 (27.84%)160 /6,004 /8,035 (2.60%)
Systemic atherosclerosis (grouped; 411.4, 433.12, 440.0, 440.1, 440.2, 440.21, 440.22, 440.9)427 /5,448 /9,456 (7.27%)157 /1,815 /3,165 (7.96%)838 /1,676 /23,289 (33.33%)627 /4,970 /8,602 (11.20%)
Premature Beats (427.6)419 /6,025 /8,887 (6.50%)159 /1,961 /3,017 (7.50%)784 /2,017 /23,002 (27.99%)144 /6,043 /8,012 (2.33%)
Coronary Atherosclerosis (411.4)378 /5,778 /9,175 (6.14%)151 /1,943 /3,043 (7.21%)739 /1,848 /23,216 (28.57%)475 /5,247 /8,477 (8.30%)
Supraventricular Premature Beats (427.61)398 /6,194 /8,739 (6.04%)119 /2,038 /2,980 (5.52%)673 /2,163 /22,967 (23.73%)94 /6,185 /7,920 (1.50%)
Atrial Fibrillation (427.21)307 /6,224 /8,800 (4.70%)86 /2,142 /2,909 (3.86%)510 /2,130 /23,163 (19.32%)344 /5,579 /8,276 (5.81%)
Hypertensive Heart Disease (401.21)274 /6,818 /8,239 (3.86%)100 /2,320 /2,717 (4.13%)480 /2,320 /23,003 (17.14%)263 /5,893 /8,043 (4.27%)
Peripheral artery disease (grouped; 440.1, 440.2, 440.21, 440.22, 440.9, 443.9)242 /6,600 /8,489 (3.54%)69 /2,259 /2,809 (2.96%)479 /2,152 /23,172 (18.21%)298 /5,764 /8,137 (4.92%)
Congestive Heart Failure (Chf) Nos (428.1)240 /6,594 /8,497 (3.51%)100 /2,212 /2,825 (4.33%)605 /2,250 /22,948 (21.19%)176 /5,943 /8,080 (2.88%)
OutcomeDiscoveryInternal TestExternal Site #1External Site #2
Cerebrovascular disease (grouped; 290.16, 430.2, 433.11, 433.12, 433.21)234 /6,794 /8,303 (3.33%)84 /2,327 /2,726 (3.48%)431 /2,367 /23,005 (15.40%)79 /6,224 /7,896 (1.25%)
Heart Failure With Preserved Ef [Diastolic Heart Failure] (428.4)221 /7,021 /8,089 (3.05%)65 /2,356 /2,716 (2.68%)529 /2,335 /22,939 (18.47%)277 /5,913 /8,009 (4.47%)
Heart Failure With Reduced Ef [Systolic Or Combined Heart Failure] (428.3)209 /6,879 /8,243 (2.95%)81 /2,299 /2,757 (3.40%)488 /2,345 /22,970 (17.23%)190 /6,065 /7,944 (3.04%)
Blood Vessel Replaced (459.7)198 /7,137 /7,996 (2.70%)79 /2,366 /2,692 (3.23%)299 /2,461 /23,043 (10.83%)110 /6,253 /7,836 (1.73%)
Paroxysmal Ventricular Tachycardia (427.12)189 /7,090 /8,052 (2.60%)53 /2,384 /2,700 (2.17%)81 /2,514 /23,208 (3.12%)10 /6,372 /7,817 (0.16%)
Atrial Flutter (427.22)178 /7,108 /8,045 (2.44%)43 /2,387 /2,707 (1.77%)187 /2,467 /23,149 (7.05%)42 /6,241 /7,916 (0.67%)
First Degree Av Block (426.21)112 /6,975 /8,244 (1.58%)41 /2,349 /2,747 (1.72%)664 /2,431 /22,708 (21.45%)98 /6,130 /7,971 (1.57%)
Varicose Veins (454.0)107 /7,306 /7,918 (1.44%)34 /2,451 /2,652 (1.37%)66 /2,603 /23,134 (2.47%)21 /6,377 /7,801 (0.33%)
Rheumatic Disease Of The Heart Valves (394.0)105 /7,347 /7,879 (1.41%)33 /2,480 /2,624 (1.31%)316 /2,282 /23,205 (12.16%)194 /6,175 /7,830 (3.05%)
Disease Of Tricuspid Valve (394.7)88 /7,339 /7,904 (1.18%)31 /2,434 /2,672 (1.26%)311 /2,457 /23,035 (11.24%)137 /6,236 /7,826 (2.15%)
Cardiac Arrest (427.42)76 /7,507 /7,748 (1.00%)19 /2,524 /2,594 (0.75%)170 /2,596 /23,037 (6.15%)37 /6,385 /7,777 (0.58%)
Arterial Embolism And Thrombosis (444.0)61 /7,396 /7,874 (0.82%)15 /2,471 /2,651 (0.60%)69 /2,628 /23,106 (2.56%)51 /6,341 /7,807 (0.80%)
Mitral Valve Disease (394.2)55 /7,263 /8,013 (0.75%)19 /2,460 /2,658 (0.77%)76 /2,575 /23,152 (2.87%)8 /6,413 /7,778 (0.12%)
Cardiogenic Shock (797.1)53 /7,485 /7,793 (0.70%)16 /2,505 /2,616 (0.63%)118 /2,617 /23,068 (4.31%)25 /6,384 /7,790 (0.39%)
Aneurysm Of Other Specified Artery (442.8)42 /7,453 /7,836 (0.56%)8 /2,522 /2,607 (0.32%)72 /2,608 /23,123 (2.69%)33 /6,341 /7,825 (0.52%)
Abdominal Aortic Aneurysm (442.11)37 /7,542 /7,752 (0.49%)10 /2,559 /2,568 (0.39%)36 /2,579 /23,188 (1.38%)5 /6,359 /7,835 (0.08%)
Orthopnea (513.32)31 /7,603 /7,697 (0.41%)3 /2,549 /2,585 (0.12%)34 /2,629 /23,140 (1.28%)6 /6,395 /7,798 (0.09%)
Neuro / cognitive
Altered Mental Status (292.4)556 /6,074 /8,701 (8.39%)160 /2,017 /2,960 (7.35%)809 /2,256 /22,738 (26.39%)193 /6,080 /7,926 (3.08%)
Dementia and cognitive disorders (grouped; 290.0, 290.1, 290.11, 290.12, 290.16, 292.2, 292.3)436 /6,246 /8,649 (6.52%)153 /2,121 /2,863 (6.73%)824 /2,172 /22,807 (27.50%)214 /5,859 /8,126 (3.52%)
Delirium Due To Conditions Classified Elsewhere (290.2)284 /6,937 /8,110 (3.93%)71 /2,305 /2,761 (2.99%)383 /2,505 /22,915 (13.26%)64 /6,360 /7,775 (1.00%)
Dementias (290.1)185 /7,231 /7,915 (2.49%)74 /2,419 /2,644 (2.97%)337 /2,480 /22,986 (11.96%)70 /6,262 /7,867 (1.11%)
Coma (348.7)171 /7,367 /7,793 (2.27%)47 /2,460 /2,630 (1.87%)87 /2,600 /23,116 (3.24%)14 /6,399 /7,786 (0.22%)
OutcomeDiscoveryInternal TestExternal Site #1External Site #2
Delirium Dementia And Amnestic And Other Cognitive Disorders (290.0)115 /7,287 /7,929 (1.55%)39 /2,447 /2,651 (1.57%)430 /2,521 /22,852 (14.57%)57 /6,224 /7,918 (0.91%)
Transient Alteration Of Awareness (292.5)91 /7,122 /8,118 (1.26%)35 /2,345 /2,757 (1.47%)144 /2,573 /23,086 (5.30%)1 /6,415 /7,783 (0.02%)
Alzheimer’s Disease (290.11)64 /7,543 /7,724 (0.84%)23 /2,531 /2,583 (0.90%)93 /2,600 /23,110 (3.45%)25 /6,381 /7,793 (0.39%)
Vascular Dementia (290.16)50 /7,624 /7,657 (0.65%)16 /2,560 /2,561 (0.62%)20 /2,638 /23,145 (0.75%)9 /6,420 /7,770 (0.14%)
Primary Open Angle Glaucoma (365.11)43 /7,546 /7,742 (0.57%)12 /2,523 /2,602 (0.47%)29 /2,621 /23,153 (1.09%)19 /6,334 /7,846 (0.30%)
Strabismus And Other Disorders Of Binocular Eye Movements (378.0)37 /7,557 /7,737 (0.49%)7 /2,556 /2,574 (0.27%)30 /2,634 /23,139 (1.13%)4 /6,404 /7,791 (0.06%)
Paralysis/Spasm Of Vocal Cords Or Larynx (473.3)34 /7,549 /7,748 (0.45%)15 /2,533 /2,589 (0.59%)37 /2,630 /23,136 (1.39%)8 /6,400 /7,791 (0.12%)
Infection and inflammation
Sepsis (994.2)470 /5,504 /9,357 (7.87%)139 /1,889 /3,109 (6.85%)623 /2,117 /23,063 (22.74%)228 /6,073 /7,898 (3.62%)
Pneumonia (480.0)437 /5,824 /9,070 (6.98%)163 /1,850 /3,124 (8.10%)1,038 /2,006 /22,759 (34.10%)281 /5,774 /8,144 (4.64%)
E. Coli (41.4)280 /6,944 /8,107 (3.88%)95 /2,350 /2,692 (3.89%)255 /2,468 /23,080 (9.36%)50 /6,325 /7,824 (0.78%)
Peritoneal Adhesions (Postoperative) (Postinfection) (568.1)257 /6,430 /8,644 (3.84%)98 /2,149 /2,890 (4.36%)172 /2,474 /23,157 (6.50%)86 /6,300 /7,813 (1.35%)
Bacteremia (38.3)250 /6,763 /8,318 (3.56%)74 /2,181 /2,882 (3.28%)284 /2,451 /23,068 (10.38%)96 /6,233 /7,870 (1.52%)
Septic Shock (994.21)247 /6,934 /8,150 (3.44%)77 /2,247 /2,813 (3.31%)290 /2,533 /22,980 (10.27%)57 /6,294 /7,848 (0.90%)
Pneumonitis Due To Inhalation Of Food Or Vomitus (501.0)201 /7,092 /8,038 (2.76%)67 /2,391 /2,679 (2.73%)424 /2,474 /22,905 (14.63%)68 /6,340 /7,791 (1.06%)
Pneumococcal pneumonia (480.11)165 /7,397 /7,769 (2.18%)57 /2,485 /2,595 (2.24%)144 /2,529 /23,130 (5.39%)32 /6,400 /7,767 (0.50%)
Empyema And Pneumothorax (506.0)153 /7,105 /8,073 (2.11%)63 /2,404 /2,670 (2.55%)214 /2,420 /23,169 (8.12%)14 /6,359 /7,826 (0.22%)
Infection With Drug-Resistant Microorganisms (41.9)151 /7,181 /7,999 (2.06%)53 /2,375 /2,709 (2.18%)214 /2,524 /23,065 (7.82%)80 /6,289 /7,830 (1.26%)
Bacterial pneumonia (480.1)127 /7,345 /7,859 (1.70%)41 /2,461 /2,635 (1.64%)219 /2,537 /23,047 (7.95%)59 /6,343 /7,797 (0.92%)
Viral pneumonia (480.2)111 /7,428 /7,792 (1.47%)36 /2,508 /2,593 (1.42%)94 /2,598 /23,111 (3.49%)25 /6,400 /7,774 (0.39%)
Viral Enteritis (8.6)107 /7,245 /7,979 (1.46%)38 /2,463 /2,636 (1.52%)79 /2,599 /23,125 (2.95%)36 /6,331 /7,832 (0.57%)
Methicillin Resistant Staphylococcus Aureus (41.12)94 /7,281 /7,956 (1.27%)35 /2,433 /2,669 (1.42%)83 /2,585 /23,135 (3.11%)34 /6,356 /7,809 (0.53%)
Methicillin Sensitive Staphylococcus Aureus (41.11)83 /7,375 /7,873 (1.11%)25 /2,430 /2,682 (1.02%)76 /2,590 /23,137 (2.85%)25 /6,365 /7,809 (0.39%)
OutcomeDiscoveryInternal TestExternal Site #1External Site #2
Endocarditis (420.3)81 /7,359 /7,891 (1.09%)33 /2,455 /2,649 (1.33%)91 /2,594 /23,118 (3.39%)9 /6,396 /7,794 (0.14%)
Mycoses (117.0)70 /7,390 /7,871 (0.94%)20 /2,437 /2,680 (0.81%)63 /2,610 /23,130 (2.36%)6 /6,411 /7,782 (0.09%)
Gram Positive Septicemia (38.2)71 /7,543 /7,717 (0.93%)26 /2,518 /2,593 (1.02%)60 /2,613 /23,130 (2.24%)8 /6,381 /7,810 (0.13%)
Gram Negative Septicemia (38.1)66 /7,491 /7,774 (0.87%)12 /2,490 /2,635 (0.48%)72 /2,612 /23,119 (2.68%)7 /6,397 /7,795 (0.11%)
Chronic Osteomyelitis (710.12)48 /7,478 /7,805 (0.64%)25 /2,506 /2,606 (0.99%)59 /2,611 /23,133 (2.21%)20 /6,383 /7,796 (0.31%)
Bronchopneumonia and lung abscess (480.5)28 /7,587 /7,716 (0.37%)22 /2,529 /2,586 (0.86%)36 /2,626 /23,141 (1.35%)24 /6,410 /7,765 (0.37%)
Urethral stricture (597.1)26 /7,620 /7,685 (0.34%)2 /2,536 /2,599 (0.08%)13 /2,629 /23,161 (0.49%)0 /6,392 /7,807 (0.00%)
Haematologic
Iron Deficiency Anemias, Unspecified Or Not Due To Blood Loss (280.1)512 /5,474 /9,345 (8.55%)177 /1,841 /3,119 (8.77%)855 /1,975 /22,973 (30.21%)298 /5,879 /8,022 (4.82%)
Anemia Of Chronic Disease (285.2)332 /6,609 /8,390 (4.78%)146 /2,163 /2,828 (6.32%)448 /2,397 /22,958 (15.75%)106 /6,238 /7,855 (1.67%)
Pancytopenia (284.1)203 /7,092 /8,036 (2.78%)66 /2,375 /2,696 (2.70%)227 /2,510 /23,066 (8.29%)121 /6,298 /7,780 (1.89%)
Neutropenia (288.11)168 /6,995 /8,168 (2.35%)59 /2,356 /2,722 (2.44%)172 /2,374 /23,257 (6.76%)85 /6,345 /7,769 (1.32%)
Secondary Thrombocytopenia (287.32)171 /7,220 /7,940 (2.31%)64 /2,400 /2,673 (2.60%)168 /2,471 /23,164 (6.37%)74 /6,311 /7,814 (1.16%)
Acquired Coagulation Factor Deficiency (286.4)149 /7,381 /7,801 (1.98%)26 /2,462 /2,649 (1.05%)192 /2,570 /23,041 (6.95%)96 /6,341 /7,762 (1.49%)
Primary Thrombocytopenia (287.31)81 /7,404 /7,846 (1.08%)22 /2,498 /2,617 (0.87%)118 /2,564 /23,121 (4.40%)6 /6,397 /7,796 (0.09%)
Metabolic / nutritional
Metabolic lipid disorders (grouped; 272.1, 272.11, 272.13)601 /3,415 /11,315 (14.97%)192 /1,147 /3,798 (14.34%)860 /702 /24,241 (55.06%)1,059 /3,618 /9,522 (22.64%)
Hypopotassemia (276.14)446 /5,318 /9,567 (7.74%)157 /1,770 /3,210 (8.15%)1,060 /1,947 /22,796 (35.25%)346 /5,763 /8,090 (5.66%)
Hyposmolality And/Or Hyponatremia (276.12)426 /5,677 /9,228 (6.98%)136 /1,923 /3,078 (6.61%)778 /1,996 /23,029 (28.05%)341 /5,948 /7,910 (5.42%)
Severe Protein-Calorie Malnutrition (260.2)450 /6,517 /8,364 (6.46%)161 /2,174 /2,802 (6.90%)501 /2,402 /22,900 (17.26%)91 /6,278 /7,830 (1.43%)
Protein-Calorie Malnutrition (260.0)414 /6,115 /8,802 (6.34%)125 /2,049 /2,963 (5.75%)556 /2,379 /22,868 (18.94%)165 /6,083 /7,951 (2.64%)
Acidosis (276.41)437 /6,491 /8,403 (6.31%)136 /2,171 /2,830 (5.90%)189 /2,303 /23,311 (7.58%)34 /6,294 /7,871 (0.54%)
Hyperpotassemia (276.13)278 /6,714 /8,339 (3.98%)94 /2,283 /2,760 (3.95%)440 /2,403 /22,960 (15.48%)145 /6,050 /8,004 (2.34%)
Fluid Overload (276.6)223 /6,858 /8,250 (3.15%)66 /2,299 /2,772 (2.79%)346 /2,486 /22,971 (12.22%)181 /6,241 /7,777 (2.82%)
OutcomeDiscoveryInternal TestExternal Site #1External Site #2
Cachexia (260.1)222 /7,263 /7,846 (2.97%)56 /2,443 /2,638 (2.24%)219 /2,590 /22,994 (7.80%)57 /6,384 /7,758 (0.88%)
Adult Failure To Thrive (260.3)213 /6,974 /8,144 (2.96%)89 /2,297 /2,751 (3.73%)372 /2,556 /22,875 (12.70%)86 /6,338 /7,775 (1.34%)
Diabetes Mellitus (250.0)153 /7,190 /7,988 (2.08%)51 /2,413 /2,673 (2.07%)134 /2,526 /23,143 (5.04%)32 /6,384 /7,783 (0.50%)
Hypocalcemia (275.51)139 /7,018 /8,174 (1.94%)66 /2,287 /2,784 (2.80%)572 /2,460 /22,771 (18.87%)142 /6,140 /7,917 (2.26%)
Hyperosmolality And/Or Hypernatremia (276.11)110 /7,294 /7,927 (1.49%)49 /2,401 /2,687 (2.00%)349 /2,524 /22,930 (12.15%)48 /6,353 /7,798 (0.75%)
Localized Adiposity (278.3)40 /7,537 /7,754 (0.53%)12 /2,524 /2,601 (0.47%)24 /2,634 /23,145 (0.90%)0 /6,407 /7,792 (0.00%)
Type 2 Diabetes With Ketoacidosis (250.21)28 /7,572 /7,731 (0.37%)9 /2,537 /2,591 (0.35%)34 /2,617 /23,152 (1.28%)38 /6,381 /7,780 (0.59%)
Musculoskeletal / trauma
Fracture Of Vertebral Column Without Mention Of Spinal Cord Injury (805.0)160 /7,038 /8,133 (2.22%)53 /2,395 /2,689 (2.17%)299 /2,474 /23,030 (10.78%)62 /6,268 /7,869 (0.98%)
Crushing Or Internal Injury To Organs (1008.0)80 /7,297 /7,954 (1.08%)36 /2,445 /2,656 (1.45%)92 /2,583 /23,128 (3.44%)18 /6,376 /7,805 (0.28%)
Pathologic Fracture Of Vertebrae (743.21)77 /7,438 /7,816 (1.02%)30 /2,489 /2,618 (1.19%)170 /2,564 /23,069 (6.22%)29 /6,348 /7,822 (0.45%)
Fracture Of Neck Of Femur (800.1)73 /7,423 /7,835 (0.97%)24 /2,497 /2,616 (0.95%)98 /2,603 /23,102 (3.63%)27 /6,368 /7,804 (0.42%)
Fracture Of Unspecified Part Of Femur (800.2)55 /7,503 /7,773 (0.73%)16 /2,529 /2,592 (0.63%)78 /2,605 /23,120 (2.91%)38 /6,369 /7,792 (0.59%)
Chondrocalcinosis (274.21)54 /7,487 /7,790 (0.72%)16 /2,523 /2,598 (0.63%)60 /2,616 /23,127 (2.24%)26 /6,362 /7,811 (0.41%)
Rupture Of Synovium (727.5)49 /7,453 /7,829 (0.65%)23 /2,502 /2,612 (0.91%)84 /2,579 /23,140 (3.15%)10 /6,355 /7,834 (0.16%)
Fracture Of Humerus (803.1)47 /7,489 /7,795 (0.62%)10 /2,525 /2,602 (0.39%)115 /2,582 /23,106 (4.26%)18 /6,366 /7,815 (0.28%)
Kyphosis (Acquired) (737.1)40 /7,557 /7,734 (0.53%)21 /2,509 /2,607 (0.83%)65 /2,613 /23,125 (2.43%)19 /6,397 /7,783 (0.30%)
Malunion And Nonunion Of Fracture (733.8)29 /7,564 /7,738 (0.38%)8 /2,539 /2,590 (0.31%)23 /2,627 /23,153 (0.87%)0 /6,376 /7,823 (0.00%)
Oncologic
Chemotherapy (197.0)246 /6,630 /8,455 (3.58%)95 /2,185 /2,857 (4.17%)302 /2,354 /23,147 (11.37%)102 /6,270 /7,827 (1.60%)
Effects Radiation Nos (990.0)225 /6,817 /8,289 (3.20%)80 /2,250 /2,807 (3.43%)420 /2,353 /23,030 (15.15%)95 /6,238 /7,866 (1.50%)
Antineoplastic And Immunosuppressive Drugs Causing Adverse Effects (963.1)215 /6,881 /8,235 (3.03%)70 /2,259 /2,808 (3.01%)290 /2,400 /23,113 (10.78%)95 /6,316 /7,788 (1.48%)
Anemia In Neoplastic Disease (285.22)218 /7,210 /7,903 (2.93%)77 /2,390 /2,670 (3.12%)339 /2,485 /22,979 (12.00%)47 /6,347 /7,805 (0.74%)
Respiratory
OutcomeDiscoveryInternal TestExternal Site #1External Site #2
Abnormal Findings Examination Of Lungs (514.0)640 /3,287 /11,404 (16.30%)224 /1,113 /3,800 (16.75%)1,126 /1,075 /23,602 (51.16%)517 /5,365 /8,317 (8.79%)
Pleurisy; Pleural Effusion (507.0)427 /5,251 /9,653 (7.52%)150 /1,752 /3,235 (7.89%)793 /1,958 /23,052 (28.83%)265 /5,977 /7,957 (4.25%)
Pulmonary Collapse; Interstitial And Compensatory Emphysema (508.0)416 /5,438 /9,477 (7.11%)154 /1,833 /3,150 (7.75%)826 /1,932 /23,045 (29.95%)343 /5,878 /7,978 (5.51%)
Respiratory Failure (509.1)384 /6,390 /8,557 (5.67%)152 /2,108 /2,877 (6.73%)790 /2,226 /22,787 (26.19%)214 /5,986 /7,999 (3.45%)
Asphyxia And Hypoxemia (1013.0)335 /6,405 /8,591 (4.97%)112 /2,110 /2,915 (5.04%)582 /2,300 /22,921 (20.19%)255 /6,093 /7,851 (4.02%)
Pulmonary Congestion And Hypostasis (503.0)319 /6,429 /8,583 (4.73%)109 /2,152 /2,876 (4.82%)486 /2,362 /22,955 (17.06%)69 /6,296 /7,834 (1.08%)
Chronic respiratory disease (grouped; 496.1, 496.2, 496.21, 496.3)226 /6,784 /8,321 (3.22%)74 /2,281 /2,782 (3.14%)430 /2,168 /23,205 (16.55%)117 /5,946 /8,136 (1.93%)
Dependence On Respirator [Ventilator] Or Supplemental Oxygen (509.8)177 /7,078 /8,076 (2.44%)54 /2,401 /2,682 (2.20%)359 /2,493 /22,951 (12.59%)49 /6,281 /7,869 (0.77%)
Chronic Airway Obstruction (496.0)164 /6,864 /8,303 (2.33%)67 /2,292 /2,778 (2.84%)400 /2,184 /23,219 (15.48%)184 /5,843 /8,172 (3.05%)
Respiratory Insufficiency (509.2)111 /7,260 /7,960 (1.51%)39 /2,407 /2,691 (1.59%)219 /2,556 /23,028 (7.89%)27 /6,388 /7,784 (0.42%)
Central/Nonobstructive Sleep Apnea (327.31)44 /7,460 /7,827 (0.59%)10 /2,527 /2,600 (0.39%)44 /2,583 /23,176 (1.67%)46 /6,307 /7,846 (0.72%)
Other
Acquired Absence Of Organs (1090.0)770 /3,520 /11,041 (17.95%)273 /1,152 /3,712 (19.16%)839 /1,449 /23,515 (36.67%)456 /5,476 /8,267 (7.69%)
Ventral Hernia (550.5)327 /6,568 /8,436 (4.74%)101 /2,207 /2,829 (4.38%)241 /2,439 /23,123 (8.99%)106 /6,227 /7,866 (1.67%)
Gastrointestinal Complications (569.2)232 /6,407 /8,692 (3.49%)67 /2,145 /2,925 (3.03%)205 /2,451 /23,147 (7.72%)106 /6,140 /7,953 (1.70%)
Incisional Hernia (550.6)184 /7,008 /8,139 (2.56%)79 /2,314 /2,744 (3.30%)121 /2,541 /23,141 (4.55%)43 /6,320 /7,836 (0.68%)
Decubitus Ulcer (707.1)143 /7,253 /7,935 (1.93%)56 /2,406 /2,675 (2.27%)299 /2,519 /22,985 (10.61%)77 /6,253 /7,869 (1.22%)
Chronic Ulcer Of Leg Or Foot (707.2)103 /7,349 /7,879 (1.38%)28 /2,480 /2,629 (1.12%)125 /2,562 /23,116 (4.65%)45 /6,264 /7,890 (0.71%)
Disorders Of Esophageal Motility (530.5)67 /7,270 /7,994 (0.91%)27 /2,464 /2,646 (1.08%)72 /2,580 /23,151 (2.71%)40 /6,379 /7,780 (0.62%)
Poisoning By Hormones And Synthetic Substitutes (962.0)54 /7,475 /7,802 (0.72%)17 /2,516 /2,604 (0.67%)56 /2,608 /23,139 (2.10%)14 /6,398 /7,787 (0.22%)

Table 9. Outcome case frequencies by cohort, grouped by outcome category.

Cells give positive / negative / excluded counts, followed by the positive percentage in parentheses. Percentages are calculated as 100×n/N100 \times n/N, where nn is the number of studies labelled as future cases and NN is the number with a valid case or control label for that outcome. Progressors had a first recorded diagnosis 1–4 years after CT; controls had no recorded diagnosis and at least four years of follow-up. All other studies were removed from the outcome-specific denominator. Related and grouped outcomes can overlap and should not be counted as independent diseases; individual Phecode endpoints are not assumed to include all clinically related subtypes. The companion CSV supplies all numerators, denominators and excluded counts. Group assignments follow the study outcome mapping.

Laboratory input types

Table 10 lists the 350 retained numerical input types and the descriptive statistics available in the preprocessing inventory. 350 types were retained from 14,890 source types, covering 77,502,851 recorded measurements (76.1% of all laboratory results in the Discovery Site, Internal Test and External Site #1 combined). Each retained type had at least 90% numeric measurements. Statistics summarize recorded measurements rather than unique patients and are not specific to a study cohort.

Measurements is the number of source records for the type before nonnumeric values were removed; Numeric is the percentage parseable as numbers. Mean, standard deviation (SD) and interquartile range (IQR; 75th minus 25th percentile) describe the numeric values retained during preprocessing, before model standardisation. Mean, SD and IQR are displayed to three significant figures. Values and units are reproduced from the existing inventory without new harmonisation; measurement frequency does not represent patient prevalence. Types with numeric encodings or measurement-related meanings should not be interpreted as continuous biological analytes. The inventory does not provide patient-level coverage, cohort-specific distributions, medians or reference intervals. Minimum and maximum values and source-measurement proportions are provided in the companion CSV. Laboratory inputs were not used at External Site #2.

References

  1. [1]Wang, H. et al. Scientific discovery in the age of artificial intelligence. Nature 620, 47–60 (2023).
  2. [2]Denny, J. C. et al. Systematic comparison of phenome-wide association study of electronic medical record data and genome-wide association study data. Nat. biotechnology 31, 1102–1111 (2013).
  3. [3]Bai, W. et al. A population-based phenome-wide association study of cardiac and aortic structure and function. Nat. medicine 26, 1654–1662 (2020).DOI
  4. [4]Gottweis, J. et al. Towards an ai co-scientist. arXiv preprint arXiv:2502.18864 (2025).
  5. [6]Sedlakova, J. et al. Challenges and best practices for digital unstructured data enrichment in health research: A systematic narrative review. PLOS digital health 2, e0000347 (2023).
  6. [7]Xiao, Q. et al. Viral oncogenesis in cancer: from mechanisms to therapeutics. Signal transduction targeted therapy 10, 151 (2025).
  7. [8]Schöckel, L. et al. Developments in x-ray contrast media and the potential impact on computed tomography. Investig. radiology 55, 592–597 (2020).
  8. [9]Jumper, J. et al. Highly accurate protein structure prediction with alphafold. nature 596, 583–589 (2021).
  9. [10]Varadi, M. et al. Alphafold protein structure database: massively expanding the structural coverage of protein-sequence space with high-accuracy models. Nucleic acids research 50, D439–D444 (2022).
  10. [11]Merchant, A. et al. Scaling deep learning for materials discovery. Nature 624, 80–85 (2023).
  11. [12]Jensen, A. B. et al. Temporal disease trajectories condensed from population-wide registry data covering 6.2 million patients. Nat. communications 5, 4022 (2014).DOI
  12. [13]Gottweis, J. et al. Accelerating scientific discovery with co-scientist. Nature 1–3 (2026).
  13. [14]Blankemeier, L. et al. Merlin: a computed tomography vision–language foundation model and dataset. Nature 652, 1318–1328 (2026).
  14. [15]Moor, M. et al. Foundation models for generalist medical artificial intelligence. Nature 616, 259–265 (2023).DOI
  15. [16]Zhou, Y. et al. A foundation model for generalizable disease detection from retinal images. Nature 622, 156–163 (2023).DOI
  16. [17]Ding, T. et al. A multimodal whole-slide foundation model for pathology. Nat. medicine 31, 3749–3761 (2025).DOI
  17. [18]Bricken, T. et al. Towards monosemanticity: Decomposing language models with dictionary learning. Transform. Circuits Thread (2023). https://transformer-circuits.pub/2023/monosemantic-features/index.html.
  18. [19]Gao, L. et al. Scaling and evaluating sparse autoencoders. arXiv preprint arXiv:2406.04093 (2024).arxiv.org/abs/2406.04093
  19. [20]Templeton, A. et al. Scaling monosemanticity: Extracting interpretable features from claude 3 sonnet. Transform. Circuits Thread (2024).
  20. [21]Zhang, Y. et al. Qwen3 embedding: Advancing text embedding and reranking through foundation models. arXiv preprint arXiv:2506.05176 (2025).arxiv.org/abs/2506.05176
  21. [22]Schuppan, D. & Afdhal, N. H. Liver cirrhosis. The Lancet 371, 838–851 (2008).
  22. [23]Vaziri, N. Dyslipidemia of chronic renal failure: the nature, mechanisms, and potential consequences. Am. J. Physiol. Physiol. (2006).
  23. [24]Oliveira, C. B. et al. Co-occurrence of chronic musculoskeletal pain and cardiovascular diseases: a systematic review with meta-analysis. Pain Medicine 21, 1106–1121 (2020).DOI
  24. [25]Vogel, A., Meyer, T., Sapisochin, G., Salem, R. & Saborowski, A. Hepatocellular carcinoma. The Lancet 400, 1345–1362 (2022).
  25. [26]De Franchis, R. et al. Baveno vii–renewing consensus in portal hypertension. J. hepatology 76, 959–974 (2022).
  26. [27]Maiwall, R. et al. Incidence, risk factors, and outcomes of transition of acute kidney injury to chronic kidney disease in cirrhosis: a prospective cohort study. Hepatology 71, 1009–1022 (2020).DOI
  27. [28]Nahon, P. et al. Bacterial infection in compensated viral cirrhosis impairs 5-year survival (anrs col12 cirr prognostic cohort). Gut 66, 330–341 (2017).
  28. [29]Foreman, M. G., Mannino, D. M. & Moss, M. Cirrhosis as a risk factor for sepsis and death: analysis of the national hospital discharge survey. Chest 124, 1016–1020 (2003).
  29. [30]Li, T. & Apte, U. Bile acid metabolism and signaling in cholestasis, inflammation, and cancer. In Advances in pharmacology, vol. 74, 263–302 (Elsevier, 2015).DOI
  30. [31]Stevens, P. E. et al. Kdigo 2024 clinical practice guideline for the evaluation and management of chronic kidney disease. Kidney international 105, S117–S314 (2024).
  31. [32]Kirtane, A. J. et al. Serum blood urea nitrogen as an independent marker of subsequent mortality among patients with acute coronary syndromes and normal to mildly reduced glomerular filtration rates. J. Am. Coll. Cardiol. 45, 1781–1786 (2005).
  32. [33]Grenier, N., Merville, P. & Combe, C. Radiologic imaging of the renal parenchyma, structure and function. Nat. Rev. Nephrol. 12, 348–359 (2016).
  33. [34]Leow, K. et al. Prognostic value of abdominal aortic calcification: a systematic review and meta-analysis of observational studies. J. Am. Hear. Assoc. 10, e017205 (2021).DOI
  34. [35]Detrano, R. et al. Coronary calcium as a predictor of coronary events in four racial or ethnic groups. New Engl. J. Medicine 358, 1336–1345 (2008).DOI
  35. [36]Fox, C. S. et al. Mitral annular calcification predicts cardiovascular morbidity and mortality: the framingham heart study. Circulation 107, 1492–1496 (2003).
  36. [37]Pickhardt, P. J. et al. Automated ct biomarkers for opportunistic prediction of future cardiovascular events and mortality in an asymptomatic screening population: a retrospective cohort study. The Lancet Digit. Heal. 2, e192–e200 (2020).DOI
  37. [38]Bos, D. et al. Atherosclerotic calcification is related to a higher risk of dementia and cognitive decline. Alzheimer’s & Dementia 11, 639–647 (2015).DOI
  38. [39]Schulz, E., Arfai, K., Liu, X., Sayre, J. & Gilsanz, V. Aortic calcification and the risk of osteoporosis and fractures. The J. Clin. Endocrinol. & Metab. 89, 4246–4253 (2004).
  39. [40]Libby, P. The changing landscape of atherosclerosis. Nature 592, 524–533 (2021).DOI
  40. [41]Dhingra, R. et al. Electrocardiographic qrs duration and the risk of congestive heart failure: the framingham heart study. Hypertension 47, 861–867 (2006).
  41. [42]Ference, B. A. et al. Low-density lipoproteins cause atherosclerotic cardiovascular disease. 1. evidence from genetic, epidemiologic, and clinical studies. a consensus statement from the european atherosclerosis society consensus panel. Eur. heart journal 38, 2459–2472 (2017).DOI
  42. [43]Bräsen, J. H. et al. Cholemic nephropathy causes acute kidney injury and is accompanied by loss of aquaporin 2 in collecting ducts. Hepatology 69, 2107–2119 (2019).DOI
  43. [44]Van Slambrouck, C. M., Salem, F., Meehan, S. M. & Chang, A. Bile cast nephropathy is a common pathologic finding for kidney injury associated with severe liver dysfunction. Kidney international 84, 192–197 (2013).DOI
  44. [45]Warny, M., Helby, J., Nordestgaard, B. G., Birgens, H. & Bojesen, S. E. Lymphopenia and risk of infection and infection-related death in 98,344 individuals from a prospective danish population-based study. PLoS medicine 15, e1002685 (2018).DOI
  45. [46]Sagar, P. et al. Intestinal obstruction promotes gut translocation of bacteria. Dis. colon & rectum 38, 640–644 (1995).
  46. [47]Biswas, A. K. & Acharyya, S. Understanding cachexia in the context of metastatic progression. Nat. Rev. Cancer 20, 274–284 (2020).DOI
  47. [48]Williams, M. D. et al. Hospitalized cancer patients with severe sepsis: analysis of incidence, mortality, and associated costs of care. Critical care 8, R291 (2004).
  48. [49]Juhl, A. R. et al. Incidentally detected splenomegaly and risk of hematologic cancer and liver disease. JAMA oncology 12, 275–284 (2026).DOI
  49. [50]Van Deursen, V. et al. Abnormal liver function in relation to hemodynamic profile in heart failure patients. J. cardiac failure 16, 84–90 (2010).DOI
  50. [51]Goldwasser, P. & Feldman, J. Association of serum albumin and mortality risk. J. clinical epidemiology 50, 693–703 (1997).
  51. [52]Don, B. R. & Kaysen, G. Poor nutritional status and inflammation: serum albumin: relationship to inflammation and nutrition. Semin. Dial. 17, 432–437 (2004).
  52. [53]Launay-Vacher, V. et al. Prevalence of renal insufficiency in cancer patients and implications for anticancer drug management: the renal insufficiency and anticancer medications (irma) study. Cancer 110, 1376–1384 (2007).
  53. [54]Sayers, E. W. A general introduction to the e-utilities. In Entrez Programming Utilities Help (National Center for Biotechnology Information, Bethesda, MD, 2009). Updated November 17, 2022.
  54. [55]Generative Expert Labs. Paperclip. Scientific literature search software (2026). Version 0.7.130, accessed October 19, 2026.
  55. [56]Blanco-Grau, A. et al. Assessing liver fibrosis using the fib4 index in the community setting. Diagnostics 11, 2236 (2021).
  56. [57]Ndumele, C. E. et al. Cardiovascular-kidney-metabolic health: a presidential advisory from the american heart association. Circulation 148, 1606–1635 (2023).DOI
  57. [58]Ghallab, A. et al. Enteronephrohepatic circulation of bile acids and therapeutic potential of systemic bile acid transporter inhibitors. J. hepatology (2025).DOI
  58. [59]Barabási, A.-L., Gulbahce, N. & Loscalzo, J. Network medicine: a network-based approach to human disease. Nat. reviews genetics 12, 56–68 (2011).DOI
  59. [60]Barnett, K. et al. Epidemiology of multimorbidity and implications for health care, research, and medical education: a cross-sectional study. The Lancet 380, 37–43 (2012).
  60. [61]VanderWeele, T. J. Principles of confounder selection. Eur. journal epidemiology 34, 211–219 (2019).DOI
  61. [62]Muen­nighoff, N., Tazi, N., Magne, L. & Reimers, N. Mteb: Massive text embedding benchmark. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, 2014–2037 (2023).
  62. [63]Bussmann, B., Nabeshima, N., Karvonen, A. & Nanda, N. Learning multi-level features with matryoshka sparse autoencoders. arXiv preprint arXiv:2503.17547 (2025).arxiv.org/abs/2503.17547
  63. [64]Thapa, R. et al. A multimodal sleep foundation model for disease prediction. Nat. Medicine 32, 752–762 (2026).
  64. [65]Chernozhukov, V. et al. Double/debiased machine learning for treatment and structural parameters. The Econom. J. 21, C1–C68 (2018).
  65. [66]Liu, M., Zhang, Y. & Zhou, D. Double/debiased machine learning for logistic partially linear model. arXiv preprint arXiv:2009.14461 (2020).arxiv.org/abs/2009.14461
  66. [67]Tan, Z. On doubly robust estimation for logistic partially linear models. Stat. & Probab. Lett. 155, 108577 (2019).
  67. [68]Bello, K., Aragam, B. & Ravikumar, P. Dagma: Learning dags via m-matrices and a log-determinant acyclicity characterization. Adv. Neural Inf. Process. Syst. 35, 8226–8239 (2022).arxiv.org/abs/2209.08037
  68. [69]Zheng, X., Aragam, B., Ravikumar, P. K. & Xing, E. P. Dags with no tears: Continuous optimization for structure learning. Adv. Neural Inf. Process. Syst. 31 (2018).arxiv.org/abs/1803.01422
  69. [70]Bouzid, K. et al. Insights into a radiology-specialised multimodal large language model with sparse autoencoders. arXiv preprint arXiv:2507.12950 (2025).arxiv.org/abs/2507.12950

Paper details

vlnc:W_3378hsns
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 4

downloads · 1

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents