WorksComputer Science
Causally Constrained and Uncertainty-Aware Cardiopulmonary Risk Prediction
Abstract
Routine chest computed tomography (CT) captures pulmonary, cardiac, and vascular phenotypes that jointly stratify lung cancer and cardiovascular disease, the two leading causes of death in lung-screening populations. Yet clinical risk calculators evaluate each disease in isolation and omit CT findings, whereas unconstrained machine learning overfits small hospital cohorts, learns treatment- and illness-confounded risk directions, and conceals uncertainty when patient records are incomplete. Here we present the Causal Envelope Agent (CE-Agent), a two-stage framework that compiles peer-reviewed epidemiological evidence into explicit constraints on risk-factor weights before any outcome is seen, while estimating local baseline risk and newly added CT phenotypes from cohort data. At deployment, closed-form uncertainty propagation accounts for missing clinical or CT inputs within milliseconds per patient, running over two orders of magnitude faster than Monte Carlo sampling and returning calibrated risk intervals that direct chart review and selective imaging to borderline patients. Across 53,452 lung-screening participants from 33 US hospitals and 6,814 multi-ethnic adults, CE-Agent outperforms cardiovascular calculators in lung screening, exceeds the Pooled Cohort Equations on hard atherosclerotic events while correcting their 1.7-fold risk overestimation, and generalizes from a single hospital or CT scanner to external sites while lowering required sample sizes up to fivefold. Applied unchanged to lung cancer, CE-Agent matches established screening models with accurate calibration even when restricted to risk-factor directions alone, and pairing both predictions improves joint cardiopulmonary triage from a single screening visit.
Introduction
Low-dose chest computed tomography (CT) is the standard of care for lung cancer screening in adults with a heavy smoking history [1, 2, 3]. The same adults carry a high burden of cardiovascular disease (CVD): smoking and aging raise the risk of both [4, 5, 6], and in the National Lung Screening Trial (NLST) they were the two leading causes of death, with comparable shares of mortality [1]. Every screening scan already images the heart and great vessels alongside the lungs, capturing pulmonary nodules, emphysema, coronary artery calcification (CAC), thoracic aortic calcification (TAC), pulmonary arterial remodeling, and cardiac chamber geometry [7, 8, 9, 10, 11, 12, 13]. Society guidelines recommend reporting coronary calcification on every noncontrast chest CT [12], yet screening eligibility is set by lung cancer risk alone [3]. Established risk calculators each address one disease with fixed coefficients that omit chest CT phenotypes, and often miscalibrate when carried to new populations [14]. This holds for lung-screening models such as the Prostate, Lung, Colorectal, and Ovarian Cancer Screening Trial 2012 model (PLCOm2012) [6] and for cardiovascular equations, including Framingham general CVD [15], the American College of Cardiology/American Heart Association (ACC/AHA) Pooled Cohort Equations (PCE) [16, 17, 18], and AHA Predicting Risk of Cardiovascular Disease Events (PREVENT) [19].
Data-driven risk models have advanced individualized prediction from imaging [13, 20] and molecular profiles [21] and have informed treatment prioritization [22]. Extending them to joint cardiopulmonary assessment at individual screening sites raises three challenges. First, individual hospitals, demographic groups, and CT scanner platforms rarely accumulate enough outcome events to estimate multi-parameter risk models reliably [14, 23]. Second, observational records mix baseline exposures with post-scan events, derived indices such as body mass index (BMI) or pulse pressure that duplicate their components, and measurements distorted by treatment or illness. Comorbidity lowers total cholesterol while raising mortality [24], statin initiation in high-risk patients [17, 25] inverts lipid associations in treated cohorts [26], and chronic obstructive pulmonary disease (COPD) enlarges the pulmonary artery [11], confounding its cardiovascular signal. Unconstrained models can therefore learn risk directions opposite to clinical biology [27, 28]. Third, outpatient records often lack recent laboratory panels or a prior chest CT at the point of care, and substituting a single imputed value into a nonlinear risk equation hides the uncertainty of the individual risk estimate [29, 30] near preventive treatment thresholds [17, 18, 25, 31].
Here we introduce the Causal Envelope Agent (CE-Agent), a two-stage framework that grounds clinical and chest CT risk modeling in peer-reviewed epidemiological evidence (Figure 1). Its central object, the Causal Envelope, states in advance how each predictor may relate to an outcome: post-scan events and composite indices that duplicate their components are excluded, established risk factors keep their published direction and range, and local baseline risk and newly added CT phenotypes are estimated from cohort data. Stage I compiles this envelope from published hazard and odds ratios before any outcome label is seen, and Stage II fits each predictor’s weight inside it. At deployment, an analytic formula evaluated once per patient propagates missing-data uncertainty, directing chart review and selective imaging to patients whose unmeasured inputs could change a clinical decision [32, 33, 34, 35]. The name reflects its role as the deterministic, fixed-path core of a broader cardiopulmonary agentic framework: CE-Agent executes five modules in a fixed, auditable sequence (clinical-record reading, CT-biomarker extraction, prior compilation, constrained estimation, and uncertainty propagation).
Figure 1. Overview of the Causal Envelope Agent (CE-Agent). a, Stage I, before any outcome is seen: a directed acyclic graph (DAG) screen removes composite indices and post-scan information; published hazard and odds ratios (HR, OR) per standard deviation (SD) assign each predictor a range, point value, direction, or free estimation. b, Stage II fits weights within these constraints, zeroes those contradicting the published direction, and estimates new CT phenotypes locally. c, Closed-form uncertainty propagation returns a risk, its 90% interval, and a triage decision. BP, blood pressure; Ht, height; Wt, weight; BMI, body mass index; CAC, coronary artery calcification; COPD, chronic obstructive pulmonary disease; PA, pulmonary artery.
We evaluate it in two complementary US cohorts totaling 60,266 adults. NLST includes 53,452 adults with a heavy smoking history, randomized to low-dose CT (LDCT) or chest radiography at 33 screening hospitals [1]. The Multi-Ethnic Study of Atherosclerosis (MESA) includes 6,814 adults from four racial/ethnic groups imaged on four CT scanner platforms [36, 37]. We first establish accuracy, single-hospital learning, protection against confounding, and uncertainty handling on cardiovascular outcomes in both cohorts. We then apply the same framework, unchanged, to lung cancer in NLST and combine the two predictions into joint cardiopulmonary triage from a single screening encounter, which adds cardiovascular prevention to lung cancer screening.
Results
Prognostic accuracy and calibration across lung screening and primary prevention
Published evidence improved cardiovascular risk models in both lung screening (NLST) and multi-ethnic primary prevention (MESA). Before any outcome was seen, Stage I retained 13 predictors in NLST and 22 in MESA and assigned each a published range or value, a direction only, or free estimation (Figure 1a and Figure 2a). In single-hospital NLST models, fitting within the Causal Envelope (Figure 1b) raised the mean area under the receiver operating characteristic curve (AUC) for cardiovascular mortality at the other 32 hospitals from 0.656 without constraints to 0.725 (Figure 3a). In the MESA test set, it raised 10-year AUC from 0.789 to 0.807, ahead of ridge shrinkage toward zero [38] and of the same model with weights fixed at published values (Figure 3d,Figure 3e and Table A4).
Figure 2. The Causal Envelope: constraints, geometry, and simulated interventions. a, Standardized log-odds weights by constraint type, from unconstrained fits (Unconstr.) and CE-Agent (CE). b, Schematic: ridge () shrinkage pulls an established risk factor and the confounded pulmonary artery (PA) diameter ratio toward zero; the Causal Envelope holds the first in range and zeroes the second. c, Risk ratios (RR) for MESA cardiovascular mortality under a simulated 1 mmol/L reduction in low-density lipoprotein (LDL) cholesterol, against Cholesterol Treatment Trialists’ (CTT) randomized controlled trials (RCTs; 95% confidence interval, CI) [39]. SES, socioeconomic status; Cox PH, Cox proportional hazards; LR, logistic regression; XGB-Cox, Cox-boosted trees.
Figure 3. Agreement with published evidence and held-out performance. a,d, Discrimination against penalty on departures from published bounds (, unconstrained; , CE-Agent) or, for ridge, from zero, in single-hospital NLST (a) and MESA (d); CV, cross-validation; right axis, mean bound deviation. b,e, Test-set concordance index (-index; 95% confidence interval, CI) and 10-year for CE-Agent, variants with weights penalized toward published ranges (soft) or fixed at published values (prior-locked), and comparators; dashes, calculators target composite endpoints with nonfatal events (calibration in Table A2 and Table A6). c,f, Time-dependent AUC (-AUC) for cardiovascular mortality in NLST under competing mortality [40] (c; shading, gain over PCE; inset Int. -AUC spans 1–13 yr) and in MESA [41] (f; gain over unconstrained). , competing-risk expected-to-observed ratio; AUC, 10-year AUC; PCE, ACC/AHA Pooled Cohort Equations; PREVENT, AHA risk equations; Int., integrated.
| Model / Specification | AUC (95% CI) | / | Int. -AUC | Slope (E/O) |
| CE-Agent (hard, ){sup}†` | 0.8068 (0.758–0.855) | 0.8157 / 0.8046 | 0.8423 | 0.777 (1.07) |
| CE-Agent (soft, ) | 0.8051 (0.752–0.854) | 0.8145 / 0.8074 | 0.8411 | 0.833 (1.07) |
| CE-Agent (hard, binary CAC) | 0.8107 (0.763–0.858) | 0.8194 / 0.8035 | 0.8418{sup}§` | 0.826 (1.07) |
| CE-Agent (hard, ordinal CAC) | 0.8072 (0.759–0.855) | 0.8162 / 0.8032 | — | 0.788 (1.06) |
| CE-Agent (prior-locked) | 0.7940 (0.738–0.847) | 0.8032 / 0.8005 | 0.8416 | 0.466 (1.08) |
| Unconstrained 22-Predictor Model | 0.7886 (0.729–0.842) | 0.7979 / 0.7952 | 0.8246 | 0.768 (1.07) |
| Zero-Locked Ridge (8 free) | 0.7130 (0.642–0.779) | 0.7211 / 0.7282 | — | 0.726 (1.04){sup}∗` |
| AHA PREVENT (10-yr) | 0.8174 (0.769–0.865) | 0.8273 / 0.8044 | 0.8428 | 1.742 (—){sup}‡` |
| ACC/AHA ASCVD-PCE | 0.8131 (0.767–0.857) | 0.8209 / 0.7945 | 0.8306 | 1.096 (—){sup}‡` |
| Framingham General CVD | 0.8149 (0.764–0.864) | 0.8204 / 0.7720 | 0.8306 | 1.149 (—){sup}‡` |
Table A4. MESA test-cohort cardiovascular mortality: CE-Agent specifications versus clinical calculators (). Bold values denote overall column-best values across non-ablation models ( and closest to zero); survival-increment-weighted integrated -AUC spans 1–15 yr (horizon curves across 2–15 yr in Figure 3f). Rightmost column reports calibration slope alongside complete-case 10-yr E/O (competing-risk E/O in Table A2).
In the held-out NLST test cohort (), CE-Agent trained across all 33 hospitals achieved a concordance index (-index) of 0.731 and accurate 10-year calibration (expected-to-observed event ratio , accounting for competing non-cardiovascular deaths [40]). It outperformed the Framingham, PCE, and PREVENT calculators while matching unconstrained survival and machine learning (ML) models pooled across all 33 hospitals (Figure 3b,Figure 3c and Table A1 and Table A3). Because NLST did not measure blood pressure, lipids, or kidney function, those inputs were set to standard untreated US adult reference values for the three calculators (Methods).
| Model / Specification | AUC (95% CI) | -index (95% CI) | Brier | E/O |
| CE-Agent (compiled hard bounds, primary) | 0.723 (0.700–0.744) | 0.731 (0.713–0.749) | 0.0433 | 1.02 |
| CE-Agent (compiled soft bounds, CV ) | 0.723 (0.700–0.744) | 0.731 (0.713–0.749) | 0.0433 | 1.02 |
| CE-Agent (point-anchored bounds) | 0.721 (0.700–0.742) | 0.729 (0.711–0.747) | 0.0436 | 1.02 |
| CE-Agent (prior-locked) | 0.717 (0.696–0.738) | 0.725 (0.707–0.744) | 0.0435 | 1.02 |
| Logistic Regression | 0.725 (0.702–0.747) | 0.734 (0.713–0.749) | 0.0433 | 1.06 |
| XGBoost Classifier | 0.724 (0.702–0.745) | 0.735 (0.717–0.751) | 0.0434 | 1.04 |
| Random Forest / RSF | 0.725 (0.703–0.746) | 0.736 (0.716–0.752) | — | 1.05 |
| Cox Proportional Hazards | 0.719 (0.697–0.740) | 0.728 (0.706–0.746) | — | 1.04 |
| ACC/AHA ASCVD-PCE | 0.683 (0.659–0.705) | 0.693 (0.672–0.709) | — | — |
| AHA PREVENT (10-yr) | 0.680 (0.655–0.703) | 0.689 (0.668–0.706) | — | — |
| Framingham General CVD | 0.667 (0.644–0.690) | 0.674 (0.653–0.691) | — | — |
Table A1. NLST test-cohort cardiovascular mortality: CE-Agent versus ML baselines and clinical calculators (). Ten-year Aalen–Johansen competing-risk AUC (525 cardiovascular deaths and 1,684 competing deaths), full-follow-up -index (750 cardiovascular deaths over 13.4 yr), Brier score, and 10-yr E/O across CE-Agent specifications, pooled 13-predictor survival and ML baselines, and clinical calculators (extended metrics in Table A3).
| Model / Specification | (95% CI) | - / CS | Slope () | () |
| CE-Agent (hard, primary) | 0.7225 (0.700–0.744) | 0.7212 / 0.7386 | 0.951 (1.02) | 6.8 |
| CE-Agent (soft, CV ) | 0.7225 (0.700–0.744) | 0.7212 / 0.7386 | 0.951 (1.02) | 6.8 |
| CE-Agent (sign-only) | 0.7225 (0.700–0.744) | 0.7213 / 0.7387 | 0.950 (1.02) | 6.8 |
| CE-Agent (point-anchored) | 0.7208 (0.700–0.742) | 0.7215 / 0.7379 | 0.805 (1.02) | 6.4 |
| CE-Agent (prior-locked) | 0.7166 (0.696–0.738) | 0.7179 / 0.7350 | 0.773 (1.02) | 6.6 |
| Random Forest (pooled) | 0.7251 (0.703–0.746) | 0.7236 / 0.7410 | 1.143 (1.05) | 6.4 |
| Logistic Regression (pooled) | 0.7248 (0.702–0.747) | 0.7223 / 0.7397 | 0.934 (1.06) | 6.5 |
| XGBoost Classifier (pooled) | 0.7240 (0.702–0.745) | 0.7234 / 0.7404 | 0.865 (1.04) | 6.9 |
| Cox PH (CIF, pooled) | 0.7188 (0.697–0.740) | 0.7186 / 0.7355 | 1.958 (1.04) | 5.5 |
| ACC/AHA ASCVD-PCE | 0.6834 (0.659–0.705) | 0.6816 / 0.6973 | 1.060 (—) | 4.9 |
| AHA PREVENT (10-yr) | 0.6801 (0.655–0.703) | 0.6778 / 0.6943 | 1.546 (—) | 4.4 |
| Framingham General CVD | 0.6675 (0.644–0.690) | 0.6659 / 0.6790 | 0.992 (—) | 3.8 |
Table A3. Table A3 NLST test-cohort cardiovascular mortality: time-dependent discrimination, calibration, and net benefit (). Bold marks overall column-best values ( and closest to zero; 1–10 yr -AUC() curves in Figure 3c).
In the held-out MESA cohort (), where blood pressure, fasting lipids, and kidney function were directly measured, the 22-predictor CE-Agent outperformed every survival and ML baseline built on the same predictors and matched both the clinical calculators and an expanded 58-predictor (63-column) Cox proportional hazards model, while staying calibrated for cardiovascular death (; Figure 3e,Figure 3f and Table A2, Table A4, and Table A5). Evaluated directly on each calculator’s composite target, CE-Agent exceeded PCE in integrated time-dependent AUC (-AUC) [41] on hard atherosclerotic cardiovascular disease (ASCVD; 0.802 versus 0.777, ; Methods) while correcting PCE’s 1.7-fold risk overestimation ( versus 1.70; Table A2 and Table A6). On the broad cardiovascular composite matched to Framingham, CE-Agent likewise exceeded Framingham in integrated -AUC (0.813 versus 0.789, ) and maintained accurate calibration (expected-to-observed ratio 0.97 versus 1.44; Table A6).
| Model / Specification | AUC (95% CI) | -index (95% CI) | Brier | E/O |
| Cardiovascular mortality (; 49 10-yr and 147 total deaths) | ||||
| CE-Agent (hard, , primary) | 0.807 (0.758–0.855) | 0.805 (0.773–0.835) | 0.0239 | 1.08 |
| CE-Agent (soft, ) | 0.805 (0.752–0.854) | 0.807 (0.776–0.838) | 0.0240 | 1.08 |
| CE-Agent (hard, binary CAC) | 0.811 (0.763–0.858) | 0.804 (0.772–0.833) | 0.0237 | —† |
| CE-Agent (prior-locked) | 0.794 (0.738–0.847) | 0.801 (0.767–0.832) | 0.0255 | 1.09 |
| Unconstrained 22-Predictor Model | 0.789 (0.729–0.842) | 0.795 (0.761–0.829) | 0.0245 | 1.08 |
| Cox Proportional Hazards | 0.797 (0.737–0.852) | 0.801 (0.767–0.836) | 0.0237 | 1.07 |
| ACC/AHA ASCVD-PCE | 0.813 (0.767–0.857) | 0.795 (0.765–0.825) | —‡ | —‡ |
| AHA PREVENT (10-yr) | 0.817 (0.769–0.865) | 0.804 (0.774–0.836) | —‡ | —‡ |
| Framingham General CVD | 0.815 (0.764–0.864) | 0.772 (0.740–0.804) | —‡ | —‡ |
| Composite hard ASCVD (; 137 10-yr and 283 total events) | ||||
| CE-Agent (hard bounds, primary) | 0.774 (0.738–0.811) | 0.753 (0.727–0.780) | 0.0682 | 1.04 |
| CE-Agent (soft bounds) | 0.776 (0.742–0.815) | 0.756 (0.729–0.783) | 0.0663 | 1.02 |
| Unconstrained 22-Predictor Model | 0.773 (0.739–0.813) | 0.752 (0.725–0.779) | 0.0665 | 1.02 |
| Cox Proportional Hazards | 0.772 (0.735–0.810) | 0.752 (0.725–0.778) | 0.0665 | 0.96 |
| ACC/AHA ASCVD-PCE | 0.750 (0.711–0.785) | 0.730 (0.701–0.758) | 0.0724 | 1.70 |
Table A2. MESA test-cohort cardiovascular mortality and hard ASCVD: CE-Agent versus unconstrained models and clinical calculators. Ten-year discrimination () and Brier score on uncensored 10-yr status ( for mortality; 1,769 for hard ASCVD), full-follow-up -index, and 10-yr competing-risk E/O (168 and 95 competing 10-yr deaths; extended benchmarks in Figure 3f and Table A4–Table A6).
| ML Architecture | AUC (22/58) | (22/58) | Int. -AUC | Brier | Slope (22/58) |
| Cox Proportional Hazards | 0.7972 / 0.8108 | 0.8059 / 0.8216 | 0.8373 | 0.02367 | 1.401 / 0.937 |
| Random Survival Forest | 0.7832 / 0.7855 | 0.7970 / 0.8080 | 0.8236 | 0.02391 | 1.036 / 1.086 |
| Logistic Regression | 0.7876 / 0.7510 | 0.7962 / 0.7614 | 0.8209 | 0.02481 | 0.732 / 0.520 |
| Cox-Boosted Trees | 0.7793 / 0.8009 | 0.7907 / 0.8121 | 0.8192 | 0.02507 | 0.641 / 0.613 |
| DeepSurv Neural Net | 0.7760 / 0.7601 | 0.7857 / 0.7708 | 0.8067 | 0.02504 | 0.617 / 0.329 |
| XGBoost Classifier | 0.7658 / 0.7411 | 0.7753 / 0.7521 | 0.8017 | 0.02454 | 0.646 / 0.538 |
| Decision Tree | 0.7133 / 0.6803 | 0.7204 / 0.6881 | 0.7362 | 0.02539 | 0.594 / 0.253 |
Table A5. MESA cardiovascular mortality: seven model families on the 22-predictor CE-Agent set versus an expanded 58-predictor (63-column) Exam 1 set (). Columns compare 22-predictor / 58-predictor test performance (14 model-by-feature configurations) alongside 22-predictor integrated -AUC (1–15 yr) and Brier score.
| Endpoint & Specification | AUC (95% CI) | Int. -AUC | Brier | Slope (E/O) | |
| Hard ASCVD (137/283 events; 95 competing 10-yr non-CVD deaths) | |||||
| CE-Agent (hard bounds) | 0.7740 (0.738–0.811) | 0.7531 | 0.8020 | 0.06820 | 0.706 (0.93) |
| CE-Agent (soft bounds) | 0.7765 (0.742–0.815) | 0.7561 | 0.8019 | 0.06630 | 0.904 (0.92) |
| ACC/AHA ASCVD-PCE | 0.7503 (0.711–0.785) | 0.7302 | 0.7774 | 0.07241 | 0.799 (1.55) |
| Hard CVD (138/287 events; 94 competing 10-yr non-CVD deaths) | |||||
| CE-Agent (hard bounds) | 0.7740 (0.738–0.811) | 0.7520 | 0.8012 | 0.06878 | 0.701 (0.95) |
| ACC/AHA ASCVD-PCE | 0.7475 (0.709–0.782) | 0.7271 | 0.7725 | 0.07291 | 0.790 (1.54) |
| Broad CVD (181/371 events; 85 competing 10-yr non-CVD deaths) | |||||
| CE-Agent (hard bounds) | 0.7870 (0.755–0.817) | 0.7625 | 0.8132 | 0.08474 | 0.721 (0.97) |
| ACC/AHA ASCVD-PCE | 0.7512 (0.718–0.783) | 0.7363 | 0.7762 | 0.08746 | 0.806 (1.18) |
| Framingham General CVD | 0.7635 (0.731–0.794) | 0.7362 | 0.7885 | 0.08655 | 1.012 (1.44) |
Table A6. MESA adjudicated composite endpoints: CE-Agent versus PCE and Framingham (). Hard ASCVD is defined in Methods; hard CVD is the adjudicated hard cardiovascular endpoint released with MESA; broad CVD comprises myocardial infarction, stroke, congestive heart failure, or cardiovascular death. Group headers list 10-yr / full-follow-up event counts and competing 10-yr non-cardiovascular deaths (uncensored 10-yr status: for hard ASCVD and hard CVD; for broad CVD).
Generalization across hospitals, demographics, and scanners
Models developed at one hospital or on one CT scanner generalized to the others (Table A8–Table A9). Trained at any single NLST hospital, CE-Agent outperformed eight survival and ML baselines at the other hospitals in all 264 hospital–baseline pairs (Figure 4a,Figure 4c). Trained on one of six MESA racial/ethnic or CT scanner groups with at least 20 training events, it outperformed the baselines in 43 of 48 comparisons, including every transfer between scanners (Figure 4d,Figure 4f). A single-hospital or single-scanner CE-Agent matched the external discrimination of unconstrained models pooled across 32 hospitals or three scanners (Figure 4b,Figure 4e) and outperformed methods designed for transfer to new sites [42, 43, 44] (Table A18).
Figure 4. Fig. 4 Generalization across environments and scaling with pooled centers. a,d, External AUC of models trained at one NLST hospital (a) or one of six MESA racial/ethnic or scanner groups with ≥ 20 training events (d); wins, training environments favoring CE-Agent. b,e, External discrimination after pooling NLST hospitals (b) and MESA scanner platforms (e; band, interquartile range, IQR). Partial, some published ranges retained; bounds freed, range-constrained predictors unconstrained. c,f, External AUC gain of CE-Agent over each baseline by NLST hospital (c) and MESA group (f). RSF, random survival forest; GBSA, gradient-boosted survival analysis; XGBoost, gradient-boosted trees; DeepSurv, neural-network survival model; Med., median; Ref., reference; Hisp., Hispanic.
Table A8 NLST models trained at one hospital and tested at the other 32: discrimination, minimum sample size, and net benefit (). Each architecture has 1,056 single-hospital external evaluations; multi-hospital scaling is in Figure 4b and Table A10.
| Model Architecture | AUC (IQR) | CE-Agent Wins | Excess NB () | |
| CE-Agent (point-anchored bounds) | 0.727 (0.726–0.727) | 264/264 | 1,110 | 0.83 |
| Random Survival Forest | 0.686 (0.667–0.703) | 33/33*** | 13,275 | — |
| Cox Proportional Hazards | 0.683 (0.648–0.703) | 33/33*** | 13,692 | 0.22 |
| Gradient-Boosted Survival | 0.671 (0.655–0.688) | 33/33*** | 15,777 | 0.09 |
| XGBoost Classifier | 0.658 (0.500–0.679) | 33/33*** | 18,790 | 0.29 |
| Cox-Boosted Trees | 0.653 (0.500–0.683) | 33/33*** | 19,943 | — |
| Logistic Regression | 0.643 (0.602–0.689) | 33/33*** | 23,029 | 0.37 |
| Decision Tree | 0.625 (0.594–0.641) | 33/33*** | 30,332 | 0.23 |
| DeepSurv Neural Network | 0.606 (0.556–0.642) | 33/33*** | 42,300 | 0.18 |
Table A8.
Table A9 MESA models trained on one racial/ethnic or scanner group and tested on the others: discrimination, minimum sample size, and net benefit (). Training groups are the six racial/ethnic and CT scanner subgroups with training events (34–81 per fold); finite-sample scaling is in Figure 4e and Table A11.
| Model Architecture | AUC (IQR) | CE-Agent Wins | Excess NB† () | |
| CE-Agent (hard bounds) | 0.801 (0.787–0.807) | 43/48 | 2,237 | 2.36 |
| Random Survival Forest | 0.795 (0.784–0.804) | 4/6 | Undef. | 2.24 |
| Cox Proportional Hazards | 0.796 (0.789–0.799) | 4/6 | 8,380 | 2.25 |
| Gradient-Boosted Survival | 0.776 (0.756–0.784) | 6/6 | Undef. | 1.43 |
| XGBoost Classifier | 0.743 (0.734–0.761) | 6/6 | Undef. | 2.17 |
| Cox-Boosted Trees | 0.794 (0.784–0.797) | 5/6 | Undef. | 2.04 |
| Logistic Regression | 0.762 (0.719–0.779) | 6/6 | 8,380 | 2.30 |
| Decision Tree | 0.677 (0.646–0.690) | 6/6 | Undef. | 1.33 |
| DeepSurv Neural Network | 0.756 (0.716–0.783) | 6/6 | Undef. | 1.81 |
Table A9.
| Benchmark | Protocol | CE-Agent | Baseline | Effect Size |
| (N × λ) inverse-sample-size scaling, adaptive weight selection, and separation control | ||||
| 1. NLST 1/ Weight Decay | centers (up to 36.4k training participants) | Subsample grid : | Slope rises as | |
| 2. NLST Sample-Size Selector | () if | Oracle gap across all | Pure inner CV (): (AUC 0.720) | AUC vs. fixed hard at |
| 3. NLST Prior-Shift Resilience | Synthetic shift SD (28 cells) | Inner CV (row 2 rule at ): oracle gap | Fixed hard bounds: up to | Inner CV detects shift in 28/28 cells |
| 4. MESA Free-Column | Unit-normal prior () on unpriced cols | 0/50 separated; : 0.7407, : 0.7607 | Unreg. cols: 29/50 (0.6512), 13/50 (0.7171) | +0.0895 (), +0.093 Chinese fold, +0.101 scanner fold |
| 5. MESA Adaptive Schedule | () across 12 cells | Oracle gap (12/12 CIs cover 0) | Pure inner CV (): oracle gap | Hard ASCVD (): 0.294 0.096 |
| Causal invariance, DRO, Steyerberg updating, and Gaussian MAP controls | ||||
| 6. Causal Anchor Regression [42] | across 33 NLST & 16 MESA OOD folds | NLST med. 0.7268; MESA med. 0.7752 | Anchor best : NLST 0.6619; MESA 0.7466 | 33/33 NLST (+0.0649 in medians); 15/16 MESA (med. paired +0.0345) |
| 7. Label-Free & CVaR DRO [43] | Oracle across 7 radii & 7 CVaR tail levels | NLST med. 0.7268; MESA med. 0.7752 | Oracle DRO: NLST 0.6460; MESA 0.7446 / 0.7475 | 33/33 NLST; 15/16 MESA (med. paired +0.0214 / +0.0262) |
| 8. Steyerberg Offset Updating [44] | Prior-offset LR & base-margin XGBoost on CE-Agent’s predicted log-odds | Prevalence-only envelope: 0.7268 / 0.7752 | Offset LR / XGB: NLST 0.6465 / 0.6929; MESA 0.7406 | Outperforms LR / XGB in 33/33 & 28/33 NLST, 16/16 MESA |
| 9. Gaussian Point-Prior MAP | Width-matched , | KKT exact : 95.1%–98.6% of fits | Gaussian MAP exact : 0.0% (68–115/200 TAC flip) | Retains full AUC (), preserves sparsity |
Table A18. Scaling the penalty with sample size and comparison with robust-learning, model-updating, and Gaussian-prior methods in NLST and MESA.
The constraints lowered the minimum sample size required for reliable model development [23] by 5.4-fold in NLST and 2.5- to 3.7-fold in MESA (Figure A1b,Figure A1e). On their own data, 12 of 33 NLST hospitals met this minimum, against none without constraints. External calibration slopes (1.0 indicating ideal calibration) stayed within the target range (0.80 to 1.25) for 28 of 33 single-hospital NLST models, against 8 without constraints (Table A10).
Figure A1. Clinical utility, minimum sample size, and calibration stability in NLST (a–c) and MESA (d–f). a,d, Decision curves for CE-Agent, unconstrained models, Cox PH, XGBoost, clinical calculators (d), and treat-all or treat-none strategies in single-hospital ( hospital) NLST external evaluation (a) and held-out MESA hard ASCVD (d; NB, net benefit; , decision threshold; shading, of 5% to 10%). b,e, Minimum sample size under the criteria of Riley et al. [23] as constraints are added, in NLST (b; , free slope parameters; soft, ; Site LR, site-level logistic regression; Max site, largest single hospital) and for MESA hard ASCVD and cardiovascular mortality (e; , effective number of free parameters; Unconstr., unconstrained model; Train pool, training participants with known 10-year status; strata, the four racial/ethnic groups and four scanner cohorts). c,f, Calibration slopes as training data grow from 1 to 32 pooled NLST hospitals (c) and from 250 to 4,821 MESA participants (f); band, 0.80 to 1.25. Unconstrained model. CE-Agent’s predictors fitted without published bounds; Unconstrained LR, logistic regression baseline.
| NLST Regime | CE-Agent Unboxed* /Unc.{sup}†` | Cox PH | RSF | GBSA | XGB | DSurv | |
| External OOD discrimination (mean AUC across 33 − centers) | |||||||
| Single Hospital | 0.7246 | 0.6977 | 0.6676 | 0.6761 | 0.6620 | 0.6178 | 0.6007 |
| Pooled Hospitals | 0.7246 | 0.7134 | — | — | — | — | — |
| Pooled Hospitals | 0.7266 | 0.7228 | — | — | — | — | — |
| Pooled Hospitals | 0.7258 | 0.7244 | — | — | — | — | — |
| Pooled Hospitals | 0.7251 | 0.7245 | — | — | — | — | — |
| LOCO (32 → 1) | 0.7256 | 0.7251 | 0.7256 | 0.7260 | — | 0.7280 | 0.7271 |
| Calibration slope error (), sample-size criteria, and DCA | |||||||
| |Med. slope − 1| | 0.178 | 0.512{sup}†` | — | — | — | 0.400 | — |
| |Med. slope − 1| | 0.173 | 0.183{sup}†` | — | — | — | 0.468 | — |
| |Med. slope − 1| | 0.160 | 0.072{sup}†` | — | — | — | 0.388 | — |
| |Med. slope − 1| | 0.166 | 0.035{sup}†` | — | — | — | 0.260 | — |
| |Med. slope − 1| | 0.166 | 0.021{sup}†` | — | — | — | 0.163 | — |
| |Med. slope − 1| | 0.153 | 0.017{sup}†` | — | — | — | 0.068 | — |
| Riley Free Params / Cols. | 4.0 | 12.0{sup}†` | 23–34 | 23–34 | 23–34 | 23–34 | 23–34 |
| Riley Required | 1,110 | 6,043{sup}†` | 13,692 | 13,275 | 15,777 | 18,790 | 42,300 |
| DCA Int. Excess NB (×) | 0.83 | 0.50{sup}†` | 0.22 | — | 0.09 | 0.29 | 0.18 |
| DCA Mean Excess NB (×) | 4.36 | 2.42{sup}†` | 1.03 | — | 0.97 | 4.16 | 2.17 |
| DCA Positive Cells (/330) | 289/330 | 196/330{sup}†` | 213/330 | — | 125/330 | 132/330 | 131/330 |
Table A10. NLST models pooled across 1 to 32 training hospitals: external discrimination, calibration, minimum sample size, and net benefit ().
CE-Agent delivered consistent clinical net benefit across hospitals and cohorts. In decision curve analysis [45], which weighs true against false positives across clinical risk thresholds, each of the 33 single-hospital NLST models gave higher mean net benefit at the other 32 hospitals than treat-all and treat-none strategies across thresholds of 5% to 10% (Figure A1a). For MESA hard ASCVD, CE-Agent led the eight baselines in net benefit integrated across decision thresholds (Figure A1d and Table A9). NLST-trained weights applied directly in MESA, adjusted only for baseline prevalence, achieved the highest concordance of any model carried across cohorts (Table A7 and Figure A2a).
| Transported Architecture | AUC (95% CI) | Brier | Slope (E/O) | |
| CE-Agent (compiled hard bounds) | 0.8029 (0.750–0.851) | 0.8066 | 0.02372 | 1.126 (1.06) |
| CE-Agent (point-anchored soft bounds) | 0.7961 (0.741–0.850) | 0.8052 | 0.02376 | 0.954 (1.07) |
| CE-Agent (point-anchored hard bounds) | 0.7940 (0.739–0.847) | 0.8043 | 0.02379 | 0.913 (1.07) |
| Unconstrained Logistic Reg. | 0.8029 (0.750–0.852) | 0.8023 | 0.02371 | 1.011 (1.06) |
| DeepSurv Neural Network | 0.8022 (0.748–0.852) | 0.8030 | 0.02371 | 1.026 (1.06) |
| Cox Proportional Hazards | 0.7921 (0.735–0.845) | 0.7977 | 0.02403 | 2.107 (1.05) |
| Cox-Boosted Trees (XGBoost) | 0.7761 (0.718–0.832) | 0.7866 | 0.02408 | 0.871 (1.05) |
| Random Survival Forest (RSF) | 0.7713 (0.710–0.827) | 0.7817 | — | — |
| XGBoost Classifier | 0.7699 (0.715–0.823) | 0.7807 | 0.02425 | 0.927 (1.05) |
| Classification Decision Tree | 0.7674 (0.720–0.814) | 0.7554 | 0.02420 | 1.286 (1.05) |
Table A7. NLST-trained models applied to MESA cardiovascular mortality on 12 shared predictors (). Models are trained exclusively on NLST () and evaluated on the MESA test cohort without refitting weights, recalibrating only the intercept to MESA prevalence.
Figure A2. Transport from NLST to MESA and projection of intermediate biomarkers in MESA. a, NLST-trained weights applied to the held-out MESA test cohort on 12 shared predictors with prevalence recalibration only: AUC and calibration slope (band, 0.80 to 1.25) for CE-Agent with hard, soft, or point-anchored (anchored) constraints and for seven survival and ML baselines. b, Test coefficient of determination () for predicting MESA’s measured intermediate biomarkers from their parent variables with training participants (bars) and with the full training cohort (; Full), for direction-constrained fitting (CE-Agent), ordinary least squares (OLS), multilayer perceptrons (MLP), and random forests. TAC, thoracic aortic calcification; eGFR, estimated glomerular filtration rate; CAC, coronary artery calcification; hs-CRP, high-sensitivity C-reactive protein. c, Partial slopes aligned to the expected sign (negative, reversed; 95% confidence intervals) of parent-to-biomarker links whose observed direction was reversed, adjusted for DAG parents only and after adding demographics and then medications (PM, fine particulate matter; SBP, systolic blood pressure). d, Expected-to-observed ratio (E/O; band, 0.90 to 1.10) and 10-year concordance () with all five chest CT phenotypes masked. CE-Agent (CE) with CT projected from clinical predictors (proj.), alone or down-weighted by projection variance (var. gate) or by the reliability score (rel. gate), is compared with observed CT, a model retrained without CT, population-mean fill, histogram-based gradient-boosted decision trees that route missing values natively (HistGBDT NaN), and ML imputers (RF, random forest).
Integrating chest CT phenotypes without confounded weights
CE-Agent incorporates new three-dimensional (3D) chest CT phenotypes while keeping confounded associations out of the model. Adding automatically measured left ventricular volume, pulmonary artery (PA) diameter ratio, and TAC raised cross-validated AUC from 0.759 to 0.774 among NLST participants with soft-tissue reconstructions (; Figure 5b), and five electrocardiogram-gated CT phenotypes gave parallel gains in MESA (Figure 5e). For the PA diameter ratio, which enlarges in COPD, the envelope specifies only the expected direction; Stage II zeroed this confounded weight in 89% of NLST cross-validation fits (Figure 2a,Figure 2b and Figure 5c). In MESA, where statin therapy and chronic illness invert crude lipid and blood-pressure slopes, published interval bounds kept six established risk factors—including total cholesterol—positive, whereas a direction-only envelope collapsed them to zero in 63% to 85% of fits (Figure 5f and Table A14–Table A15).
Figure 5. Composite-index screening, added CT phenotypes, and zeroed confounded weights. a,d, Discrimination of CE-Agent, unconstrained baselines, and models retaining composite indices (+BMI; +Shared-parent, sharing a parent measurement; +All 4 or +All 10 comp., all composites) in single-hospital ( hospital) NLST fits (a) and 250-participant MESA subsamples (d); collinearity, condition number (a, ratio to CE-Agent). b,e, Discrimination with added CT phenotypes: NLST low-dose CT (LDCT), soft-tissue () or any () reconstruction kernel (b); five MESA measures, for cardiovascular disease (CVD) death and hard atherosclerotic CVD (ASCVD; e). c,f, Standardized log-odds weights from unconstrained fits, fits with CT directions free (sign-free CT), a direction-only envelope (sign-only), and CE-Agent; percentages (f), sign-only fits with the weight zeroed. LV, left ventricular; TAC, thoracic aortic calcification; 3D, three-dimensional; Ao. diam., ascending aortic diameter.
| Experiment | Protocol | CE-Agent | ML Baseline | Effect Size |
| 1. Interval-Bound Sample Scaling | hospitals (4 active bounds) | : 0.7246 → : 0.7256 | Four bounds freed: 0.6977 → 0.7251; Cox: 0.6676 → 0.7256 | +0.0268 (, ) → +0.0005 () |
| 2. Factorial Bound Attribution () | All 16 bound coalitions across 33 centers () | Joint 4-bound | Sum of singletons ; Four bounds freed 0.6977 | 83.0% main-effect share; interaction |
| 3. CAC Provenance Audit | Replace the NLST CAC estimate [8] with the ITALUNG trial estimate [65] () | NLST estimate 0.7246; ITALUNG estimate 0.7248 | Free CAC 0.7134; mean RSF 0.6761; Cox 0.6676; LR 0.6376 | ; free-CAC model above 8/8 baselines |
| 4. Sign-Inversion Perturbation | All sign coalitions (0 → 9 inverted) | 0 flipped: med. AUC = 0.7246 | 3 flipped: 0.596; 6 flipped: 0.431; 9 flipped: 0.290 | (0 → 9 inverted) |
| 5. Positive Scale Resilience | Global prior multiplier (25× span) | Held-out [0.7297, 0.7306] | Sign inversion (480× larger) | across 25× span |
| 6. Algebraic Equivalence | Include derived vascular-aging composite () | Screened: | Included: , | (Proposition 1) |
| 7. Composite Relaxation | Screened vs. progressive relaxation at (33 centers) | Screened, published weights: 0.7251 (-ratio 1.0×) | Relax BMI: 0.7190; Shared-parent: 0.6985; All 4 free: 0.6947 (-ratio 321×) | (30/33 centers) |
| 8. 3D LDCT Expansion | Training LDCT subcohorts (; 6,467; 1,243), 5-fold CV and cross-site: +3 LDCT | AUC: 0.748–0.763 → 0.763–0.775 (soft-tissue 5-fold CV: 0.759 → 0.774) | Clinical-only CE-Agent (13 predictors): 0.748–0.763; PCE held-out -index: 0.6926 | to +0.017 (6/6 comparisons) |
| 9. Zeroing of Reversed Weights | COPD-confounded PA diameter ratio (1,040 CV fits) | in 928/1,040 (89.2%) | Direction constraint lifted: | 89.2% (928/1,040) zeroed* |
Table A14. Mechanistic ablations of the Causal Envelope in NLST ( training; rows 5–6, held-out test).
| Experiment | Protocol | CE-Agent | ML Baseline | Effect Size |
| 1. Interval-Bound Sample Scaling | (5 active bounds) | : 0.7171 → Full: 0.8068 | Five bounds freed: 0.6025 → 0.7983; Unc.: 0.5855 → 0.7886 | +0.1146 → +0.0569 → +0.0195 → +0.0129 → +0.0086 |
| 2. Factorial Bound Attribution () | All 32 coalitions at (50 draws) & full split | : ; Full +0.0086 | : (95.3%); Full (101%) | 95.3%–101% additive; interaction |
| 3. CAC Provenance Audit | Binary free/locked vs. Detrano interval [37] vs. Drop CAC | CVD bin. free 0.8107 (locked 0.7983); 0.8068 | Unc. 0.7886; RSF 0.7832; Drop CAC: ASCVD (95% CI −0.017 to −0.005) | +0.0043 on ASCVD at (40/50 draws)* |
| 4. Sign-Inversion Perturbation | 0 → 14 inverted signs (CVD death & Hard ASCVD) | 0 flipped: 0.8068 (CVD), 0.7740 (ASCVD) | 14 flipped: 0.4774 (CVD), 0.5602 (ASCVD); Unc. 6/14 flipped | (CVD), −0.214 (ASCVD) |
| 5. Positive Scale Resilience | Global prior multiplier (0.2 ×–5×) | 0.787–0.810 across 0.5 ×–2× | Zero-locked ridge (8 free): (−0.094) | (16× smaller vs. sign flip) |
| 6. Algebraic Equivalence | Include derived composites ( test) | Screened: | Included: , | (Proposition 1) |
| 7. Composite Relaxation | Screened vs. progressive relaxation at & (200 event-stratified draws) | Screened: 0.7150 (), 0.7662 () | All 10 free: 0.6299 (, ), 0.7277 () | () |
| 8. 5-Phenotype CT Progression | No CT → +CAC → +TAC → +Chamber → +Ao. diam. → +Liver fat | ASCVD 0.7576 → 0.7744; 0.7983 → 0.8069 | Unc. LR: 0.7468 → 0.7725 (ASCVD), 0.7920 → 0.7876 (CVD) | ASCVD (95% CI +0.006 to +0.027) |
| 9. Zeroing of Reversed Weights | 213 fits (200 bootstrap, 5 cross-validation, and 8 single-group) | ; | Unconstrained flips 6/14 (42.9%) signs () | 63.4%–84.5% zeroed across 6 predictors |
Table A15. Ablations of the Causal Envelope in MESA ().
Composite-index screening and published bounds each improved small-sample accuracy (Figure 5a,Figure 5d and Figure A3a,Figure A3b,Figure A3e,Figure A3f). Excluding redundant composite indices while retaining their underlying primary clinical measurements raised external AUC by 0.030 in single-hospital NLST fits and by 0.085 in 250-participant MESA subsamples.
Figure A3. Contribution of published ranges, the CAC prior source, and predictor directions in NLST (a–d) and MESA (e–h). a,e, External discrimination across sample size for CE-Agent with all, some (partial), or none (bounds freed) of the published ranges enforced, across 1 to 32 pooled NLST hospitals (a) and MESA subsamples of 250 to 4,821 participants (e); shading, all versus none. Comparators are RSF, Cox PH, and an unconstrained fit (the CE-Agent model without priors in a; unconstrained logistic regression in e). b,f, Gain in discrimination from all published ranges together, from the sum of single-predictor ranges, and from each range alone, in single-hospital NLST fits (b) and 500-participant MESA subsamples (f; PM, fine particulate matter; Resp., respiratory). c,g, CE-Agent under alternative sources for the CAC weight (the NLST-derived [8] and ITALUNG [65] estimates, a free CAC weight, and all four interval bounds freed in c) and alternative CAC codings (including the Detrano et al. [37] log-CAC bound in g), compared with RSF, Cox PH, and unconstrained logistic regression. d,h, Discrimination when the directions of subsets of predictors are inverted (median; band, range in d and interquartile range in h), against intact CE-Agent and Cox PH; rescaling all priors changed discrimination little ( index) across a 25× span in NLST, d; AUC 0.787–0.810 across 0.5× to 2× in MESA, h). CE, CE-Agent.
Under simulated reductions in smoking, body weight, cholesterol, and systolic blood pressure [5, 39, 46, 47], CE-Agent’s predicted risk fell or held steady for every participant in both cohorts, whereas unconstrained ML baselines predicted paradoxical risk increases in up to half of patients in NLST and 46%–100% on lipid reduction in MESA (Figure 2c and Figure A4c,Figure A4g). In MESA, lowering low-density lipoprotein (LDL) cholesterol by 1 mmol/L gave a risk ratio of 0.865, matching the pooled randomized-trial estimate for vascular mortality (0.86 [39]; Figure 2c), and lowering systolic blood pressure by 10 mmHg gave 0.876, consistent with trial estimates for all-cause mortality (0.87) and major cardiovascular events (0.80 [47]; Table A19).
Figure A4. Penalty scaling, robust-learning comparators, simulated risk-factor reductions, and two-stage CT triage in NLST (a–d) and MESA (e–h). a,e, External discrimination of CE-Agent (CE) with hard bounds, the penalty (adaptive -to-CV rule in a; pure in e), or cross-validated (CV) soft bounds against the unconstrained model (left axis), with the AUC-optimal penalty weight (amber, right axis), across pooled NLST hospitals (a) and MESA training subsamples (e; shaded band, small-sample gain from the penalty). b,f, Single-environment external discrimination of CE against Anchor Regression [42], and conditional value-at-risk (CVaR) distributionally robust optimization (DRO) [43], and prior-offset logistic regression (LR) model updating [44] and its gradient-boosted decision tree (GBDT) extension across 33 NLST hospitals (b) and 16 MESA demographic and scanner transfers (f). c,g, Share of patients whose predicted risk rose under a simulated risk-factor reduction in seven models: BMI and smoking (pack-years, pack-yr) reductions in single-hospital NLST fits (c) and LDL cholesterol and systolic blood pressure (SBP) reductions in MESA (g). d,h, Net benefit per 1,000 patients (/1k) against the share with CAC read, in NLST screening (d) and in MESA adults not on lipid-lowering therapy (h; error bar, guideline 95% CI; markers at 28.5% offset for visibility); dotted line, guideline net benefit. CE reads no or all CAC, reads CAC only when the 90% CT-uncertainty interval crosses the threshold (selective), or reads as many scans as the guideline (budget-matched); PCE is used alone (recalibrated in NLST) or refers intermediate-risk patients for CAC (guideline).
| Benchmark | Protocol | CE-Agent | Baseline | Clinical Effect |
| Counterfactual interventional direction consistency across single-environment and pooled fits | ||||
| 1. NLST Smoking & BMI () | Cessation, pack-yr, BMI [5, 46] | 0.0% rises (0/33 , 0/50 subsample rev.) | rev.: 9% quit, 21% pack-yr, 41% BMI (51.6% rises) | LOCO BMI rev. in 4/8 ML (55%–62% rises)* |
| 2. MESA LDL ( mmol/L) | Counterfactual LDL ( with recorded lipids) vs. CTT RCT [39] | 0.0% discordant rises (RR = 0.865, 0/48 fits) | 20/24 ML configs reversed (46%–100% rises) | Eliminates statin-confounded lipid inversion |
| 3. MESA SBP ( mmHg) | Counterfactual SBP vs. Ettehad RCT [47] | 0.0% discordant rises (RR = 0.876, 0/48 fits) | RSF 31.3%, RF 40.2%, XGB-Cox 50.8%, DeepSurv 53.2% rises | Prevents tree/neural local non-monotonicity |
| Two-stage clinical CT reading and statin triage efficiency | ||||
| 4. NLST Opportunistic LDCT Triage | Primary-prev. LDCT train (, 31 non-zero-event hospital fits, ) [48, 49] | LOCO +2.52/1k (18.3% CT); +1.48/1k | Intermediate PCE CAC: +1.38/1k (40.5% CT); ML < 0 | −55% reads (+1.14/1k, 95% CI 0.23–2.13, ); 27/31 vs. ML† |
| 5. MESA 2018 ACC/AHA CAC Triage | Lipid-therapy-free test ages 45–75 (, ) [17] | Equal-budget (28.5% CT): +16.0/1k; Two-stage selective CT (7.1% CT, 24.9% statin): +13.0/1k | Guideline PCE CAC: +16.1/1k (28.5% CT, 42.2% statin)† | Parity at 28.5% CT; −75% CT & −41% statin within 95% CI |
Table A19. Simulated risk-factor reductions and two-stage triage in NLST and MESA.
Real-time uncertainty propagation, outpatient adaptation, and targeted CT triage
Records are often incomplete at the point of care. With up to 60% of predictors missing, CE-Agent’s closed-form uncertainty propagation (Figure 1c) reproduced Monte Carlo predictive intervals with mean relative error up to 2.3%, in 1 to 3 ms per patient on a single core (Figure 6a,Figure 6e and Tables A16–A17). On synthetic risk-factor graphs, runtime grew linearly with model size, reaching 18.4 ms at 1,000 variables. Across sizes it was 83- to 377-fold faster than 4,000-draw Monte Carlo sampling (353-fold at 1,000 variables), with near-identical risk estimates (Figure 7a,Figure 7b).
Figure 6. Uncertainty propagation, reliability scoring, and uncertainty-guided triage in NLST (a–d) and MESA (e–h). a,e, Relative error of 90% interval widths () from closed-form propagation, 400- and 4,000-draw Monte Carlo (MC), and single-value (plug-in) imputation, against 100,000-draw MC; 0%–60% missing completely at random (MCAR) or at random given BMI (MAR); monotonicity violations, risk rising when a harmful predictor is zeroed. b,f, Calibration slopes for reliability scoring, its untuned version, population-mean fill, and plug-in imputation (band, 0.80–1.25). c,g, Extra discrimination loss at 40% missingness versus CE-Agent; labels, overconfident false negatives (FN; events predicted below half the training event rate) per 1,000 (/1k): most uncertain fifth, 5%–40% missingness (c); all participants, MCAR (g). d,h, Discrimination gain among retained records with up to 30% reviewed, ranked by missing-data uncertainty (), missing-predictor count, total interval width, or at random.
Figure 7. Per-patient runtime from 22 to 1,000 variables. a, Single-core runtime (median, IQR) for 4,000-draw MC and CE-Agent closed-form propagation on synthetic risk-factor graphs; inset, CE-Agent curve. b, Logarithmic axis; closed-form propagation was 83- to 377-fold faster (353-fold at 1,000 variables), with Pearson between risk estimates. , links.
To prevent missing inputs from distorting predictions, CE-Agent weights each predictor by a reliability score that down-weights missing or high-variance imputations (Methods). This score kept calibration slopes near 1.0, preserved discrimination at 40% missingness, and reduced overconfident false negatives (events assigned low risk) at least twofold versus Cox, boosted-tree, logistic, and decision-tree baselines (Figure 6b,Figure 6c,Figure 6f,Figure 6g). When all five chest CT phenotypes were unmeasured at simulated MESA outpatient visits, projecting them from routine clinical predictors gave an expected-to-observed event ratio of 0.96, against 0.81 under population-mean filling (Figure A2d and Tables A12–A13).
The same uncertainty estimate directs chart review and selective imaging. Reviewing the 30% of records with the largest missing-data uncertainty raised AUC among the remaining records by 0.013 in NLST and 0.061 in MESA (Figure 6d,Figure 6h). In two-stage selective CT triage using published weights (Methods), CE-Agent requested a CAC read only when its 90% pre-CT risk interval—computed with CT predictors treated as unmeasured—crossed the 5% 10-year fatal-CVD threshold [48, 49]. Among CVD-free LDCT-arm participants in the NLST training set (), this rule required
55% fewer radiologist reads than CAC referral of everyone at intermediate risk by recalibrated PCE while raising net benefit by 1.14 per 1,000 (95% CI, 0.23 to 2.13; Figure A4d). For guideline statin decisions [17] among MESA adults not on lipid-lowering therapy (), CE-Agent matched the net benefit of guideline CAC referral at the same scan budget (+16.0 versus +16.1 per 1,000) and reached +15.4 without imaging, exceeding PCE alone (+14.0; Figure A4h and Table A19).
Lung cancer prediction and joint cardiopulmonary triage
Applied unchanged to NLST lung cancer ( held-out participants; Figure 8 and Table A20–Table A22), CE-Agent achieved accurate 6-year calibration (E/O = 0.99, versus 0.86 for published PLCOm2012) and matched recalibrated PLCOm2012 in competing-risk AUC (AUC = 0.708 versus 0.704; difference +0.004, 95% CI −0.008 to +0.017, meeting the prespecified −0.02 criterion; Figure 8a). It matched or outperformed survival and ML baselines built on the same 11 predictors, achieved comparable accuracy when restricted to risk-factor directions alone (AUC = 0.710), and maintained accurate calibration (E/O = 0.98) and discrimination when refitted to 10-year lung cancer mortality (Table A20).
Figure 8. Lung cancer prediction and joint cardiopulmonary triage in NLST. a, Test-set 6-year competing-risk AUC (; 95% CI) and E/O for CE-Agent, baselines on the same 11 predictors, and PLCOm2012 [6] recalibrated to NLST (recal.; dashed line) or published (pub.). b, Six-year AUC (AUC6) of single-hospital models evaluated across the remaining 32 hospitals. c, AUCcr in test-set LDCT participants () with radiologist-reported nodules and emphysema added. d, Training-cohort calibration slopes with 0%–60% of predictors missing (band, 0.80–1.25; published weights). e, Risk groups at 1.5% lung cancer and 5% fatal CVD cutoffs [48, 49, 50]. f, Summed net benefit per 1,000 (95% CI). USPSTF, US Preventive Services Task Force [3]; two-stage, CT read only when the 90% pre-CT risk interval crosses a cutoff; clinic-only, lung cancer CE-Agent without CT, paired with PCE; PLCOm2012 and PCE recalibrated unless marked pub.
| Model / Specification | LOCO LDCT | Test AUC | Slope (E/O) | 10-yr Death§ |
| CE-Agent (hard, primary) | 0.727 (0.708–0.745)* | 0.708 (0.689–0.728) | 0.99 (0.99) | 0.714 (0.693–0.735) |
| CE-Agent (soft, ) | 0.728 (0.709–0.746)* | 0.709 (0.691–0.729) | 0.99 (0.99) | 0.716 (0.696–0.738) |
| CE-Agent (sign-only)† | 0.728 (0.708–0.746)* | 0.710 (0.691–0.729) | 0.98 (0.99) | 0.715 (0.695–0.737) |
| Cox PH (CIF, 11 pred.) | 0.727 (0.707–0.746) | 0.711 (0.691–0.731) | 2.76 (1.00) | 0.713 (0.693–0.735) |
| Logistic Reg. (11 pred.) | 0.727 (0.708–0.745) | 0.710 (0.690–0.729) | 0.92 (1.10) | 0.712 (0.692–0.734) |
| XGBoost (11 pred.) | 0.726 (0.707–0.744) | 0.709 (0.688–0.729) | 0.87 (1.08) | 0.706 (0.685–0.727) |
| DeepSurv (11 pred.) | 0.725 (0.705–0.744) | 0.708 (0.686–0.728) | 0.92 (1.01) | — |
| RSF (11 pred.) | 0.710 (0.690–0.729) | 0.695 (0.674–0.715)‡ | 0.90 (1.02) | — |
| GBSA (11 pred.) | — | 0.666 (0.644–0.689)‡ | 4.19 (1.00) | — |
| PLCOm2012 (recal.) [6] | 0.704 (0.685–0.725) | 0.704 (0.684–0.724) | 0.92 (0.99) | 0.723 (0.703–0.742) |
| PLCOm2012 (published) [6] | 0.704 (0.685–0.725) | 0.704 (0.684–0.724) | 0.92 (0.86) | 0.723 (0.703–0.742) |
Table A20. NLST lung cancer: CE-Agent, PLCOm2012, and ML baselines in the test cohort and in leave-one-center-out validation in the LDCT arm.
| Strategy / Quadrant | CT / Flag | All (1.5%/5%) | LDCT (1.5%/5%) | All (2%/5%) | LDCT (2%/5%) |
| Joint cardiopulmonary triage net benefit per 1,000 participants (95% CI) | |||||
| Dual CE-Agent (full CT) | 100% / 83.1% | +30.2 (26.8–33.9) | +33.4 (28.6–38.9) | +26.6 (23.2–30.0) | +29.2 (24.4–34.5) |
| Dual CE-Agent (sign-only) | 100% / 82.0% | +30.0 (26.5–33.6) | +32.8 (27.9–38.3) | +26.4 (23.0–29.9) | +29.0 (24.1–34.2) |
| Two-stage selective CT † | 61.1% / 84.9% | +29.6 (26.2–33.3) | +32.6 (27.7–37.9) | +26.1 (22.8–29.6) | +28.5 (24.0–33.8) |
| PLCOm2012 + PCE (recal.) | 0% / 80.4% | +28.6 (25.3–32.1)∗ | +31.8 (27.3–37.1) | +25.2 (22.0–28.7) | +27.1 (22.5–32.7) |
| PLCOm2012 (pub.) + PCE (recal.) | 0% / 73.6% | +28.7 (25.4–32.3)∗ | +31.6 (26.9–37.0) | +24.7 (21.6–28.1)∗ | +26.4 (21.9–31.4) |
| Clinic-only (no CT) | 0% / 87.4% | +28.4 (25.1–32.0)∗ | +31.4 (26.7–36.8) | +24.4 (21.1–28.0)∗ | +26.6 (22.0–31.8) |
| USPSTF + PCE (recal.) | 0% / 99.4% | +27.5 (24.2–31.2)∗ | +30.8 (26.2–36.2)∗ | +22.5 (19.2–26.3)∗ | +25.9 (21.2–31.3)∗ |
| USPSTF + PCE (pub., 7.5%) | 0% / 99.4% | +18.4 (14.6–22.4)∗ | +20.8 (15.7–26.8)∗ | +13.5 (9.6–17.5)∗ | +15.9 (10.7–21.9)∗ |
| Held-out dual CE-Agent risk quadrants at 1.5% lung cancer & 5.0% fatal-CVD cutoffs () | |||||
| Dual high (LungCVD) | 17.8% (2,832) | 170 (29.5%) 6-yr lung cancers; 226 (42.9%) 10-yr CVD deaths | |||
| Lung-only (LungCVD) | 65.3% (10,401) | 385 (67.0%) lung cancers (CVD-missed); 243 (46.5%) CVD deaths | |||
| CVD-only (LungCVD) | 0.8% (135) | 2 (0.3%) lung cancers; 12 (2.3%) CVD deaths (45.2% CVD total) | |||
| Dual low (LungCVD) | 16.1% (2,563) | 18 (3.2%) 6-yr lung cancers; 44 (8.3%) 10-yr CVD deaths |
Table A22. Joint cardiopulmonary triage in the NLST test cohort: combined net benefit of eight strategies and risk quadrants (; LDCT arm ).
The same properties held across single-hospital training, baseline CT expansion, and incomplete records. Trained at a single hospital, CE-Agent won 246 of 264 external comparisons with same-hospital ML baselines (Figure 8b). Among held-out LDCT participants (), adding radiologist-reported nodules and visual emphysema [7] raised AUC from 0.685 to 0.708 (; Figure 8c). With up to 60% of predictors missing in the training cohort (, using published weights), reliability scoring kept calibration slopes between 0.91 and 1.07 (Figure 8d).
Pairing the lung cancer and cardiovascular predictions from the same screening encounter identified a substantial dual-risk subgroup (Figure 8e,Figure 8f and Table A22). At cutoffs of 1.5% 6-year lung cancer risk and 5.0% 10-year fatal-CVD risk [48, 49, 50], participants above both thresholds comprised 17.8% of the held-out cohort yet accounted
for 42.9% of cardiovascular deaths and 29.5% of lung cancers. Whereas US Preventive Services Task Force (USPSTF) criteria [3] qualified nearly all participants (99.4%), CE-Agent focused lung follow-up on 83.1% while capturing 96.5% of 6-year lung cancers.
With CE-Agent, the combined net benefit of lung cancer and cardiovascular decisions reached +30.2 per 1,000, outperforming USPSTF criteria paired with recalibrated PCE (+27.5 per 1,000, ; Figure 8f). Selective two-stage reading—scoring CT features only when a participant’s pre-CT risk interval crossed either cutoff—required CT reads for 61.1% of participants while preserving full-scan net benefit (difference −0.58 per 1,000, 95% CI −1.26 to 0.00, within the prespecified −2.0 criterion). Within the LDCT arm (), where the screening scan itself supplies all CT inputs, CE-Agent again outperformed USPSTF criteria plus recalibrated PCE (+33.4 versus +30.8 per 1,000, ; Table A22).
Discussion
Risk stratification from routine clinical records and chest CT needs models that respect clinical timing, learn reliably from a single hospital’s cohort, and quantify uncertainty when records are incomplete. CE-Agent compiles peer-reviewed epidemiological evidence into a Causal Envelope before any outcome fitting and pairs constrained estimation with closed-form uncertainty propagation. The result is calibrated, uncertainty-aware prediction of cardiovascular risk in a national lung-screening trial and a multi-ethnic primary-prevention cohort, and of lung cancer risk in the screening trial.
Published epidemiological bounds guide a model differently from generic statistical shrinkage (Figure 2b). Generic shrinkage penalties [38, 51] pull every coefficient toward zero regardless of medical evidence, leaving small-sample models exposed to variance and to confounding-driven sign reversals [27]. The Causal Envelope instead removes post-scan events and collinear composite indices before fitting, holds established risk factors within their published ranges, and lets baseline incidence and the effect sizes of newly added CT phenotypes adapt to the local cohort. In small cohorts, the constraints leave fewer free parameters and raise expected discrimination, so established sample-size criteria [23] are met with a 2.5- to 5.4-fold smaller development sample, enough for 12 of 33 NLST hospitals to develop a model on their own data. In large multi-center cohorts, the constrained model matches the discrimination of unconstrained and accounted for more than two in five cardiovascular deaths. Where binary eligibility criteria qualified nearly every participant, individual lung cancer risk focused follow-up on five in six while capturing 96.5% of 6-year lung cancers (Table A22). In the LDCT arm, both assessments drew on nodule, emphysema, and coronary calcium descriptors from the screening scan itself, with no additional radiation or scan time. Reading CT only for borderline cases—about three in five participants—preserved the net benefit of reading every scan.
CE-Agent’s modular design opens several directions for clinical translation. First, risk is computed from a weighted sum of clinical and CT phenotypes, which clinicians can audit and which supports closed-form uncertainty propagation. This layer can sit directly on automated 3D thoracic and abdominal CT segmentation [13, 52, 53] and on imaging foundation models [20] for screening, inpatient, and emergency scans. Second, building on transfer across MESA groups, automated prior compilation can incorporate international multi-ancestry registries, proteomic panels [21], and genomic panels. Third, extending validation beyond the two US cohorts studied here, prospective studies that add clinic vital signs, laboratory panels, longitudinal health records, and serial chest CT to screening programs can measure real-world effects on nodule management, statin prescribing, and cardiopulmonary outcomes.
In summary, CE-Agent shows that peer-reviewed epidemiological knowledge can be compiled into explicit constraints and paired with real-time uncertainty propagation at deployment. Grounding estimation in accumulated medical evidence yields cardiovascular models that stay calibrated and generalize from a single hospital or CT scanner to the others, alongside calibrated lung cancer prediction matching established screening models. It keeps risk directions consistent with established biology, accommodates incomplete outpatient records, and turns a single chest CT into an opportunistic dual-disease screening exam.
Methods
Ethics oversight
This study analyzed de-identified secondary data from two National Institutes of Health (NIH) clinical cohorts conducted in accordance with the Declaration of Helsinki. NLST (ClinicalTrials.gov identifier NCT00047385) was approved by the institutional review board at each of the 33 participating US screening centers, and all participants provided written informed consent [1]. MESA (NCT00005487) was approved by the institutional review boards of all six participating field centers and the National Heart, Lung, and Blood Institute (NHLBI) coordinating center, with written informed consent obtained from all participants [36]. Access to de-identified NLST and MESA datasets was granted under approved National Cancer Institute (NCI) Cancer Data Access System and NHLBI Biologic Specimen and Data Repository Information Coordinating Center (BioLINCC) data-use agreements.
Problem formulation and two-stage architecture
We consider a training cohort of participants with baseline clinical and chest CT records across candidate variables and adjudicated binary cardiopulmonary outcomes . NLST models are fitted to event status over follow-up and aligned to the 6- or 10-year competing-risk horizon, whereas MESA models are fitted directly to 10-year status, with deaths from other causes coded as non-events and participants censored earlier excluded (Appendix C). Within our cardiopulmonary agentic framework, CE-Agent forms the deterministic, fixed-path core executing specialized clinical-record-reading, CT-biomarker-extraction, prior-compilation, constrained-estimation, and uncertainty-propagation modules in two stages (Figure 1). The clinical-record reader maps each cohort’s baseline fields to model variables, enters recorded values as observed, and flags missing fields, which carry their population prior into the uncertainty propagation described below. In Stage I (Figure 1a), structural screening selects admissible baseline predictors (re-indexed as ). Each observed entry is standardized to using center and positive scale (empirical training moments in MESA and pre-specified cohort scales in NLST; Appendix C), forming . Verified published effect ranges are then compiled into a convex parameter set, the Causal Envelope , before viewing outcome labels. In Stage II (Figure 1b,Figure 1c), CE-Agent estimates the unpenalized baseline intercept and standardized log-odds weights within . At deployment, analytic variance propagation outputs calibrated risk , 90% predictive interval width , and missing-data log-odds uncertainty .
Causal structural screening and missing-biomarker projection
Candidate variables are organized on a domain directed acyclic graph (DAG) encoding temporal ordering and biological hierarchy (Appendix B). Structural screening excludes post-baseline events to prevent outcome leakage and separates derived linear composite scores from directly measured intermediate biomarkers. Any derived linear composite formed with fixed weights from retained standardized parent columns () lies in the column space of . Excluding therefore leaves full-rank risk predictions unchanged while preventing small-sample collinearity and Fisher information singularity (Proposition 1, Appendix B). Ratio indices such as BMI, which are nearly collinear with their retained parents, are screened by the same rule in the NLST cardiovascular model. By contrast, directly measured intermediate biomarkers with standardized column —including blood pressure, lipids, glucose, high-sensitivity C-reactive protein (CRP), kidney function, chronic respiratory status, and thoracic CT phenotypes—capture subclinical biological variation beyond upstream clinical factors. Chronic systemic inflammation, represented by a pre-specified linear summary of smoking exposure and BMI (the upstream inflammation score), enters the cardiovascular models as a single column; BMI, which is not a direct predictor in the NLST cardiovascular model, contributes through this score. When a continuous intermediate biomarker is unmeasured at deployment, CE-Agent projects its conditional mean and variance recursively from its parents, using sign-constrained training weights , parent moments equal to when observed, and the training residual variance :
Here , and the variance recursion treats parent projection errors as independent. Binary intermediate biomarkers analogously use sign-constrained logistic projection and delta-method variance (Appendix B). When all chest CT biomarkers are unmeasured at simulated outpatient visits, CE-Agent substitutes the unattenuated expectations , which preserve the cohort risk scale; variants attenuate these projections by their variance or by the reliability score defined below (Figure A2d and Table A13).
| Inference Regime | Full | AUC | Brier | E/O (95% CI) | Slope |
| 1. Observed CT | 0.8046 | 0.8068 | 0.02390 | 1.07 (0.84–1.47) | 0.777 |
| 2. Population-Mean Fill () | 0.7872 | 0.7954 | 0.02363 | 0.81 (0.63–1.11) | 0.907 |
| 3. Projected CT (rel.-gated) | 0.7884 | 0.7961 | 0.02360 | 0.84 (0.65–1.16) | 0.884 |
| 4. Projected CT (var.-gated) | 0.7903 | 0.7970 | 0.02359 | 0.95 (0.74–1.31) | 0.826 |
| 5. Projected CT (unattenuated) | 0.7904 | 0.7971 | 0.02360 | 0.96 (0.74–1.32) | 0.823 |
| 6. Retrained No-CT (18 features) | 0.7903 | 0.7983 | 0.02356 | 1.06 (0.82–1.45) | 0.926 |
| 7. External ML (84 configs) | 0.673–0.800 | 0.676–0.806 | 0.0235–0.0258 | 0.59–1.19 | 0.43–3.09 |
Table A13. MESA cardiovascular mortality when all five CT phenotypes are unavailable at prediction (). Row 1 provides the observed-CT reference; bold values denote best CE-Agent or retrained single-model missing-CT performance.
Literature prior compilation and compatibility screening
Published epidemiological studies report hazard, odds, or risk ratios across clinical increments , such as smoking intensity, hypertension status, alcohol intake, educational attainment, particulate air pollution, chronic respiratory disease, or thoracic and coronary calcification increments [6, 8, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63]. Supporting publications for each predictor–outcome link were cataloged before model fitting, with PubMed abstracts and source tables archived with identifier, access date, and checksum. In the pooled NLST cardiovascular specification, regular-expression verification against archived verbatim PubMed abstracts extracts the published log-effect interval and point estimate directly [55, 56, 58]; lung-cancer and point-anchored specifications additionally compile verified intervals and point priors from archived full-text tables and meta-analyses [6, 8, 54, 57, 59, 60, 61, 62, 63]. When several archived estimates qualified for one predictor, the estimate with the longest follow-up, and then the narrowest interval, was selected, and alternatives were logged with the selection rationale. Hazard, odds, and risk ratios all serve as log-odds effects because the outcomes are rare (cumulative incidence below 10%) [64].
CE-Agent rescales these effects to standardized local cohort units via and records the established biological direction . A four-criterion compatibility check (Appendix B) evaluates study population compatibility, covariate over-adjustment, unit-scale alignment, and anatomical match. Predictors with verified directional evidence form , while predictors without it, such as sex and race/ethnicity in the lung cancer model and cardiac chamber area and aortic diameter in the MESA cardiovascular mortality model, remain unconstrained in . A new CT phenotype with an expected direction but no compatible published effect size, such as the PA diameter ratio, carries the sign constraint alone, so its effect size is estimated from local data. Within , predictors passing all four checks form the interval-bounded subset with rescaled bounds . Established clinical risk factors fixed at published point estimates form , and predictors flagged on any compatibility check retain directional sign constraints in . A fixed weight is either or, in single-hospital NLST fits and in MESA, a weight obtained by numerically differentiating published risk equations [15, 16, 19]. For intercept and weight vector , the compiled Causal Envelope is (Figure 1a and Figure 2a):
In primary pooled multi-center NLST estimation of cardiovascular (, three interval bounds) and lung cancer (, five interval bounds) risk, event counts far exceed , so predictors with compatible published intervals enter and is empty. A sign-only lung cancer variant keeps directional constraints on eight predictors without the PLCOm2012 numerical intervals (Appendix C). In single-hospital NLST fits and in MESA (), core clinical risk factors are fixed in , which reserves local degrees of freedom for pulmonary, laboratory, and CT biomarkers. The point-anchored NLST specification used for single-hospital results fixes 9 of its 13 weights (eight clinical and CAC factors and the upstream inflammation score) and fits 4 behavioral and pulmonary predictors within . The prior-locked specification fixes every weight at its published value and sets only the intercept from training prevalence. Because the primary CAC weight was derived in NLST [8], single-hospital fits were repeated with a CAC weight from an independent screening trial [65]; external discrimination changed by less than 0.001 AUC (Figure A3c).
Constrained risk estimation and sample-size efficiency
We define the directional sign cone , linear log-odds , and empirical logistic negative log-likelihood . CE-Agent computes by solving the convex optimization problem:
Here is a continuously differentiable squared-hinge penalty outside , and remains unpenalized (in NLST competing-risk evaluations, is subsequently shifted by an additive constant so mean predicted risk matches training-cohort Aalen–Johansen [66] cumulative incidence at the 6- or 10-year horizon; Appendix C). The coefficient provides optional weak per-SD ridge stabilization solely on predictors without interval or point priors ( in primary fits and in small-sample MESA subcohorts; Appendices B–C). Taking yields the primary hard-bound estimator with exact membership and zero hyperparameter tuning, whereas finite keeps the total prior precision fixed, so its influence relative to the data declines as the cohort grows. The penalty path of Figure 3a,Figure 3d applies the box and point penalties without the sign cone and keeps only the upstream inflammation score fixed, so is the otherwise unconstrained maximum-likelihood fit. By Karush–Kuhn–Tucker (KKT) optimality conditions, directional sign constraints preserve concordant empirical associations and set on sign-only predictors whenever the empirical gradient favors the sign opposite to , as when confounding reverses the association (Figure 2b and Proposition 2). Restricting parameters to lowers the minimum development sample size of Riley et al. [23] through two routes: fewer effective degrees of freedom () and higher anticipated discrimination of the constrained model, with the anticipated Cox–Snell converted from each specification’s median held-out area under the receiver operating characteristic curve (AUC) [67] (equation (11)).
Analytic uncertainty propagation and selective clinical triage
At deployment, each predictor in patient record carries mean , missing-data variance , and missingness indicator , taking when observed and imputed or projected moments when missing. Propagation runs over either the Stage II fitted weights or the Stage I graph, which carries published weights and an intercept set to training prevalence (Appendices B–C). With denoting the logistic sigmoid, a single reverse-topological pass with out-of-fold gate parameters computes each predictor’s reliability gate , reported as the reliability score in Results. The same pass returns the total pathway sensitivities , the summed effect of predictor on the log-odds over all directed paths to the outcome. Each link weight entering is multiplied by the gate of its source predictor (Appendix B), so missing or uncertain predictors contribute less to the propagated log-odds mean and missing-data variance :
where captures latent-pathway residual variance when latent nodes are modeled (Appendix B). With training-prevalence intercept variance , fixed-size Gauss–Hermite quadrature over and safeguarded Newton–bisection inversion of the log-odds mixture cumulative distribution (; full mixture quadrature in equation (9), Appendix B) yield calibrated risk , missing-data log-odds uncertainty , and 90% predictive interval width :
By first-order delta expansion, , where is the standard deviation of the predicted log-odds, combining baseline, missing-data, and parameter variance (equation (B5)). This width is largest near whether or not a record is complete. Uncertainty-guided chart review therefore ranks records by —or its cohort-measured missing-predictor component (Appendix B)—to isolate sensitivity to unmeasured features, and two-stage CT triage reads CT features only when the 90% CT-uncertainty interval, on the log-odds scale with all CT-derived predictors treated as unmeasured, contains a decision threshold (Appendix C).
Per-patient runtime scales linearly with the number of DAG variables and directed links . Monte Carlo propagation with draws redraws every input node and re-evaluates every link for each draw and then sorts the outcomes, in time. CE-Agent instead accumulates pathway sensitivities in one reverse-topological pass over a fixed parameter grid, and the integration and quantile root-finding that follow operate on a fixed-size scalar mixture whose cost does not depend on or , giving time per patient. In single-record scoring, each added variable cost 0.016 ms, against 6.49 ms for 4,000-draw sampling (approximately 400-fold lower), and total runtime was 83- to 377-fold lower (Figure 7a and Appendix B).
Study cohorts, benchmarks, and statistical analysis
NLST enrolled 53,454 heavy smokers aged 55–74 years across 33 US screening hospitals [1]; we analyzed the 53,452 participants in the released trial dataset. Cardiovascular death was a death whose recorded cause fell in the circulatory chapter (I00–I99) of the International Classification of Diseases, 10th revision. Causes of death came from the trial’s extended follow-up, which includes National Death Index linkage through 2015; deaths from other causes were competing events. Lung cancer incidence was a confirmed diagnosis, and lung cancer death followed the trial’s endpoint verification process where available and the death certificate otherwise. Cardiovascular and lung cancer analyses shared a deterministic 70%/30% split derived from participant identifiers into 37,521 training and 15,931 held-out test participants. With up to 13.4 years of mortality follow-up, the training set had 1,789 cardiovascular and 1,663 lung cancer deaths, and the test set had 750 and 747. Over the trial’s cancer-incidence follow-up (median 6.6 years), 1,440 training and 618 test participants developed lung cancer. In the test set, 14,971 participants had uncensored 10-year vital status, including 525 cardiovascular and 1,684 competing deaths (Table A1).
Within the baseline LDCT arm, 18,519 training and 7,786 held-out participants experienced 732 and 338 full-follow-up incident lung cancers, respectively. Lung cancer evaluations assessed 6-year competing-risk incidence (1,350 training and 575 held-out cases within 6 years; LDCT arm 694 and 317 cases, accounting for pre-diagnosis mortality) and 10-year competing-risk mortality (1,242 training and 535 held-out deaths within 10 years) across 11 screened baseline predictors. Joint cardiopulmonary triage paired 6-year lung cancer thresholds (1.5% primary and 2.0% sensitivity) [50, 68] with the 5.0% 10-year fatal-CVD threshold [48, 49]. In the NLST 3D LDCT subcohorts (5,269 training participants with soft-tissue reconstructions and 6,467 with any reconstruction kernel), thoracic aortic calcium volume, left ventricular volume, proximal main-to-sinotubular-junction PA diameter ratio, and quantitative emphysema were measured automatically on baseline chest CT from TotalSegmentator [52] and lungmask [69] segmentations in a mean of 45.6 s per scan on one graphics processing unit (GPU; Table C1).
| Pipeline stage | Extracted output | Hardware | Mean (median) | Range / mode |
| DICOM to NIfTI conversion | 3D NIfTI volume | CPU | 2.5 s (2.2 s) | Shared input |
| Sybil [80] | 3D lung-cancer CNN (runtime ref.) | 1× RTX A6000 | 3.2 s (3.2 s) | Concurrent |
| lungmask [69] | 3D emphysema (LAA) | 1× RTX A6000 | 7.6 s (7.2 s) | Concurrent |
| TotalSegmentator (total) | Thoracic aortic calcium ( HU) | 1× RTX A6000 | 26.8 s (25.8 s) | Cumulative |
| TotalSegmentator (heartchambers_highres) | Left ventricular chamber volume | 1× RTX A6000 | 30.6 s (29.4 s) | Cumulative |
| TotalSegmentator (pulmonary_report) | Pulmonary artery diameter ratio | 1× RTX A6000 | 45.6 s (45.7 s) | 39.4–51.7 s |
| CE-Agent scoring & uncertainty | Calibrated , , | 1 CPU core | 1.10–1.23 ms | Batch (2.1–2.6 ms single) |
Table C1. Per-scan runtime of automated CT feature extraction and risk scoring.
MESA [36, 37] Exam 1 enrolled 6,814 White, Chinese American, Black, and Hispanic adults aged – years free of clinical CVD at baseline, with a median follow-up of 18.9 years (maximum 20.9 years) in 6,795 participants. Electrocardiogram-gated cardiac CT provided five phenotypes: CAC, TAC, cardiac chamber area, ascending aortic diameter, and liver attenuation. Cardiovascular death was an adjudicated death attributed to coronary heart disease, stroke, or other atherosclerotic or cardiovascular disease; deaths from other or unknown causes were competing events. Using the same deterministic 70%/30% split, cardiovascular mortality analyses comprised 4,821 training and 1,974 held-out test participants (4,731 and 1,932 with uncensored 10-year status; 49 test cardiovascular deaths within 10 years and 147 overall, with 168 competing deaths within 10 years). Adjudicated composite hard ASCVD (myocardial infarction, stroke, or death from coronary heart disease or stroke) analyses comprised 4,812 training and 1,971 test participants (4,301 and 1,769 with uncensored 10-year status; 137 test events within 10 years and 283 overall, with 95 competing deaths within 10 years; Table A2 and Table A6). Two-stage cardiovascular CT triage was evaluated in both cohorts (Table A19).
The NLST analysis used baseline CVD-free LDCT participants in the training cohort (), each scored with published weights and an intercept estimated at the other 32 hospitals; 1,256 participants with diabetes were routed directly to statins and 14,471 candidates were triaged. The MESA analysis used test adults aged – years not on lipid-lowering therapy at baseline (), with the no-CT risk recalibrated in the large to the training event rate; 178 participants with diabetes or LDL mg/dL (4.9 mmol/L) were routed directly to statins and 1,287 candidates were triaged.
Clinical comparators comprised PLCOm2012 [6] and 2021 USPSTF criteria [3] for lung cancer, alongside Framingham general CVD [15], ACC/AHA PCE [16], and the AHA PREVENT base 10-year ASCVD equation [19] for CVD, evaluated with their published coefficients and after training-prevalence logit recalibration. The cardiovascular calculators were built for composite endpoints that include nonfatal events (general CVD for Framingham; ASCVD for PCE and PREVENT). Calibration with published coefficients was therefore assessed on MESA’s adjudicated composite endpoints matched to each calculator: hard ASCVD for PCE and PREVENT, and for Framingham the broad cardiovascular composite, the closest counterpart of general CVD (Table A2 and Table A6). For cardiovascular death, the calculators were compared on discrimination, and their calibration is reported after recalibration. Because NLST recorded no blood pressure, lipid, or renal measurements, NLST evaluations supplied the three cardiovascular calculators with each participant’s recorded age, sex, current smoking status, and diabetes history (plus race/ethnicity in PCE) alongside standard US untreated adult reference values: 197 mg/dL (5.1 mmol/L) total cholesterol, 50 mg/dL (1.3 mmol/L) high-density lipoprotein (HDL) cholesterol, 130 mmHg systolic blood pressure, and 90 mL/min/1.73 m estimated glomerular filtration rate (eGFR; PREVENT only). The HDL cholesterol, blood pressure, and eGFR values are the centering values of the PREVENT equations [19]. In MESA Exam 1, all three calculators used directly measured seated blood pressure, fasting lipids, and the 2009 Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) eGFR supplied with the MESA dataset [70]. In both cohorts, antihypertensive and statin indicators were set to untreated, so that each calculator returns pre-treatment risk.
Machine learning comparators comprised Cox proportional hazards (Cox PH) [71], random survival forests (RSF) [72], gradient-boosted survival analysis (GBSA) [73], XGBoost and XGBoost-Cox [74], DeepSurv neural networks [75], decision trees, random forests [76], and logistic regression (LR) on the CE-Agent predictor set and on expanded feature sets; single-hospital NLST and single-group MESA comparisons used the eight baselines listed in Table A8 and Table A9. Domain-generalization and prior-informed comparators included Anchor Regression [42], and conditional value-at-risk (CVaR) distributionally robust optimization (DRO) [43], prior-offset logistic model updating [44] and its boosted-tree extension, and width-matched Gaussian maximum a posteriori (MAP) priors (Appendix C). Performance across 1–15 years was evaluated using Harrell’s C-index [77] for the cause-specific event (deaths from other causes censored), inverse-probability-of-censoring weighted time-dependent AUC (t-AUC) [41], and competing-risk [40]. Calibration, overall accuracy, and clinical utility were summarized by competing-risk and Aalen–Johansen , binary Brier score [78], logistic calibration slope [14], and decision curve analysis (DCA) net benefit [45] (equation (C1)). Tests were two-sided, and 95% confidence intervals came from 1,000-replicate participant-level bootstrap resampling unless stated otherwise. For lung cancer discrimination against recalibrated PLCOm2012, preregistered noninferiority required the lower limit of the two-sided 95% bootstrap interval of the difference to exceed −0.02, and for 6-year incidence also a positive point estimate. For two-stage against full CT reading, it required the lower limit of the net-benefit difference to exceed −2.0 per 1,000 participants, with fewer than 70% of scans read and a positive point difference from equal-budget random reading. In NLST, primary competing-risk cardiovascular evaluation, equivalence testing between compiled and point-anchored literature bounds (Table A1), and the lung cancer and joint cardiopulmonary triage analyses were preregistered in time-stamped, version-controlled analysis plans before the held-out test split was scored once on each endpoint. External MESA replication and mechanistic or pathway analyses outside those plans, including single-hospital lung cancer fits, added radiologist CT findings, cardiovascular CAC triage, LDCT-arm triage, and missing-data propagation benchmarks, are secondary, exploratory, or post hoc analyses.
Reporting summary
Further information on research design, cohort eligibility, and statistical evaluation protocols is provided in Appendices § 5–C and the Nature Portfolio Reporting Summary linked to this article.
Data availability
Individual-level clinical and chest CT data from NLST are available to approved researchers upon application through the NCI Cancer Data Access System (https://cdas.cancer.gov/nlst/). MESA Exam 1 clinical, laboratory, CT, and adjudicated cardiovascular outcome data are available to qualified investigators through the NHLBI BioLINCC repository (https://biolincc.nhlbi.nih.gov/studies/mesa/) and the MESA Coordinating Center (https://www.mesa-nhlbi.org/). All PubMed abstracts and epidemiological effect estimates compiled in Stage I are publicly accessible and archived in the project repository.
Code availability
Source code for CE-Agent, automated PubMed prior compilation, constrained convex estimation, deterministic Gauss–Hermite uncertainty propagation, figure generation, and cross-cohort benchmark reproduction is publicly available at https://github.com/Yuxiang-Lai117/CP_Agent_Paper. Baseline models and evaluation ran in Python 3.13 with NumPy 2.5.1, SciPy 1.18.0, scikit-learn 1.9.0, scikit-survival 0.28.0, lifelines 0.30.3, XGBoost 3.3.0, and PyTorch 2.13.0; CT phenotypes were extracted with TotalSegmentator 2.17.0 and lungmask 0.2.20 (Appendix C).
Acknowledgements
We thank the participants, investigators, and study staff of the National Lung Screening Trial and the Multi-Ethnic Study of Atherosclerosis, as well as the National Cancer Institute (NCI), the National Heart, Lung, and Blood Institute (NHLBI), and the National Institutes of Health (NIH) for providing access to trial and cohort resources. The manuscript reflects the views of the authors and does not necessarily represent the official positions of the NCI, NHLBI, or NIH. This research is supported in part by the National Institutes of Health under Award Numbers 1R01HL183223, R01DE033512 and R01CA272991.
Author contributions
Y.L., L.Z., and X.Y. conceived the study and designed the Causal Envelope methodology. Y.L. and J.Z. implemented the automated literature compiler, constrained estimation solver, analytic uncertainty propagation module, and multi-cohort evaluation pipelines. Y.L. and J.Z. performed the statistical, cross-institution, and clinical triage analyses across NLST and MESA. Y.L. drafted the manuscript and designed the visualizations with critical revision from J.Z., L.Z., and X.Y. L.Z. and X.Y. supervised the project. All authors approved the final manuscript.
Competing interests
The authors declare no competing interests.
Extended results
Abbreviations in Tables A1–A22: AUC, area under the receiver operating characteristic curve; -AUC, time-dependent AUC; C, Harrell’s concordance index; E/O, expected-to-observed ratio; cr, competing-risk; CAC, coronary artery calcification; LDCT, low-dose CT; PCE, Pooled Cohort Equations; LR, logistic regression; RSF, random survival forest; XGB, XGBoost; OOD, out-of-distribution; , 90% interval width; , missing-input log-odds uncertainty; USPSTF, US Preventive Services Task Force.
Held-out accuracy, composite endpoints, and transfer
Tables A1–A7 extend Figure 3. They compare CE-Agent with clinical calculators, its own point-anchored and prior-locked variants, and seven survival and ML model families, on held-out cardiovascular mortality and adjudicated composite endpoints; the last table applies NLST-trained models to MESA.
Note: Bold marks the best value in each column (|E/O closest to zero for calibration). 95% CIs are percentile bootstrap intervals over held-out participants (-index: 20,000 replicates for CE-Agent rows and 200 for baselines and calculators, in place of the default 1,000). Brier, full-follow-up binary Brier score; RSF, random survival forest; ACC/AHA ASCVD-PCE, American College of Cardiology/American Heart Association Pooled Cohort Equations; PREVENT, Predicting Risk of Cardiovascular Disease Events; CV , soft-bound penalty selected by five-fold inner cross-validation; point-anchored bounds, established clinical factors anchored at rescaled literature point estimates, with imaging and pulmonary predictors bounded in ; prior-locked, weights fixed at published point estimates without outcome fitting. Preregistered equivalence of compiled and point-anchored bounds (compiled minus point-anchored; paired bootstrap 95% CI): (−0.0039 to +0.0072; bounds 0.01) and Brier (−5.16 to −1.30 ; bounds 0.01); both intervals lie within their bounds. The 10-yr competing-risk Brier score (Brier) is 0.0320 for compiled CE-Agent, logistic regression, and random forest, 0.0321 for XGBoost, and 0.0322 for Cox PH. For logistic regression, XGBoost, and the forest row, -index and Brier come from models trained on full-follow-up labels, and AUC and E/O from classifiers trained on 10-yr status; the forest row pairs a random survival forest with a random forest classifier. CE-Agent competing-risk predictions apply a logit shift aligned to the training-cohort Aalen–Johansen 10-yr incidence; Cox PH reports competing-risk cumulative incidence. Clinical calculators use published coefficients, with US untreated adult reference values for unmeasured blood pressure, lipids, and eGFR. They target composite endpoints that include nonfatal events, so Brier score and E/O are not scored against cardiovascular death (matched-endpoint calibration in Table A2 and Table A6); after training-prevalence logit recalibration all three reach E/O (Methods).
Note: Bold marks the best value in each column per endpoint ( closest to zero for calibration). CAC, coronary artery calcification; ASCVD, atherosclerotic cardiovascular disease; ACC/AHA ASCVD-PCE, American College of Cardiology/American Heart Association ASCVD Pooled Cohort Equations; PREVENT, Predicting Risk of Cardiovascular Disease Events. †Binary CAC coding has complete-case 10-yr E/O = 1.07 (Table A4); competing-risk E/O was evaluated on the primary continuous specification with training-cohort Aalen–Johansen logit anchoring. On mortality, the 10-yr discrimination of CE-Agent is comparable to that of the published clinical equations (). ‡The published equations target composite cardiovascular endpoints that include nonfatal events, so their Brier score and E/O are reported on the matched endpoint (PCE on hard ASCVD, lower block, where published PREVENT has E/O; Framingham in Table A6) and not against cardiovascular death; after train prevalence-logit recalibration the equations reach E/O–1.06 on mortality and 0.94 for PCE on hard ASCVD. Cox PH reports competing-risk cumulative incidence. 22-predictor vs. 58-predictor (63-column) ML benchmarks are in Table A5.
Note: Int. -AUC / CS, integrated competing-risk and cause-specific time-dependent AUC across 1–13 yr; E/O, 10-yr Aalen–Johansen competing-risk expected-to-observed ratio (E/O = 0.99 on uncensored 10-yr status across all CE-Agent models); CIF, competing-risk cumulative incidence function (unadjusted cause-specific Cox PH: AUC = 0.7192, slope 1.797, E/O = 1.11, NB = 6.2); NB, net benefit () at risk threshold (95% CI 4.7–8.8 for hard/soft CE-Agent); ICI, integrated calibration index (0.0019 for CE-Agent, 0.0024–0.0109 for pooled ML, and 0.0010–0.0066 for recalibrated calculators). Discrimination of CE-Agent is comparable to that of the 33-center pooled ML models ( across paired replicates; full-follow-up C-index 0.7306 for CE-Agent, 0.7327–0.7337 for 10-yr-fitted pooled LR/XGBoost/RF, and 0.7278 for Cox PH). The published calculators target composite cardiovascular endpoints that include nonfatal events, so E/O against cardiovascular death is not reported for them (1.02 after training-prevalence recalibration; Table A1), and their NB and ICI are given after recalibration.
Note: , 10-yr Harrell’s -index; , full-follow-up Harrell’s -index; {sup}†primary specification; {sup}§evaluated under the locked binary CAC prior (; Figure A3g). {sup}∗Zero-locked ridge is a control ablation that shrinks the 13 anchored and interval-bounded clinical and literature predictors toward zero (inflammation locked at its published point weight) while fitting the 8 unconstrained predictors ($p_{\mathrm{eff}} = 8.0$). Kaplan–Meier {citep}[79] net E/O is 1.03 for CE-Agent and unconstrained fits, 1.04 prior-locked, and 1.01–1.02 for the calculators after prevalence recalibration. {sup}‡The published calculators target composite cardiovascular endpoints that include nonfatal events, so their E/O against cardiovascular death is not reported; calibration on matched endpoints is in Table A2 and Table A6. Brier score is 0.02389 for ordinal CAC and 0.02459 for zero-locked ridge (other models in Table A2). Integrated cause-specific -AUC spans 1–15 yr, weighted by Kaplan–Meier survival increments (competing-risk integrated -AUC vs. 0.7986 unconstrained); hard-bound CE-Agent reaches 0.842, PREVENT 0.843, and unconstrained estimation 0.825 (difference +0.018; 95% CI −0.004 to +0.040). Dashes, integrated -AUC not computed.
Note: Bold values denote column-best baseline performance ( closest to zero for calibration). Abbreviations: , 10-yr Harrell’s -index; Int. -AUC, integrated cause-specific time-dependent AUC (1–15 yr, weighted by Kaplan–Meier survival increments as in Table A4 and Table A6) across the 22 predictors.
Note: Bold marks the best value in each column per endpoint (|slope − 1| and |E/O − 1| closest to zero for calibration). On hard ASCVD, the integrated -AUC of CE-Agent (hard bounds) exceeds that of PCE by 0.025 (P = 0.010). 10-yr concordance equals 0.7671 (hard) and 0.7692 (soft) vs. 0.7459 (PCE) on Hard ASCVD, 0.7665 vs. 0.7426 on Hard CVD, and 0.7776 vs. 0.7442 (PCE) and 0.7567 (Framingham) on Broad CVD; the rightmost column reports complete-case 10-yr E/O (competing-risk E/O hard ASCVD in Table A2). PCE targets hard ASCVD; Framingham targets general CVD, for which broad CVD is the closest adjudicated composite, and on which the integrated -AUC of CE-Agent exceeds that of Framingham by 0.025 (P = 0.012).
Note: Bold values denote overall column-best performance (|slope − 1| and |E/O − 1| closest to zero); denotes full-follow-up Harrell’s -index and E/O denotes complete-case 10-yr expected-to-observed ratio (— on RSF denotes nonparametric hazard ranking without a parametric logit intercept for prevalence recalibration).
Generalization, sample size, and clinical utility
Figure A1 and Table A8–Table A11 extend Figure 4. They report out-of-distribution (OOD) performance across hospitals, racial/ethnic groups, and scanners, scaling with training size, minimum sample size [23], and net benefit [45].
| MESA Regime | CE-Agent | Unc. | Cox PH | RSF | GBSA | XGB-Cox | XGB |
| Finite-sample mortality discrimination (mean AUC, 20 draws) | |||||||
| (6 deaths) | 0.7173 | 0.5549 | 0.7526 | 0.7487 | 0.7106 | 0.7055 | 0.5000 |
| (13 deaths) | 0.7740 | 0.6789 | 0.7800 | 0.7667 | 0.7319 | 0.7445 | 0.7063 |
| (26 deaths) | 0.7950 | 0.7429 | 0.7886 | 0.7672 | 0.7433 | 0.7605 | 0.7299 |
| (52 deaths) | 0.8000 | 0.7659 | 0.7951 | 0.7674 | 0.7465 | 0.7663 | 0.7250 |
| Full Train () | 0.8068 | 0.7886 | 0.7972 | 0.7832 | — | 0.7793 | 0.7658 |
| Cross-group OOD AUC, slope error (), sample-size criteria, and DCA | |||||||
| Hispanic OOD () | 0.7776 | 0.7184 | 0.7835 | 0.7717 | 0.7535 | 0.7919 | 0.7326 |
| Scanner OOD () | 0.7848 | 0.7238 | 0.7079 | 0.7823 | 0.7352 | 0.7769 | 0.7381 |
| LOGO Mean (CVD Death) | 0.8051 | 0.8046 | 0.8054 | 0.7965 | 0.7946 | 0.8215 | 0.7827 |
| LOGO Mean (Hard ASCVD) | 0.7765 | 0.7846 | 0.7773 | 0.7734 | 0.7619 | 0.7773 | 0.7672 |
| () | 0.720 | 0.973 | 0.160 | — | — | — | — |
| () | 0.418 | 0.775 | 0.125 | — | — | — | 0.378 |
| () | 0.344 | 0.516 | 0.272 | — | — | — | 0.448 |
| (Hispanic) | 0.330 | 0.513 | 0.219 | 0.051 | 0.727 | 0.391 | 0.459 |
| (Scanner) | 0.400 | 0.677 | 0.620 | 0.397 | 0.090 | 0.497 | 0.546 |
| Riley (CVD Death) | 2{,}237 | 8,380 | 8,380 | Undef. | Undef. | Undef. | Undef. |
| Riley (Hard ASCVD) | 1{,}202 | 3,055 | 3,055 | Undef. | Undef. | Undef. | Undef. |
| Excess NB (, CVD) | 4.67 | 3.68 | 5.37 | 4.43 | 2.67 | 3.81 | 2.62 |
| Excess NB (, ASCVD) | 2.36 | 2.35 | 2.25 | 2.24 | 1.43 | 2.04 | 2.17 |
Table A11. MESA models by training size ( to 4,821) and by racial/ethnic and scanner group: CE-Agent versus 31 baseline configurations ().
Note: Bold marks the best value in each column. IQR, interquartile range across the 33 single-hospital mean external AUC values (median outside parentheses); CE-Agent Wins, single-hospital comparisons won by CE-Agent against each row’s baseline (264/264 across the eight baselines; ***, two-sided paired Wilcoxon signed-rank test); , minimum required sample size under Riley et al. [23]; Excess NB (), integrated excess net benefit () over max(treat-all, 0) across 330 institution-threshold cells. Point-anchored CE-Agent ( free slope parameters out of 12 active slopes; the thirteenth weight, on the upstream inflammation score, is fixed in every specification) requires , met by 12/33 hospitals on their own data; an unconstrained 12-parameter clinical/CT model requires 6,043, and the baselines, with 23–34 encoded predictor columns, a median of 13,275–42,300 (0/33 hospitals each). Net benefit is positive in 289/330 cells for CE-Agent (33/33 hospitals across 5%–10%) and in 123–213/330 cells for baselines. Random survival forest and Cox-boosted trees give relative hazard ranks without 10-yr absolute risks (—).
Note: Bold marks the best value in each column. IQR, interquartile range (median OOD AUC across the six folds outside parentheses); CE-Agent Wins, subgroup OOD folds won by CE-Agent against each row’s baseline (43/48, 89.6%, across the eight baselines: 24/24 scanner shifts, ; 19/24 demographic shifts, ; two-sided paired Wilcoxon signed-rank tests); , minimum required sample size on cardiovascular mortality ( vs. 22.0 unconstrained; hard ASCVD vs. 3,055, met by 3/4 racial/ethnic and 2/4 scanner subcohorts); Undef., tree and neural models without a fixed number of parameters. †Integrated excess net benefit () on hard ASCVD across decision thresholds of 1% to 15% (Figure A1d). Mortality excess net benefit on all 31 baseline configurations (153/186 mortality and 193/217 hard ASCVD wins) are in Table A11.
Note: Bold marks the best value in each row. PH, proportional hazards; RSF, random survival forest; GBSA, gradient-boosted survival analysis; XGB, XGBoost; DSurv, DeepSurv neural network; LOCO, leave-one-center-out; DCA, decision curve analysis; NB, net benefit; Mean Excess NB, median across hospitals of average excess net benefit over positive-benefit thresholds. *Unboxed (four bounds freed; Figure A3a) leaves the four interval-bounded predictors unconstrained while keeping published point anchors and sign constraints on the remaining predictors (0.6977 mean external AUC at , vs. 0.6559 for fully unconstrained 12-parameter logistic estimation); †fully unconstrained 12-parameter model (Figure A1a–Figure A1c). Survival and ML baselines report median across the 33 hospitals using 23–34 encoded predictor columns (Table A8). Slope rows hold the single-hospital () point-anchored specification fixed across ; the sample-size-adaptive relaxation converges to unit slope (1.00) as grows (Table A18). At , 28/33 centers (84.8%; 29/33 at ) have calibration slopes within [0.80, 1.25], vs. 8/33 (24.2%) unconstrained, and 12/33 (36.4%) meet on their own data, vs. 0/33 unconstrained. XGB-Cox (not tabulated) reaches 0.6211 at and 0.7298 at .
Note: Bold marks the best value in each row. Unc., unconstrained 22-predictor model; GBSA, gradient-boosted survival analysis; XGB-Cox, Cox-boosted trees; XGB, XGBoost; Undef., no fixed number of parameters; , training events (approximate mean training deaths in parentheses for ); LOGO, leave-one-group-out (4 racial/ethnic groups and 4 scanner cohorts); Excess NB, decision-curve integrated excess net benefit. With 50 rather than 20 subsampling draws, CE-Agent mean AUC is 0.7171, 0.7794, 0.7949, 0.8009 at (DeepSurv 20-draw mean 0.6888, 0.7319, 0.7460, 0.7658). Across 31 baseline configurations, CE-Agent won 58.5%–84.2% of pairwise subsample comparisons on CVD mortality (363/620 at , 464/620 at , 510/620 at , and 522/620 at ), 74.7%–89.7% on hard ASCVD (463/620, 556/620, and 550/620 at , 500, and 1,000), and 66.9%–91.4% of cross-group OOD and LOGO comparisons. With (CVD) and 9.0 (ASCVD) vs. 22 unconstrained, CE-Agent meets in 3/4 racial/ethnic and 2/4 scanner cohorts on hard ASCVD. Excess net benefit: CVD death, vs. for ASCVD-PCE (); hard ASCVD, vs. for PREVENT () and for ASCVD-PCE ().
Intermediate biomarkers and missing CT in MESA
Figure A2 and Table A12–Table A13 evaluate zero-shot cross-cohort transport from NLST to MESA (Table A7), sign-constrained estimation across MESA’s 10 directly measured intermediate clinical, laboratory, and CT biomarkers (mediators) compared against unconstrained regression and seven regularized, tree-ensemble, and neural baselines, and outpatient deployment when all five thoracic CT biomarkers are unmeasured.
| Mediator (Metric) | Test (C / U) | Wins | Confounding Audit | |
| 1. ln(TAC + 1) () | 0.2990 / 0.2990 | +0.0047 | 43.5% / 56–100% | 2/2 concordant |
| 2. CKD-EPI eGFR () | 0.2710 / 0.2771 | +0.0070 | 48.0% / 63–100% | DM: +0.093 → +0.087 |
| 3. ln(CAC + 1) () | 0.2429 / 0.2467 | +0.0204 | 75.5% / 44–99% | CRP: −0.060 → +0.015 |
| 3b. CAC > 0 (AUC) | 0.7663 / 0.7677 | +0.0127 | 77.5% / 52–100% | PM: −0.035 → −0.006 |
| 4. ln(CRP + 1) () | 0.1591 / 0.1591 | +0.0065 | 27.5% (100% ≥ U) | 2/2 concordant |
| 5. CT Liver Fat () | 0.0883 / 0.0883 | +0.0000 | 100% concordant | 1/1 concordant |
| 6. Ao. Diameter () | 0.0680 / 0.0680 | +0.0070 | 27.5% (100% ≥ U) | 2/2 concordant |
| 7. Chamber Area () | 0.0296 / 0.0296 | +0.0008 | 4.0% (100% ≥ U) | 1/1 concordant |
| 8. Systolic BP () | 0.0171 / 0.0250 | +0.0009 | 23.5% / 38–100% | Alc.: −0.088 → −0.021 |
| 9. Diabetes (AUC) | 0.6331 / 0.6331 | +0.0067 | 5.0% (100% ≥ U) | 1/1 concordant |
| 10. Respiratory Disease (AUC) | 0.5543 / 0.5543 | +0.0193 | 35.5% (100% ≥ U) | 1/1 concordant |
Table A12. Direction-constrained prediction of 10 intermediate biomarkers from their parent variables in MESA, with a confounding audit ( training, test). Constrained fitting gives no parent sign reversals, versus mean reversal rates of 10%–27% across unconstrained regression and seven ML baselines at .
Note: Bold in Test (C / U) marks the higher held-out score when C and U differ to four decimal places; bold adjusted slopes mark > 75% attenuation of confounding or a sign reversal toward the causal direction. , coefficient of determination; C, sign-constrained; U, unconstrained; , mean test gain (C − U) across 200 subsamples of size (non-negative across all 10 mediators); Wins, strict win rate vs. U / range of win rates across seven ML baselines (ridge, lasso, elastic net, random forest, gradient-boosted trees, XGBoost, and multilayer perceptron; 100% ≥ U, C matches or exceeds U on all 200 draws); Confounding Audit, standardized parent slope on reversed U edges adjusted for DAG parents only → after adding demographics and medications (raw slopes; expected sign negative for DM- and SBP-to-eGFR, positive otherwise; SBP-to-eGFR +0.007 → +0.009, Figure A2c; PM slope on continuous ln(CAC + 1); 76% of alcohol-to-SBP and 82% of PM-to-CAC confounding attenuated); CKD-EPI, Chronic Kidney Disease Epidemiology Collaboration; eGFR, estimated glomerular filtration rate; CRP, high-sensitivity C-reactive protein; BP, blood pressure; PM, fine particulate matter (≤ 2.5 µm); DM, diabetes mellitus; Alc., alcohol; Ao., ascending aortic; Respiratory Disease, self-reported emphysema or asthma. Observed parents in the domain DAG (Methods; 22 links, with binary CAC > 0 reusing the seven CAC parents): TAC (age, lipids); eGFR (age, DM, SBP); CAC (age, smoking, lipids, DM, liver fat, CRP, PM); CRP (smoking, BMI); liver fat (BMI); Ao. diameter (age, SBP); chamber area (SBP); SBP (alcohol, BMI); diabetes (BMI); respiratory disease (smoking).
Note: Unattenuated conditional projection uses in equation (1) with observed-entry gate weight ; variance-gated (var.-gated) scales by , and reliability-gated (rel.-gated) scales by . Row 6 omits the 4 direct CT columns from the 22-predictor outcome model (the fifth masked CT phenotype, liver attenuation, enters as an upstream parent of CAC). Relative to Population-Mean Fill (row 2), unattenuated projection (row 5) gives paired bootstrap changes of and (95% CI +0.118 to + 0.209) and moves calibration-in-the-large (CITL) from +0.233 to +0.046 (, 95% CI −0.200 to − 0.174); 10-year is 0.8052 and 0.8074 for rows 2 and 5, 0.8157 with observed CT, and 0.8083 for the retrained no-CT model. Row 7 spans 12 external ML architectures across 6 imputers plus retrained no-CT (84 configurations; range 0.686–0.816; Brier 0.02348–0.02579). Unattenuated CE-Agent has lower and higher than 68/84 (81.0%) configurations; reliability-gated and retrained no-CT CE-Agent have lower than 60/84 and 71/84 configurations, respectively; reliability gating gives the lowest expected calibration error (ECE = 0.0081). Histogram-based gradient-boosted decision trees with native missing-value handling reach (), E/O = 0.72, and slope 1.149.
Ablations of the Causal Envelope
Tables A14–A15 and Figure A3 report mechanistic ablations across NLST and MESA, complementing Figure 5. Each ablation removes or alters one part of the envelope: the published ranges, the CAC source [8, 37], the predictor directions, composite-index screening (Proposition 1), added CT phenotypes, or the zeroing of reversed weights (Proposition 2).
Note: Bold values highlight primary CE-Agent metrics. Abbreviations: , marginal single-bound AUC gain; , interaction for bound set from the inclusion–exclusion (Möbius) decomposition of all bound combinations; , Pearson correlation; , log-odds difference; -ratio, design-matrix condition-number ratio relative to screened matrix; and , standardized pulmonary artery (PA) diameter ratio weights with and without the direction constraint (other predictors under hard bounds); CV, cross-validation. *Across the 1,040 CV fits, proximal main PA-to-sino-tubular junction ratio is zeroed in 668/780 fits and direct ascending-aorta-normalized ratio sensitivity fits () are zeroed in 260/260 (100%).
Note: Bold values highlight primary CE-Agent metrics. Abbreviations: Unc., unconstrained model; , marginal single-bound gain; , gain with all bounds applied; , pairwise interaction from the inclusion–exclusion (Möbius) decomposition of bound combinations; bin., binary CAC coding; , log-odds difference; , design-matrix condition number; Ao. diam., ascending aortic diameter; , KKT dual multiplier satisfying complementary slackness (equation (B3)). *In Row 3, the Detrano et al. [37] coronary interval bound improves small-sample Hard ASCVD discrimination at (+0.0043, winning 40/50 draws). In full-sample Hard ASCVD (, 137 test events), Stage 1 compatibility screening assigns CAC a directional sign constraint (or soft bound) because Detrano et al. report a coronary-only endpoint, which differs from the coronary-plus-stroke composite (soft-bound , Table A6). All cell discrimination values denote 10-year ; small-sample experiments () average 20–50 independent stratified subsample draws (200 in row 7, with 6 and 13 events per draw).
Uncertainty propagation and missing data
Tables A16–A17 report missing-data benchmarks in NLST and MESA across 11 missingness settings (0%–60%; missing completely at random [MCAR] or at random given BMI [MAR]), complementing Figure 6 and Table A21.
| Evaluation Dimension | Protocol / Cohort | CE-Agent | Comparators | Key Finding |
| 1. Single-Hospital OOD () | 33 single-site fits ×8 ML baselines (264 cells) | 246 / 264 (93.2%) wins; med. per-site AUC = 0.665 (0.679 pooled external) | Sign-only: 244/264 (92.4%), med. per-site 0.657; RSF 0.633, Cox 0.635, GBSA 0.631, LR 0.620, XGB 0.611, DeepSurv 0.586 | +0.052 med. AUC across external sites () |
| 2. Multi-Center Scaling | 1, 2, 4, 8, 16, 32 hospitals (33 subsets/level) | AUC: 0.679 → 0.702 → 0.708 → 0.713 → 0.718 → 0.715 | Unconstrained 11-pred.: 0.653 → 0.694 → 0.704 → 0.713 → 0.718 → 0.708; XGB: 0.629 → 0.712 | decays 0.283 → 0.006; slope within [0.80, 1.25] in 36% → 70% → 88% → 100% → 100% → 70% |
| 3. Domain-Gen. & Riley | external sites vs. DRO / Anchor / updating | Med. 0.679; Riley = 3,296 () | Anchor: 0.633; -DRO: 0.617; CVaR-DRO: 0.629; Offset LR/XGB: 0.625/0.650; Unconstr. 11-param.: 8,383; 34-col. LR: 37,183 | 2.5–11.3× lower (5.7× at matched ); LOCO NB > 0 in 27/33 sites |
| 4. Progressive LDCT Integration | Clinical only → + LDCT nodule → + Visual emphysema | Held-out (): 0.685 → 0.700 → 0.708 (0.710 sign-only); LOCO: 0.700 → 0.726 → 0.729 | Held-out full LDCT: LR 0.701 (), XGB 0.696, PLCOm2012 0.696; LOCO full LDCT: PLCOm2012 0.704* | +0.023 held-out CT gain (; nodule step +0.015, ; full cohort +0.009, ) |
| 5. Smoking Counterfactuals | Quit smoking (18,164), −10/−20 pack-yr (27,818/17,798) | 0 / 33 (0.0%) hospital fits with any risk inversion (quit RR = 0.65) | Sign-only: 0/33 (0.0%); Unconstr. LR/Cox: 3–5/33; XGB: 16–27/33; RSF: 31–32/33; DeepSurv: 33/33 | Exact smoking monotonicity across 100% of single-site fits† |
| 6. Missing-Data Uncertainty | 11 MCAR/MAR cells (0%–60%, , Stage I lung graph) | rel. MAE 0.64–0.79% (); 1.21 ms/pt; gated slope 0.906–1.071 | MC-4000: 1.55–2.31% (93.7 ms, 77.5× slower); Plug-in slope: 0.512–0.607; Pop.-mean: 0.519–0.650 | ECE reduced 3–4× (0.0029–0.0063); full-CT risk inside the no-CT PI under unattenuated projection: 99.98% |
Table A21. NLST lung cancer: transfer across hospitals, added CT findings, simulated smoking reductions, and missing data.
Table A16. Uncertainty propagation with missing data in NLST cardiovascular mortality (Stage I graph; five-fold out-of-fold, ).
| Dimension | Metric / Setup | CE-Agent | Baselines | Summary |
| 1. Interval Accuracy | Rel. MAE vs. -draw MC () | 2.30% [2.05–2.57%] Cond. imp. () | MC-4000: 3.06% [2.74–3.33%]; MC-400: 9.07%; single Gaussian: 8.03%; : 64.3–79.3% | 11/11 cells; lung graph: 0.64–0.79% |
| 2. Predictive SD | Rel. MAE vs. -draw MC | 2.00% [1.83–2.14%] Cond. imp. | MC-4000: 5.23%; MC-400: 15.06%; : 71.4–85.4% (76.5% pooled) | 11/11 cells (2.6× lower vs. MC-4000) |
| 3. Single-Core Latency | Batch / single-record ms per patient | 0.62 ms quad. (1.23 ms end-to-end) / 1.80 ms single | MC-400: 20.8 ms; MC-4000: 203.8 ms; MC-: 5.0 s | 329× quad. speedup vs. MC-4000 |
| 4. Counterfactual Monotonicity | Spurious risk rises when setting CAC/emph = 0 | 0 / 3,990 (0.0%) violations | MC-400: 197 (CAC) & 449 (emphysema) rises | 0/3,990 (0.0%) sign violations |
| 5. Reliability Calibration Slope | Slope across 11 MCAR/MAR cells (0%–60%) | 0.959–1.047 (reliability score); ECE 0.0132 → 0.0053 | Untuned gate: 0.76–0.86; Pop.-mean: 0.78–0.89; Plug-in: 0.71–0.78 | Lowest log-loss in 11/11 cells (95% CI < 0) |
| 6. 40% Missingness Retention | Smaller AUC loss (0% → 40%) vs. baselines | Higher retention vs. all 6 trained baselines at 40% MCAR/MAR | Retention gain at MCAR 40%: Cox +0.008, XGB-Cox +0.015, XGB +0.018, RSF +0.022, Tree +0.026, LR +0.115 | 6/6 trained baselines (95% CI above zero) |
| 7. Overconfident False Negatives | Events predicted below half the training event rate, per 1,000 (O/E) in top-20% | 2.27 / 1,000 (O/E = 0.78 pooled; 0.00–0.96) | RSF 0.8, XGB-Cox 4.7, Cox 5.1, XGB 9.1, Tree 19.4, LR 22.3/1,000 | 2.1–9.8× lower vs. five of six baselines |
| 8. Uncertainty-Guided Triage | Retained AUC across 0%–30% review | +0.011 (20%) to +0.013 (30% ) | Total width ( with ): −0.057 (20%) to −0.069 (30%) | +0.068 to +0.082 vs. triage |
Table A16.
Note: Bold values highlight primary CE-Agent metrics. Abbreviations: MCAR, missing completely at random; MAR, missing at random; Rel. MAE, relative mean absolute error; MC, Monte Carlo; , Pearson correlation; Cond. imp., out-of-fold conditional Gaussian imputation; Pop.-mean, population-mean fill; Tree, decision tree; quad., Gauss–Hermite quadrature (1.80 ms single-record); emph, emphysema; ECE, expected calibration error; LR, logistic regression; XGB, XGBoost; XGB-Cox, Cox-boosted trees; RSF, random survival forest; O/E, observed-to-expected ratio; , Spearman rank correlation. Row 6 reports difference-in-degradation () base − () CE, with 95% CI across all 6 baselines under 40% MCAR and MAR (raw 40% MCAR AUC = 0.681, +0.005 to +0.109 over baselines).
Note: Bold, primary CE-Agent metrics. MCAR, missing completely at random; MAR, missing at random; Rel. MAE, relative mean absolute error; MC, Monte Carlo; , Pearson correlation; SD, standard deviation; quad., Gauss–Hermite quadrature (2.45 ms single-record end-to-end); Untuned, reliability gate before tuning; Pop.-mean, population-mean fill; on par or better, 95% CI of the loss difference includes zero or favors CE-Agent; main baselines, Cox PH, RSF, G BSA, XGBoost, LR, and DeepSurv (Fig. 6g: RSF, Cox PH, XGB-Cox, XGBoost, decision tree, LR; all eight models favor CE-Agent by point estimate); Ext., all 40 model-and-imputer configurations; USPSTF, lung-screening-eligible subcohort [3]; , missing-predictor log-odds uncertainty; , Spearman rank correlation. Row 9 ranks participants by the 90% predictive-interval width under a fixed CT budget, a rule distinct from the CT-uncertainty-interval trigger of two-stage triage (Table A19).
Penalty scaling, robust comparators, and CT triage
Tables A18–A19 and Figure A4 show how the soft-bound penalty scales with sample size, compare CE-Agent with robust-learning and model-updating methods [42, 43, 44] and with Gaussian priors of matched width, and report simulated risk-factor reductions [5, 39, 46, 47] and two-stage CT and statin triage [17, 48, 49].
Note: Bold values highlight primary CE-Agent metrics. Abbreviations: , best-performing penalty weight (lowest out-of-sample log loss; Figure A4a,Figure A4e plot the weight with the highest AUC); , penalty weight set from published interval widths; Oracle gap, out-of-sample log-loss excess over ; CV, cross-validation; cols, columns; Unreg., unregularized; , development-calibrated prior precision scale (31.6; closed-form in Figure A4e) under free-column stabilization; , Anchor Regression invariance weight; DRO, distributionally robust optimization (Row 7 reports the best single fixed-hyperparameter median across the 33 NLST hospitals, 0.6460, whereas Figure A4b plots per-hospital hyperparameter oracles, 0.651 and 0.655); CVaR, conditional value-at-risk; med., median across 33 single-hospital NLST OOD folds (4 race and 4 scanner subgroups across cardiovascular mortality and hard ASCVD); LR, logistic regression; XGB, XGBoost; Prevalence-only envelope, prevalence-recalibrated envelope without slope refitting; Offset LR / XGB, Steyerberg prior-offset logistic revision [44] and base-margin boosted-tree updating (prior-offset GBDT in Figure A4b,Figure A4f); MAP, maximum a posteriori; KKT, Karush–Kuhn–Tucker; TAC, thoracic aortic calcification.
Note: Bold highlights primary CE-Agent metrics. pack-yr, pack-years; rises, predicted risk increases under a beneficial risk-factor reduction; rev., cohort-level direction reversals; LOCO, leave-one-center-out; CTT, Cholesterol Treatment Trialists’ Collaboration; CHD, coronary heart disease; RCT, randomized controlled trial; RR, risk ratio; SBP, systolic blood pressure; RSF, random survival forest; RF, random forest; XGB-Cox, Cox-boosted trees; , decision risk threshold (5.0% fatal-CVD risk [48, 49] in NLST; 7.5% hard ASCVD risk [17] in MESA); /1k, net benefit per 1,000 patients; Intermediate PCE CAC, CAC referral at intermediate risk by recalibrated PCE; Two-stage selective CT, CAC read only when the 90% CT-uncertainty interval crosses .
Rows 2–3: CTT [39] reports CHD mortality RR = 0.80 (99% CI 0.74–0.87) and vascular mortality RR = 0.86 (95% CI 0.82–0.90) per 1 mmol/L LDL reduction, and Ettehad et al. [47] report major CVD RR = 0.80 (0.77–0.83) and all-cause mortality RR = 0.87 (0.84–0.91) per 10 mmHg SBP reduction. †Rows 4–5 use the Stage I propagation graph (published weights). In NLST, selective CT with Stage II hard-bound weights gives +2.34/1k at 6.7% CT reads, +0.96/1k above intermediate PCE CAC referral (; +2.42/1k at 0% CT); in MESA, CE-Agent gives +15.4/1k and PCE alone +14.0/1k at 0% CT scans.
Lung cancer and joint cardiopulmonary triage
Tables A20–A22 report held-out test (; LDCT arm ) and 33-hospital out-of-center (; LDCT arm ) benchmarks for 6-year competing-risk lung cancer incidence and 10-year lung cancer mortality, complementing Figure 8. Evaluations span single-hospital generalization (), PLCOm2012-free sign-only controls, progressive LDCT integration, smoking counterfactuals, missing-data uncertainty propagation, and joint cardiopulmonary triage.
Table A20 NLST lung cancer: CE-Agent, PLCOm2012, and ML baselines in the test cohort () and in leave-one-center-out validation in the LDCT arm (). Six-year competing-risk AUC (575 held-out and 1,350 training 6-year incident lung cancers, with 618 and 1,440 in total; LDCT arm 317 and 694 6-year cases, with 338 and 732 total), held-out calibration slope and E/O, and 10-year competing-risk lung cancer mortality AUC (535 held-out and 1,242 training 10-year deaths; 747 and 1,663 total).
Note: Bold marks the best discrimination or calibration in each column (|slope − 1| and |E/O − 1| closest to zero). LOCO, 33-hospital leave-one-center-out development validation ( LDCT); CIF, cumulative incidence function; LR, logistic regression; RSF, random survival forest; GBSA, gradient-boosted survival analysis; recal., prevalence-recalibrated. *Paired development (LOCO LDCT; not preregistered) difference from recalibrated PLCOm2012: +0.023 (95% CI +0.010 to +0.036) for CE-Agent and +0.024 (+0.011 to +0.038) for sign-only CE-Agent; on the held-out test, ΔAUC = +0.004 (95% CI −0.008 to +0.017; ; preregistered noninferiority bound −0.20). †Uses only the direction of each risk factor, with no PLCOm2012 regression coefficients; the preregistered comparison of compiled with sign-only bounds gave ΔAUC = −0.0017 (95% CI −0.0040 to +0.0005), within the ±0.01 equivalence margin. ‡Held-out comparisons of CE-Agent with 11-predictor baselines: RSF, ; GBSA, . Against baselines given an expanded 19-node encoding that adds visual emphysema and missingness indicators, the 11-predictor CE-Agent stays calibrated (slope 0.99, E/O = 0.99), whereas pooled Cox CIF (AUC = 0.724; slope 2.59) and LR (0.715; E/O = 1.10) gain discrimination with poorer calibration. §The same 11-predictor envelope fitted to 10-yr lung cancer mortality (E/O = 0.98 vs. 0.93 for published PLCOm2012; noninferior to recalibrated PLCOm2012: ΔAUC = −0.009, 95% CI −0.019 to +0.001, within the preregistered −0.02 margin; two-sided ). Refitting on the mortality label raised AUC by 0.011 () over the incidence weights shifted to the 10-year horizon (0.702).
Note: Bold highlights primary CE-Agent metrics. OOD, out-of-distribution; med., median; DRO, distributionally robust optimization; CVaR, conditional value-at-risk; Anchor, Anchor Regression [42]; Offset LR/XGB, prior-offset logistic updating [44] and base-margin XGBoost; NB, net benefit; pt, patient; Pop.-mean, population-mean fill; ECE, expected calibration error; PI, 90% predictive interval. Rows 1–3, 5, and 6 and leave-one-center-out (LOCO) values are post hoc training-cohort analyses; the held-out CT gain in row 4 is exploratory. Row 1: per-site values are medians across the 33 single-hospital fits (median ) of mean AUC at the other 32 hospitals; pooled external values (also rows 2–3) score those hospitals together. CE-Agent meets Riley’s in the two largest centers (0/33 unconstrained). Pooling two or more centers raises CE-Agent to 0.702–0.718; published PLCOm2012, which needs no local fitting, leads single-site fits (per-site median 0.696, 27/33 centers). *Lung models use radiologist reads of the baseline scan (nodules; visual emphysema in the LDCT-arm extension) and are variants of Sybil [80] (Appendix C). †The inverse BMI–lung cancer association reported for PLCOm2012 [6] is a pre-diagnostic association, not a weight-loss intervention target.
Note: Combined net benefit per 1,000 participants at the preregistered 1.5% (sensitivity 2.0%) 6-year competing-risk lung cancer incidence and 5.0% 10-year fatal-CVD incidence cutoffs [48, 49, 50, 68]; Dual CE-Agent, the lung cancer and cardiovascular CE-Agent models applied together; Clinic-only, lung cancer CE-Agent without CT inputs paired with recalibrated PCE; pub., published; recal., prevalence-recalibrated. CT / Flag, fraction with CT features read / fraction flagged for lung-cancer follow-up in the cohort at 1.5%/5.0% (99.4% meet USPSTF 2021 eligibility by NLST enrollment design). Quadrant cells give uncensored 6-year lung cancer () and 10-year cardiovascular death () counts with inverse-probability-of-censoring weighted competing-risk shares. Two-sided paired bootstrap versus Dual CE-Agent in the same cohort; ; bootstrap values are exploratory. In the LDCT arm, where imaging input is read from screening CT, Dual CE-Agent itself exceeds USPSTF + PCE (recal.) by +2.6/1k (). CT features are read when input is CT-unavailable; interval-censored cutoffs are 1.5%/5.0% vs. equal-budget random CT (95% CI −1.26 to 0.00; within
Mathematical details and proofs
Domain graph, binary biomarkers, and composite indices
In structural causal screening (Methods), the candidate domain graph is a directed acyclic graph with node set and directed edge set , in which each edge is marked as an established dependency (supported by at least one cited reference) or a plausible, context-specific one. Each node carries a timing index distinguishing baseline from post-baseline variables and a clinical role (decision nodes are clinical actions, such as starting a drug, and are excluded from baseline predictors). Each derived linear composite is formed from its standardized baseline parents with fixed weights as . Adding for to the baseline linear predictor shifts the parent weights to :
For missing binary intermediate biomarkers with training prevalence and standard deviation , sign-constrained logistic regression on standardized parent moments yields conditional probability , standardized conditional mean , and delta-method variance , complementing continuous linear projection in equation (1). For latent mediators inferred linearly from standardized upstream parents with residual variance , the downstream edge into the outcome carries reliability gate .
Proposition 1 (Shared-parent composite rank deficiency) *Let the standardized baseline feature matrix over have full column rank , and let . If , then , and for any diagonal weight matrix the unregularized Fisher matrix of is singular. Moreover, for all-ones , equals on every participant under equation (6). Conversely, if and only if .
Proof. When , lies in the column space of the parent columns, giving . Defining , where has entries for and otherwise, gives . Hence for any diagonal weight matrix , proving singularity of . Substituting into gives the result by equation (6). Conversely, writing shows , and hence , if and only if the sum outside lies outside . ∞
Unit conversion and compatibility screening
In literature prior compilation (Methods), the pooled NLST cardiovascular pipeline accepts published effect sizes via verbatim regular-expression verification inside archived PubMed abstracts [55, 56, 58], while lung-cancer and point-anchored specifications additionally compile verified intervals and point priors from archived source tables [6, 8, 54, 57, 59, 60, 61, 62, 63]. When predictor undergoes a coding transformation prior to standardization , the general conversion factor in equation (2) evaluated at baseline reference level is , which simplifies to under identity coding ; is first oriented to the direction of the cohort coding. After rescaling, the four-criterion compatibility check assigns predictor to a directional sign constraint in with rather than an interval bound in whenever any of four conditions holds, one for each check in Figure 1a (study population, covariate adjustment, unit-scale alignment, and anatomical match). First, population mismatch flags priors derived in disease-specific cohorts that differ from a screening population, such as studies of the pulmonary-to-aortic diameter ratio and exacerbations in patients with COPD [11]. Second, over-adjustment flags estimates conditioned on downstream mediators or medication-confounded biomarkers. Third, unit-scale mismatch flags categorical contrasts—such as Agatston [81] CAC tiers [8] or ordinal visual calcification grades [9, 65]—applied to continuous or binary cohort codings. Fourth, anatomical mismatch flags distinct vascular beds or segmentation definitions. For example, NLST’s proximal main pulmonary artery-to-sinotubular junction diameter ratio differs anatomically from published pulmonary-to-ascending-aorta ratios and is flagged, whereas MESA’s log-transformed TAC per SD matches the scaling Bos et al. used for aortic arch calcification [54] and enters .
Soft bounds and zeroing of reversed weights
For finite penalty weight in equation (3) with , the soft-bound estimator minimizes over :
subject to for all , with in primary fits and the sample-size-adaptive per-SD unit-normal prior (weighted by each gated design column’s empirical variance) on unbounded predictors in small-sample MESA scaling experiments. The continuously differentiable squared-hinge violation vanishes throughout and penalizes excursions quadratically outside the interval. Because a published 95% interval corresponds to Gaussian prior standard deviation and quadratic negative-log-prior coefficient on summed negative log-likelihood , the per-patient penalty weight is set from the published interval widths and scales inversely with cohort size as , where equals 1,576.5 on NLST cardiovascular mortality and 224.5 on NLST lung cancer. Taking (with ) enforces for all and for , recovering the hard-bound limit over . For each sign-only predictor with dual multiplier , first-order optimality over at satisfies the KKT conditions:
Proposition 2 (Zeroing of sign-reversed weights) Because and are convex and continuously differentiable on and both and are non-empty closed convex polyhedra, any local minimizer of equations (3) and (B2) for , including the primary hard-bound limit (, ), is globally optimal and zeroes sign-reversed coordinates : whenever confounding or competing mortality reverses the empirical gradient against the published causal sign so that , complementary slackness forces (whereas an interval-bounded predictor with and clamps at its literature boundary nearer zero as ).
Proof. Convexity of , , and together with linear constraints makes the KKT conditions necessary and sufficient for global optimality. On any sign-only coordinate , stationarity gives (reducing directly to equation (B3) when ). Because and , multiplying by yields . Whenever , positivity and force . □
Quadrature and variance decomposition
In analytic uncertainty propagation (Methods), each observed-predictor node carries moments when observed, under population-mean fill (with imputed binary predictors carrying standardized Bernoulli variance clipped at , where is the imputed probability that predictor equals 1), or under out-of-fold conditional Gaussian imputation. Each latent mediator carries standardized residual variance with . Active directed edges from node to direct child comprise the outcome and any mediator children that are latent or missing for record , so observed intermediate biomarkers block indirect upstream paths without double-counting. Each active edge carries the reliability-gated weight , where collects the gate parameters and the edge-scale multipliers. The base edge weight is on direct outcome edges, 0.5 on edges through latent mediators, and, on edges with no published or fitted coefficient, 0.6 on direct outcome edges (0.5 times the established weight for plausible dependencies). The factor collects unit-mean log-normal multipliers () with on the global outcome scale and the latent-edge scale and on the plausible-to-established weight ratio. Out-of-fold fitted gate parameters vary around the fitted center (initialized at ) with normal standard deviations and clipped at zero. Reverse-topological recursion from via yields and .
At each of the Gauss–Hermite nodes over the three gate parameters (weights ), the nodes over the plausible-to-established weight ratio and the latent-edge scale are moment-matched into a single Gaussian component on the logit scale. Incorporating training-prevalence intercept variance , marginal risk integrates over a 16-node Gauss–Hermite rule (with 16 inner Gauss–Hermite nodes for ), while the log-odds mixture cumulative distribution function integrates over a 64-point uniform normal-density grid on to resolve steep tail transitions:
where denotes the standard normal cumulative distribution function. Safeguarded Newton–bisection (60 maximum iterations) on logit solves and to obtain monotone risk quantiles .
By the law of total variance, the marginal log-odds variance decomposes into baseline, input, and parameter channels:
With standard normal 95th percentile , a first-order Taylor expansion of around the predictive mean yields in equation (5). At the central parameter vector , , decomposing total input variance into the complete-cohort baseline from latent nodes and graph predictors that the cohort did not record (0.3284 in NLST cardiovascular mortality and 0.1884 in MESA when all cohort-measured predictors are observed) and cohort-measured missing-predictor variance for cohort-measured predictors . Per patient on a graph , assembling the node moments visits nodes and links once, and vectorized reverse-topological sensitivity accumulation performs operations per link and node in time. Subsequent scalar logit-mixture moment integration ( evaluations) and safeguarded Newton–bisection root-finding ( iterations over mixture components) perform a fixed number of operations, , independent of and .
-draw Monte Carlo simulation instead takes time. Figure 7 benchmarks both across with 8.5% latent mediators, 30% binary input nodes, 25% missingness, and on a single pinned CPU core; single-patient speed-ups over 4,000-draw Monte Carlo at these sizes were 83, 140, 217, 271, 362, 371, 377, and 353-fold.
Effective degrees of freedom and minimum sample size
Around with logistic Hessian weights and active prior curvature , effective degrees of freedom compress from to (equaling the number of fitted parameters, , in the point-anchored single-hospital NLST cardiovascular specification and median in single-hospital NLST lung cancer fits). Under the binary-outcome sample-size standards of Riley et al. [23] (Methods), minimum sample size is determined from , prevalence , Cox–Snell , and maximum :
Here enforces global shrinkage , limits expected Nagelkerke optimism to , and bounds marginal outcome prevalence error within . Criterion C1 binds across cohorts, so rises with and falls with the anticipated of each specification, converted from its median held-out AUC at the pooled training prevalence [67]. In NLST cardiovascular mortality (), yields versus 6,043 unconstrained (), a 5.4-fold reduction combining a 3.0-fold gain from fewer parameters with a 1.8-fold gain from higher anticipated discrimination (12 parameters at the constrained would require 3,329). In NLST 6-year lung cancer ( on uncensored single-hospital records), yields versus 8,383 unconstrained (11 parameters) and 37,183 (34 expanded columns). In MESA, versus 22.0 yields versus 8,380 on cardiovascular mortality and versus 22.0 yields versus 3,055 on hard ASCVD.
Experimental protocols
Standardization, optimization, and decision curves
In MESA, predictors are standardized using empirical training-fold moments , whereas NLST predictors use pre-specified cohort reference centers and scales, and each training entry is multiplied by its reliability gate (equation (4)), which downweights missing values. Constrained estimation in equation (3) minimizes the summed objective via bound-constrained limited-memory Broyden–Fletcher–Goldfarb–Shanno (L-BFGS-B) optimization [82] with projected gradient tolerance , function tolerance 0, and in primary fits (with active-set projected Newton polishing to whenever the summed projected gradient exceeds , certified by summed gradient or half-squared Newton decrement ). Under soft bounds, a sample-size-adaptive selector applies below the Riley (at anticipated ) or 10-events-per-free-parameter threshold and five-fold inner cross-validation across 13 grid points (, –, and ) once both thresholds are met. For 6-year and 10-year competing-risk evaluations [40], inverse-probability-of-censoring weights are re-estimated in each bootstrap replicate and subgroup from the Kaplan–Meier censoring distribution to match nonparametric Aalen–Johansen cumulative incidence, while NLST models are aligned to each horizon by an additive logit shift that matches the training-cohort Aalen–Johansen incidence; the same alignment is applied to the prevalence-recalibrated clinical calculators. Reliability-gate centers minimize out-of-fold logistic log-loss across pooled missingness regimes via coarse grid search (27 points in NLST cardiovascular risk and 72 points including intercept shift in MESA) followed by bounded Nelder–Mead optimization [83]. In decision curve analysis [45], net benefit from true-positive and false-positive counts and and excess net benefit over treat-all and treat-none at threshold are
Across 33 NLST hospitals, excess net benefit from training center to external center is averaged across all 32 external centers; the thresholds at which this average is positive for every training center, , include 5%–10% for CE-Agent (Table A8).
Missing data, MESA transfer, and cardiovascular triage
Across NLST and MESA Exam 1, features were masked completely at random (MCAR) at rates and under BMI-dependent missing-at-random (MAR) mechanisms with probability for BMI percentile (mean rate up to 50%, and 58.3% at the nominal 60%). Masked entries were handled by population-mean fill (moments (0, 1, 1)), out-of-fold ridge-regularized conditional Gaussian imputation, or single conditional draws paired with conditional error variance. In MESA Exam 1, the 22-predictor outcome model comprises 13 upstream clinical predictors, 8 directly measured intermediate clinical, laboratory, respiratory, and CT biomarkers, and 1 upstream-derived inflammation score (leaving 18 non-CT features when the 4 direct CT columns are omitted). The 8 directly measured intermediate biomarkers that enter the outcome model are systolic blood pressure (SBP), diabetes/glycemic status, Chronic Kidney Disease Epidemiology Collaboration estimated glomerular filtration rate (CKD-EPI eGFR), respiratory disease history (emphysema or asthma), CAC, TAC, cardiac chamber area, and ascending aortic diameter. Exam 1 measured and CT liver attenuation are modeled alongside these 8 biomarkers among the 10 supervised intermediate DAG biomarkers (Table A12), while an expanded 58-predictor (63-column) Exam 1 set adds 36 demographic, socioeconomic, electrocardiographic, laboratory, and CT covariates.
In MESA Exam 1, cross-environment experiments evaluated single-group-to-rest and leave-one-group-out transfers across all four racial/ethnic cohorts and four electron-beam and multi-detector CT scanner platforms. Outpatient missing-CT inference masked all five CT phenotypes (CAC, TAC, cardiac chamber area, ascending aortic diameter, and liver attenuation) at test time. Seven survival and ML baseline families were trained across both the 22- and 58-predictor sets (14 model-by-feature configurations) using identical standardized features and cross-validated hyperparameter grids. Domain-generalization and clinical updating comparators comprised Anchor Regression [42] across invariance weights , -divergence DRO, which needs no hospital or subgroup labels, across divergence radii and CVaR-DRO across tail fractions [43], prior-offset logistic regression [44] and base-margin XGBoost, which take CE-Agent’s predicted log-odds as the offset and the starting margin, and width-matched Gaussian priors with . Cardiovascular counterfactual audits shifted standardized risk factors one variable at a time—holding observed downstream laboratory and CT measurements fixed—across smoking cessation, pack-year reductions, and BMI in NLST (operating through the upstream BMI-to-latent-inflammation pathway in the point-anchored single-hospital model), and LDL, SBP, and smoking cessation in MESA. Two-stage cardiovascular CT triage routed participants with baseline clinical indications directly to preventive statin therapy (diabetes in primary-prevention NLST screening; diabetes or LDL (4.9 mmol/L) in baseline lipid-lowering-therapy-free MESA adults aged 45–75 years) and compared recalibrated PCE intermediate-risk referral against triggering on the pre-read 90% CT-uncertainty interval at in NLST [48, 49] and against 2018 ACC/AHA intermediate-risk (PCE 7.5%–20%) CAC referral rules at in MESA [17].
Lung cancer and joint triage protocols
In NLST lung cancer estimation (), structural screening retains five interval-bounded clinical predictors rescaled from published PLCOm2012 odds-ratio intervals [6] (age, BMI, COPD, family history of lung cancer, and smoking status; ), four sign-only predictors (smoking intensity/duration , baseline radiologist non-calcified nodule mm read , occupational exposure , and educational attainment ), and two free demographic predictors (sex and race/ethnicity). BMI is a direct predictor here, as in PLCOm2012, because height and weight are not entered separately, and visual emphysema enters only the LDCT-arm extension. To test independence from published PLCOm2012 numerical weights, an interval-free directional (sign-only) specification retains clinical consensus signs on eight predictors without numerical intervals while leaving BMI, sex, and race/ethnicity free; a strict three-sign variant retaining only non-PLCOm2012 literature directions on smoking, occupational exposure, and baseline nodules yields identical held-out discrimination.
Survival and ML comparators were trained on both the matched 11-predictor design and an expanded 19-node (34-column) baseline encoding, with missing nodule reads handled via reliability gating in CE-Agent and training-median imputation plus missingness indicators in ML baselines (128 held-out records missing scalar PLCOm2012 inputs used training medians, verified invariant on the complete-case subset). Primary 6-year competing-risk incidence (2,191.5 d) treats pre-diagnosis death as a competing event, whereas 10-year lung cancer mortality (3,652.5 d) treats non-lung-cancer death as competing and compares the incidence-fitted weights, shifted to the 10-year horizon by the additive logit term, against weights refitted on the mortality label (Table A20–Table A21).
Joint cardiopulmonary triage pairs the compiled 6-year lung cancer model with the previously fixed 10-year cardiovascular mortality model on the held-out cohort () at preregistered (1.5%, 5.0%) cutoffs [48, 49, 50], with sensitivity analyses at (2.0%, 5.0%) [68] on both the full held-out cohort and the baseline LDCT arm (; Table A22). Two-stage selective reading scores clinic-only (no CT) 10-predictor lung cancer risk and recalibrated PCE in Tier 1 and triggers Tier 2 reading of chest CT features (the nodule read for lung cancer; CAC and emphysema for cardiovascular risk) whenever on either endpoint, where is the Tier 1 risk and comes from the corresponding CE-Agent graph with all CT-derived predictors treated as unmeasured. A random-referral control reads the same number of scans, selected deterministically from participant identifiers. Missing-data propagation benchmarks evaluated the prevalence-anchored Stage I 11-predictor lung model using five-fold out-of-fold gate parameters refined via 12-point grid search and bounded Nelder–Mead steps (80 evaluations), which lifted calibration slope from 0.52–0.61 to 0.91–1.07. Confidence intervals and two-sided paired values across lung and joint-triage analyses used 1,000 participant-level bootstrap replicates without center-level clustering.
Automated CT segmentation and runtime
Baseline nongated chest LDCT scans in NLST were converted from Digital Imaging and Communications in Medicine (DICOM) series to Neuroimaging Informatics Technology Initiative (NIfTI) volumes with SimpleITK and segmented using TotalSegmentator [52] (v2.17.0 for cohort feature extraction; v2.18.0 for timing) to quantify log-transformed thoracic aortic calcium volume ( Hounsfield units, HU [81]), left ventricular chamber volume (mL), and proximal main-to-sinotubular-junction pulmonary artery diameter ratio (with direct ascending-aorta normalization evaluated in a subcohort, ). Emphysema was quantified within automated 3D lung masks (lungmask [69]). Only baseline () series were analyzed; series flagged for missing structures or field-of-view truncation were excluded, and repeated series reconstructed with the same kernel were averaged per participant. The primary subcohort is restricted to soft-tissue kernels because sharp kernels read about ninefold higher calcium volumes in participants with both reconstructions. Because at least 75.1% of the 9,346 participants with baseline 3D CT, including at least 16.0% of the 2,773 held-out participants, overlapped the training or development data of the 3D lung-cancer convolutional neural network (CNN) Sybil [80], Sybil was excluded from all risk models and timed only as a concurrent runtime reference. In the same participants, automatically measured emphysema (low-attenuation area fraction, LAA HU [84]) yields 0.663 held-out () and 0.707 leave-one-center-out (LOCO; ) AUC. Single-scan runtime from local DICOM hand-off to tabular output was measured across 10 baseline NLST scans following one warm-up scan with all models resident and executed concurrently on one 48-GB NVIDIA RTX A6000 GPU with a 40-core CPU and 100 GB of memory (Table C1). Because 3D lung-cancer CNN inference and lung-mask extraction complete in 3.2 s and 7.6 s, concurrent execution adds no wall-clock time beyond the 45.6 s cardiovascular segmentation window, while resident-model outputs matched standard command-line execution identically on lungmask and Sybil and within native GPU floating-point precision on TotalSegmentator ( relative volume difference and 803/804, or 99.9%, identical vascular diameters at 0.1 cm resolution).
Note: Chest CT stage times are cumulative wall-clock times from local DICOM hand-off across 10 baseline NLST scans (after 1 warm-up scan), with all models resident and run concurrently on one NVIDIA RTX A6000 (48 GB) GPU (40 CPU cores, 100 GB RAM; peak memory GB; TotalSegmentator [52] v2.18.0; cohort features extracted with v2.17.0). Sybil [80] is timed only as a concurrent 3D CNN reference; it is not a CE-Agent predictor because at least 75.1% of the participants with baseline 3D CT overlap its training or development data. Names in typewriter font are TotalSegmentator tasks. HU, Hounsfield units; LAA, low-attenuation lung area fraction below HU.
| Dimension | Metric / Setup | CE-Agent | Baselines | Summary |
| 1. Interval Accuracy | Rel. MAE vs. -draw MC () | 0.50% med. (0.52% mean) [0.43–0.64%] () | MC-4000: 2.13–2.90%; MC-400: 5.9–10.1%; : 69.3% | 11/11 cells (4.8× lower vs. MC-4000) |
| 2. Predictive SD | Rel. MAE vs. -draw MC | 1.28% med. (1.95% mean) [0.85–4.45%] | MC-4000: 3.1–13.6% (6.3% mean); MC-400: 8.8–28.0% | 11/11 cells (3.3× lower vs. MC-4000) |
| 3. Single-Core Latency | Quadrature / end-to-end ms per patient | 0.72 ms quad. / 1.13 ms end-to-end | MC-400: 18.9 ms; MC-4000: 184.8 ms | 257× quad. (164× end-to-end) vs. MC-4000 |
| 4. Counterfactual Monotonicity | Spurious risk rises across 17 non-zero edges | 0 / 2,720 (0.0%) violations | Monte Carlo sampling produces rank reversals | 0/2,720 (0.0%) sign violations |
| 5. Reliability Calibration Slope | Slope across 11 cells (CVD death / Hard ASCVD) | 0.881–1.083 (CVD) / 0.92–1.17 (ASCVD) | Untuned: 0.39–0.46; Pop.-mean: 0.45–0.60; Plug-in: 0.45–0.51 | 11/11 cells within [0.88, 1.17] |
| 6. 40% Missingness Retention | Smaller AUC loss (0% → 40%) vs. baselines | Higher retention vs. all 6 main baselines; on par or better vs. 40/40 under MAR | Gain (MCAR/MAR): RSF 0.021/0.044; Cox 0.103/0.051; LR 0.235/0.191 | Point estimate higher: 6/6 main; 30/40 MAR, 18/40 MCAR Ext.; on par or better, MCAR: 39/40 ASCVD, 27/40 CVD |
| 7. Overconfident False Negatives | Events predicted below half the training event rate, per 1,000 | 2.07–3.10 / 1,000 (CVD); 1.7–9.6 / 1,000 (ASCVD) | Untuned: 8.8–19.7 (CVD), 24.3–50.3 (ASCVD); Ext.: ≤ 24.3/72.4 | 2.5–11.7× lower at 40% MCAR |
| 8. Uncertainty-Guided Triage | Retained AUC across 0%–30% review | +0.061 (30% ); +0.083 / +0.104 (USPSTF ) | Triage by (): −0.046 (15%), −0.036 (20%), −0.004 (30%) | 0.8% flagged at 0% missingness |
| 9. Targeted CT Triggering | CT budget 0%–100% (Stage I graph) | 30% CT budget by or distance to 7.5% risk → AUC = 0.794–0.795 | No CT: 0.776; Random 30% CT: 0.777 (needs 70% for 0.794) | 70% CT scans spared (AUC = 0.000) |
Table A17. Uncertainty propagation with missing data in MESA: CE-Agent versus 40 model-and-imputer configurations ().
References
- [1]National Lung Screening Trial Research Team. Reduced lung-cancer mortality with low-dose computed tomographic screening. New England Journal of Medicine 365, 395–409 (2011).DOI
- [2]de Koning, H. J. et al. Reduced lung-cancer mortality with volume CT screening in a randomized trial. New England Journal of Medicine 382, 503–513 (2020).DOI
- [3]US Preventive Services Task Force. Screening for lung cancer: US Preventive Services Task Force recommendation statement. JAMA 325, 962–970 (2021).
- [4]Prospective Studies Collaboration. Age-specific relevance of usual blood pressure to vascular mortality: a meta-analysis of individual data for one million adults in 61 prospective studies. The Lancet 360, 1903–1913 (2002).
- [5]Mons, U. et al. Impact of smoking and smoking cessation on cardiovascular events and mortality among older adults: meta-analysis of individual participant data from prospective cohort studies of the CHANCES consortium. BMJ 350, h1551 (2015).DOI
- [6]Tammemägi, M. C. et al. Selection criteria for lung-cancer screening. New England Journal of Medicine 368, 728–736 (2013).
- [7]McWilliams, A. et al. Probability of cancer in pulmonary nodules detected on first screening CT. New England Journal of Medicine 369, 910–919 (2013).
- [8]Chiles, C. et al. Association of coronary artery calcification and mortality in the National Lung Screening Trial: a comparison of three scoring methods. Radiology 276, 82–90 (2015).DOI
- [9]Gondrie, M. J. A. et al. Cardiovascular disease: prediction with ancillary aortic findings on chest CT scans in routine practice. Radiology 257, 549–559 (2010).DOI
- [10]Jacobs, P. C. et al. Coronary artery calcium can predict all-cause mortality and cardiovascular events on low-dose CT screening for lung cancer. American Journal of Roentgenology 198, 505–511 (2012).DOI
- [11]Wells, J. M. et al. Pulmonary arterial enlargement and acute exacerbations of COPD. New England Journal of Medicine 367, 913–921 (2012).DOI
- [12]Hecht, H. S. et al. 2016 SCCT/STR guidelines for coronary artery calcium scoring of noncontrast noncardiac chest CT scans: a report of the Society of Cardiovascular Computed Tomography and Society of Thoracic Radiology. Journal of Cardiovascular Computed Tomography 11, 74–84 (2017).
- [13]Miller, R. J. H. et al. Predicting mortality from AI cardiac volumes mass and coronary calcium on chest computed tomography. Nature Communications 15, 2747 (2024).DOI
- [14]Van Calster, B., McLernon, D. J., van Smeden, M., Wynants, L. & Steyerberg, E. W. Calibration: the Achilles heel of predictive analytics. BMC Medicine 17, 230 (2019).DOI
- [15]D’Agostino, R. B., Sr. et al. General cardiovascular risk profile for use in primary care: the Framingham Heart Study. Circulation 117, 743–753 (2008).
- [16]Goff, D. C., Jr. et al. 2013 ACC/AHA guideline on the assessment of cardiovascular risk: a report of the American College of Cardiology/American Heart Association Task Force on Practice Guidelines. Circulation 129, S49–S73 (2014).
- [17]Grundy, S. M. et al. 2018 AHA/ACC/AACVPR/AAPA/ABC/ACPM/ADA/AGS/APhA/ASPC/NLA/PCNA guideline on the management of blood cholesterol: a report of the American College of Cardiology/American Heart Association Task Force on Clinical Practice Guidelines. Circulation 139, e1082–e1143 (2019).
- [18]Arnett, D. K. et al. 2019 ACC/AHA guideline on the primary prevention of cardiovascular disease: a report of the American College of Cardiology/American Heart Association Task Force on Clinical Practice Guidelines. Circulation 140, e596–e646 (2019).
- [19]Khan, S. S. et al. Development and validation of the American Heart Association’s PREVENT equations. Circulation 149, 430–449 (2024).
- [20]Niu, C. et al. Medical multimodal multitask foundation model for lung cancer screening. Nature Communications 16, 1523 (2025).DOI
- [21]Li, C. et al. Proteomic risk score for early prediction of kidney disease progression in individuals with APOL1 high-risk genotypes. Nature Medicine 32, 1701–1707 (2026).DOI
- [22]Demircan, K. et al. Data-driven prioritization of high-risk individuals for weight loss interventions. Nature Medicine 32, 2117–2127 (2026).DOI
- [23]Riley, R. D. et al. Calculating the sample size required for developing a clinical prediction model. BMJ 368, m441 (2020).DOI
- [24]Jacobs, D. R., Jr. et al. Report of the Conference on Low Blood Cholesterol: mortality associations. Circulation 86, 1046–1060 (1992).
- [25]US Preventive Services Task Force. Statin use for the primary prevention of cardiovascular disease in adults: US Preventive Services Task Force recommendation statement. JAMA 328, 746–753 (2022).
- [26]Sperrin, M. et al. Using marginal structural models to adjust for treatment drop-in when developing clinical prediction models. Statistics in Medicine 37, 4142–4154 (2018).
- [27]Pearl, J. Causality: Models, Reasoning, and Inference 2nd edn (Cambridge University Press, 2009).
- [28]Peters, J., Janzing, D. & Schölkopf, B. Elements of Causal Inference: Foundations and Learning Algorithms (MIT Press, 2017).
- [29]Kompa, B., Snoek, J. & Beam, A. L. Second opinion needed: communicating uncertainty in medical machine learning. npj Digital Medicine 4, 4 (2021).DOI
- [30]Riley, R. D. et al. Uncertainty of risk estimates from clinical prediction models: rationale, challenges, and approaches. BMJ 388, e080749 (2025).DOI
- [31]Pauker, S. G. & Kassirer, J. P. The threshold approach to clinical decision making. New England Journal of Medicine 302, 1109–1117 (1980).
- [32]Geifman, Y. & El-Yaniv, R. Selective classification for deep neural networks. Advances in Neural Information Processing Systems 30, 4878–4887 (2017).
- [33]Leibig, C., Allken, V., Ayhan, M. S., Berens, P. & Wahl, S. Leveraging uncertainty information from deep neural networks for disease detection. Scientific Reports 7, 17816 (2017).
- [34]Raghu, M. et al. Direct uncertainty prediction for medical second opinions. Proceedings of the 36th International Conference on Machine Learning (PMLR) 97, 5281–5290 (2019).arxiv.org/abs/1807.01771
- [35]Dvijotham, K. et al. Enhancing the reliability and accuracy of AI-enabled diagnosis via complementarity-driven deferral to clinicians. Nature Medicine 29, 1814–1820 (2023).DOI
- [36]Bild, D. E. et al. Multi-Ethnic Study of Atherosclerosis: objectives and design. American Journal of Epidemiology 156, 871–881 (2002).
- [37]Detrano, R. et al. Coronary calcium as a predictor of coronary events in four racial or ethnic groups. New England Journal of Medicine 358, 1336–1345 (2008).DOI
- [38]Hoerl, A. E. & Kennard, R. W. Ridge regression: biased estimation for nonorthogonal problems. Technometrics 12, 55–67 (1970).
- [39]Cholesterol Treatment Trialists’ (CTT) Collaboration. Efficacy and safety of more intensive lowering of LDL cholesterol: a meta-analysis of data from 170 000 participants in 26 randomised trials. The Lancet 376, 1670–1681 (2010).
- [40]Blanche, P., Dartigues, J.-F. & Jacqmin-Gadda, H. Estimating and comparing time-dependent areas under receiver operating characteristic curves for censored event times with competing risks. Statistics in Medicine 32, 5381–5397 (2013).DOI
- [41]Uno, H., Cai, T., Tian, L. & Wei, L.-J. Evaluating prediction rules for t-year survivors with censored regression models. Journal of the American Statistical Association 102, 527–537 (2007).
- [42]Rothenhäusler, D., Meinshausen, N., Bühlmann, P. & Peters, J. Anchor regression: heterogeneous data meet causality. Journal of the Royal Statistical Society: Series B (Statistical Methodology) 83, 215–246 (2021).
- [43]Duchi, J. C. & Namkoong, H. Learning models with uniform performance via distributionally robust optimization. The Annals of Statistics 49, 1378–1406 (2021).
- [44]Steyerberg, E. W., Borsboom, G. J. J. M., van Houwelingen, H. C., Eijkemans, M. J. C. & Habbema, J. D. F. Validation and updating of predictive logistic regression models: a study on sample size and shrinkage. Statistics in Medicine 23, 2567–2586 (2004).
- [45]Vickers, A. J. & Elkin, E. B. Decision curve analysis: a novel method for evaluating prediction models. Medical Decision Making 26, 565–574 (2006).
- [46]Prospective Studies Collaboration. Body-mass index and cause-specific mortality in 900 000 adults: collaborative analyses of 57 prospective studies. The Lancet 373, 1083–1096 (2009).DOI
- [47]Ettehad, D. et al. Blood pressure lowering for prevention of cardiovascular disease and death: a systematic review and meta-analysis. The Lancet 387, 957–967 (2016).
- [48]Conroy, R. M. et al. Estimation of ten-year risk of fatal cardiovascular disease in Europe: the SCORE project. European Heart Journal 24, 987–1003 (2003).
- [49]De Backer, G. et al. European guidelines on cardiovascular disease prevention in clinical practice: Third Joint Task Force of European and Other Societies on Cardiovascular Disease Prevention in Clinical Practice. European Heart Journal 24, 1601–1610 (2003).
- [50]Tammemägi, M. C. et al. Evaluation of the lung cancer risks at which to screen ever- and never-smokers: screening rules applied to the PLCO and NLST cohorts. PLOS Medicine 11, e1001764 (2014).DOI
- [51]Tibshirani, R. Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society: Series B (Methodological) 58, 267–288 (1996).
- [52]Wasserthal, J. et al. TotalSegmentator: robust segmentation of 104 anatomic structures in CT images. Radiology: Artificial Intelligence 5, e230024 (2023).DOI
- [53]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 Digital Health 2, e192–e200 (2020).DOI
- [54]Bos, D. et al. Comparison of atherosclerotic calcification in major vessel beds on the risk of all-cause and cause-specific mortality: the Rotterdam Study. Circulation: Cardiovascular Imaging 8, e003843 (2015).DOI
- [55]Aune, D., Huang, W., Nie, J. & Wang, Y. Hypertension and the risk of all-cause and cause-specific mortality: an outcome-wide association study of 67 causes of death in the National Health Interview Survey. BioMed Research International 2021, 9376134 (2021).DOI
- [56]Ronksley, P. E., Brien, S. E., Turner, B. J., Mukamal, K. J. & Ghali, W. A. Association of alcohol consumption with selected cardiovascular disease outcomes: a systematic review and meta-analysis. BMJ 342, d671 (2011).
- [57]Khaing, W., Vallibhakara, S. A., Attia, J., McEvoy, M. & Thakkinstian, A. Effects of education and income on cardiovascular outcomes: a systematic review and meta-analysis. European Journal of Preventive Cardiology 24, 1032–1042 (2017).
- [58]González Gutiérrez, J. et al. Emphysema at baseline low-dose CT lung cancer screening predicts death from chronic obstructive pulmonary disease and cardiovascular disease up to 25 years later. Radiology 316, e250949 (2025).DOI
- [59]Heo, S. et al. Effect modification by sex for associations of fine particulate matter (PM₂.₅) with cardiovascular mortality, hospitalization, and emergency room visits: systematic review and meta-analysis. Environmental Research Letters 17, 053006 (2022).DOI
- [60]Mattila, T. et al. Mortality of asthma, COPD, and asthma-COPD overlap during an 18-year follow up. Respiratory Medicine 207, 107112 (2023).
- [61]Li, Y., Li, F., Wang, G., Zeng, Q. & Xie, P. Additive impact of chronic obstructive pulmonary disease (COPD) and cardiovascular disease (CVD) on all-cause and disease-specific mortality: a longitudinal nationwide population-based study. BMC Pulmonary Medicine 25, 275 (2025).DOI
- [62]Navarese, E. P. et al. Causal effects of inflammation on long-term mortality: a Mendelian randomization study. Aging 18, 45–59 (2026).
- [63]Tasdighi, E. et al. Association between cigarette smoking status, intensity, and cessation duration with long-term incidence of nine cardiovascular and mortality outcomes: the Cross-Cohort Collaboration (CCC). PLOS Medicine 22, e1004561 (2025).DOI
- [64]Zhang, J. & Yu, K. F. What’s the relative risk? A method of correcting the odds ratio in cohort studies of common outcomes. JAMA 280, 1690–1691 (1998).
- [65]Mascalchi, M. et al. Moderate-severe coronary calcification predicts long-term cardiovascular death in CT lung cancer screening: the ITALUNG trial. European Journal of Radiology 145, 110040 (2021).DOI
- [66]Aalen, O. O. & Johansen, S. An empirical transition matrix for non-homogeneous Markov chains based on censored observations. Scandinavian Journal of Statistics 5, 141–150 (1978).
- [67]Riley, R. D., Van Calster, B. & Collins, G. S. A note on estimating the Cox–Snell R² from a reported C statistic (AUROC) to inform sample size calculations for developing a prediction model with a binary outcome. Statistics in Medicine 40, 859–864 (2021).DOI
- [68]Tammemägi, M. C. et al. Selection of individuals for lung cancer screening based on risk prediction model performance and economic factors – the Ontario experience. Lung Cancer 156, 31–40 (2021).DOI
- [69]Hofmanninger, J. et al. Automatic lung segmentation in routine imaging is primarily a data diversity problem, not a methodology problem. European Radiology Experimental 4, 50 (2020).DOI
- [70]Levey, A. S. et al. A new equation to estimate glomerular filtration rate. Annals of Internal Medicine 150, 604–612 (2009).
- [71]Cox, D. R. Regression models and life-tables. Journal of the Royal Statistical Society: Series B (Methodological) 34, 187–202 (1972).
- [72]Ishwaran, H., Kogalur, U. B., Blackstone, E. H. & Lauer, M. S. Random survival forests. The Annals of Applied Statistics 2, 841–860 (2008).
- [73]Friedman, J. H. Greedy function approximation: a gradient boosting machine. The Annals of Statistics 29, 1189–1232 (2001).DOI
- [74]Chen, T. & Guestrin, C. XGBoost: a scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 785–794 (2016).DOI
- [75]Katzman, J. L. et al. DeepSurv: personalized treatment recommender system using a Cox proportional hazards deep neural network. BMC Medical Research Methodology 18, 24 (2018).DOI
- [76]Breiman, L. Random forests. Machine Learning 45, 5–32 (2001).
- [77]Harrell, F. E., Jr., Lee, K. L. & Mark, D. B. Multivariable prognostic models: issues in developing models, evaluating assumptions and adequacy, and measuring and reducing errors. Statistics in Medicine 15, 361–387 (1996).
- [78]Brier, G. W. Verification of forecasts expressed in terms of probability. Monthly Weather Review 78, 1–3 (1950).
- [79]Kaplan, E. L. & Meier, P. Nonparametric estimation from incomplete observations. Journal of the American Statistical Association 53, 457–481 (1958).
- [80]Mikhael, P. G. et al. Sybil: a validated deep learning model to predict future lung cancer risk from a single low-dose chest computed tomography. Journal of Clinical Oncology 41, 2191–2200 (2023).DOI
- [81]Agatston, A. S. et al. Quantification of coronary artery calcium using ultrafast computed tomography. Journal of the American College of Cardiology 15, 827–832 (1990).
- [82]Byrd, R. H., Lu, P., Nocedal, J. & Zhu, C. A limited memory algorithm for bound constrained optimization. SIAM Journal on Scientific Computing 16, 1190–1208 (1995).
- [83]Nelder, J. A. & Mead, R. A simplex method for function minimization. The Computer Journal 7, 308–313 (1965).
- [84]Gevenois, P. A., de Maertelaer, V., De Vuyst, P., Zanen, J. & Yernault, J. C. Comparison of computed density and macroscopic morphometry in pulmonary emphysema. American Journal of Respiratory and Critical Care Medicine 152, 653–657 (1995).











