betaFeedback

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 (L2L_2) 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 λ\lambda on departures from published bounds (λ=0\lambda= 0, unconstrained; λ=∞\lambda= \infty, 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 (CC-index; 95% confidence interval, CI) and 10-year E/OcrE/O_{\mathrm{cr}} 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 (tt-AUC) for cardiovascular mortality in NLST under competing mortality [40] (c; shading, gain over PCE; inset Δ\DeltaInt. tt-AUC spans 1–13 yr) and in MESA [41] (f; gain over unconstrained). E/OcrE/O_{\mathrm{cr}}, competing-risk expected-to-observed ratio; AUC10_{10}, 10-year AUC; PCE, ACC/AHA Pooled Cohort Equations; PREVENT, AHA risk equations; Int., integrated.

Model / SpecificationAUC10_{10} (95% CI)C10C_{10} / CfullC_{\mathrm{full}}Int. tt-AUCSlope (E/O)
CE-Agent (hard, ln⁡(CAC+1)\ln(\mathrm{CAC}+1)){sup}†`0.8068 (0.758–0.855)0.8157 / 0.80460.84230.777 (1.07)
CE-Agent (soft, ln⁡(CAC+1)\ln(\mathrm{CAC}+1))0.8051 (0.752–0.854)0.8145 / 0.80740.84110.833 (1.07)
CE-Agent (hard, binary CAC)0.8107 (0.763–0.858)0.8194 / 0.80350.8418{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 / 0.80050.84160.466 (1.08)
Unconstrained 22-Predictor Model0.7886 (0.729–0.842)0.7979 / 0.79520.82460.768 (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 / 0.80440.84281.742 (—){sup}‡`
ACC/AHA ASCVD-PCE0.8131 (0.767–0.857)0.8209 / 0.79450.83061.096 (—){sup}‡`
Framingham General CVD0.8149 (0.764–0.864)0.8204 / 0.77200.83061.149 (—){sup}‡`

Table A4. MESA test-cohort cardiovascular mortality: CE-Agent specifications versus clinical calculators (n=1,974n = 1{,}974). Bold values denote overall column-best values across non-ablation models (∣slope−1∣|\mathrm{slope} - 1| and ∣E/O−1∣|\mathrm{E/O} - 1| closest to zero); survival-increment-weighted integrated tt-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/Ocr_{\mathrm{cr}} in Table A2).

In the held-out NLST test cohort (n=15,931n = 15{,}931), CE-Agent trained across all 33 hospitals achieved a concordance index (CC-index) of 0.731 and accurate 10-year calibration (expected-to-observed event ratio E/Ocr=1.02E/O_{\mathrm{cr}} = 1.02, 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 / SpecificationAUCcr_{\mathrm{cr}} (95% CI)CC-index (95% CI)BrierE/Ocr_{\mathrm{cr}}
CE-Agent (compiled hard bounds, primary)0.723 (0.700–0.744)0.731 (0.713–0.749)0.04331.02
CE-Agent (compiled soft bounds, CV λ\lambda)0.723 (0.700–0.744)0.731 (0.713–0.749)0.04331.02
CE-Agent (point-anchored bounds)∗^{*}0.721 (0.700–0.742)0.729 (0.711–0.747)0.04361.02
CE-Agent (prior-locked)0.717 (0.696–0.738)0.725 (0.707–0.744)0.04351.02
Logistic Regression0.725 (0.702–0.747)0.734 (0.713–0.749)0.04331.06
XGBoost Classifier0.724 (0.702–0.745)0.735 (0.717–0.751)0.04341.04
Random Forest / RSF0.725 (0.703–0.746)0.736 (0.716–0.752)—†^{\dagger}1.05
Cox Proportional Hazards0.719 (0.697–0.740)0.728 (0.706–0.746)—†^{\dagger}1.04
ACC/AHA ASCVD-PCE0.683 (0.659–0.705)0.693 (0.672–0.709)—‡^{\ddagger}—‡^{\ddagger}
AHA PREVENT (10-yr)0.680 (0.655–0.703)0.689 (0.668–0.706)—‡^{\ddagger}—‡^{\ddagger}
Framingham General CVD0.667 (0.644–0.690)0.674 (0.653–0.691)—‡^{\ddagger}—‡^{\ddagger}

Table A1. NLST test-cohort cardiovascular mortality: CE-Agent versus ML baselines and clinical calculators (n=15,931n = 15{,}931). Ten-year Aalen–Johansen competing-risk AUCcr_{\mathrm{cr}} (525 cardiovascular deaths and 1,684 competing deaths), full-follow-up CC-index (750 cardiovascular deaths over 13.4 yr), Brier score, and 10-yr E/Ocr_{\mathrm{cr}} across CE-Agent specifications, pooled 13-predictor survival and ML baselines, and clinical calculators (extended metrics in Table A3).

Model / SpecificationAUCcr\mathrm{AUC}_{\mathrm{cr}} (95% CI)tt-AUCcr\mathrm{AUC}_{\mathrm{cr}} / CSSlope (E/Ocr\mathrm{E}/\mathrm{O}_{\mathrm{cr}})NB5%\mathrm{NB}_{5\%} (10−310^{-3})
CE-Agent (hard, primary)0.7225 (0.700–0.744)0.7212 / 0.73860.951 (1.02)6.8
CE-Agent (soft, CV λ\lambda)0.7225 (0.700–0.744)0.7212 / 0.73860.951 (1.02)6.8
CE-Agent (sign-only)0.7225 (0.700–0.744)0.7213 / 0.73870.950 (1.02)6.8
CE-Agent (point-anchored)0.7208 (0.700–0.742)0.7215 / 0.73790.805 (1.02)6.4
CE-Agent (prior-locked)0.7166 (0.696–0.738)0.7179 / 0.73500.773 (1.02)6.6
Random Forest (pooled)0.7251 (0.703–0.746)0.7236 / 0.74101.143 (1.05)6.4
Logistic Regression (pooled)0.7248 (0.702–0.747)0.7223 / 0.73970.934 (1.06)6.5
XGBoost Classifier (pooled)0.7240 (0.702–0.745)0.7234 / 0.74040.865 (1.04)6.9
Cox PH (CIF, pooled)0.7188 (0.697–0.740)0.7186 / 0.73551.958 (1.04)5.5
ACC/AHA ASCVD-PCE0.6834 (0.659–0.705)0.6816 / 0.69731.060 (—)∗^{*}4.9∗^{*}
AHA PREVENT (10-yr)0.6801 (0.655–0.703)0.6778 / 0.69431.546 (—)∗^{*}4.4∗^{*}
Framingham General CVD0.6675 (0.644–0.690)0.6659 / 0.67900.992 (—)∗^{*}3.8∗^{*}

Table A3. Table A3 NLST test-cohort cardiovascular mortality: time-dependent discrimination, calibration, and net benefit (n=15,931n = 15{,}931). Bold marks overall column-best values (∣slope−1∣|\mathrm{slope} - 1| and ∣E/Ocr−1∣|\mathrm{E}/\mathrm{O}_{\mathrm{cr}} - 1| closest to zero; 1–10 yr tt-AUC(tt) curves in Figure 3c).

In the held-out MESA cohort (n=1,974n = 1{,}974), 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 (E/Ocr=1.08E/O_{\mathrm{cr}} = 1.08; 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 (tt-AUC) [41] on hard atherosclerotic cardiovascular disease (ASCVD; 0.802 versus 0.777, P=0.010P = 0.010; Methods) while correcting PCE’s 1.7-fold risk overestimation (E/Ocr=1.04E/O_{\mathrm{cr}} = 1.04 versus 1.70; Table A2 and Table A6). On the broad cardiovascular composite matched to Framingham, CE-Agent likewise exceeded Framingham in integrated tt-AUC (0.813 versus 0.789, P=0.012P = 0.012) and maintained accurate calibration (expected-to-observed ratio 0.97 versus 1.44; Table A6).

Model / SpecificationAUC10_{10} (95% CI)CC-index (95% CI)BrierE/Ocr_{\mathrm{cr}}
Cardiovascular mortality (n=1,974n = 1{,}974; 49 10-yr and 147 total deaths)
CE-Agent (hard, ln⁡(CAC+1)\ln(\mathrm{CAC} + 1), primary)0.807 (0.758–0.855)0.805 (0.773–0.835)0.02391.08
CE-Agent (soft, ln⁡(CAC+1)\ln(\mathrm{CAC} + 1))0.805 (0.752–0.854)0.807 (0.776–0.838)0.02401.08
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)0.02551.09
Unconstrained 22-Predictor Model0.789 (0.729–0.842)0.795 (0.761–0.829)0.02451.08
Cox Proportional Hazards0.797 (0.737–0.852)0.801 (0.767–0.836)0.02371.07
ACC/AHA ASCVD-PCE0.813 (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 CVD0.815 (0.764–0.864)0.772 (0.740–0.804)—‡—‡
Composite hard ASCVD (n=1,971n = 1{,}971; 137 10-yr and 283 total events)
CE-Agent (hard bounds, primary)0.774 (0.738–0.811)0.753 (0.727–0.780)0.06821.04
CE-Agent (soft bounds)0.776 (0.742–0.815)0.756 (0.729–0.783)0.06631.02
Unconstrained 22-Predictor Model0.773 (0.739–0.813)0.752 (0.725–0.779)0.06651.02
Cox Proportional Hazards0.772 (0.735–0.810)0.752 (0.725–0.778)0.06650.96
ACC/AHA ASCVD-PCE0.750 (0.711–0.785)0.730 (0.701–0.758)0.07241.70

Table A2. MESA test-cohort cardiovascular mortality and hard ASCVD: CE-Agent versus unconstrained models and clinical calculators. Ten-year discrimination (AUC10\mathrm{AUC}_{10}) and Brier score on uncensored 10-yr status (n=1,932n = 1{,}932 for mortality; 1,769 for hard ASCVD), full-follow-up CC-index, and 10-yr competing-risk E/Ocr_{\mathrm{cr}} (168 and 95 competing 10-yr deaths; extended benchmarks in Figure 3f and Table A4–Table A6).

ML ArchitectureAUC10_{10} (22/58)C10C_{10} (22/58)Int. tt-AUCBrierSlope (22/58)
Cox Proportional Hazards0.7972 / 0.81080.8059 / 0.82160.83730.023671.401 / 0.937
Random Survival Forest0.7832 / 0.78550.7970 / 0.80800.82360.023911.036 / 1.086
Logistic Regression0.7876 / 0.75100.7962 / 0.76140.82090.024810.732 / 0.520
Cox-Boosted Trees0.7793 / 0.80090.7907 / 0.81210.81920.025070.641 / 0.613
DeepSurv Neural Net0.7760 / 0.76010.7857 / 0.77080.80670.025040.617 / 0.329
XGBoost Classifier0.7658 / 0.74110.7753 / 0.75210.80170.024540.646 / 0.538
Decision Tree0.7133 / 0.68030.7204 / 0.68810.73620.025390.594 / 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 (n=1,974n = 1{,}974). Columns compare 22-predictor / 58-predictor test performance (14 model-by-feature configurations) alongside 22-predictor integrated tt-AUC (1–15 yr) and Brier score.

Endpoint & SpecificationAUC10_{10} (95% CI)CfullC_{\mathrm{full}}Int. tt-AUCBrierSlope (E/O)
Hard ASCVD (137/283 events; 95 competing 10-yr non-CVD deaths)
CE-Agent (hard bounds)0.7740 (0.738–0.811)0.75310.80200.068200.706 (0.93)
CE-Agent (soft bounds)0.7765 (0.742–0.815)0.75610.80190.066300.904 (0.92)
ACC/AHA ASCVD-PCE0.7503 (0.711–0.785)0.73020.77740.072410.799 (1.55)
Hard CVD (138/287 events; 94 competing 10-yr non-CVD deaths)
CE-Agent (hard bounds)0.7740 (0.738–0.811)0.75200.80120.068780.701 (0.95)
ACC/AHA ASCVD-PCE0.7475 (0.709–0.782)0.72710.77250.072910.790 (1.54)
Broad CVD (181/371 events; 85 competing 10-yr non-CVD deaths)
CE-Agent (hard bounds)0.7870 (0.755–0.817)0.76250.81320.084740.721 (0.97)
ACC/AHA ASCVD-PCE0.7512 (0.718–0.783)0.73630.77620.087460.806 (1.18)
Framingham General CVD0.7635 (0.731–0.794)0.73620.78850.086551.012 (1.44)

Table A6. MESA adjudicated composite endpoints: CE-Agent versus PCE and Framingham (n=1,971n = 1,971). 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: n=1,769n = 1,769 for hard ASCVD and hard CVD; n=1,778n = 1,778 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 (n=37,521n = 37{,}521). Each architecture has 1,056 single-hospital external evaluations; multi-hospital scaling is in Figure 4b and Table A10.

Model ArchitectureAUC (IQR)CE-Agent WinsNmin⁡N_{\min}Excess NB (10−310^{-3})
CE-Agent (point-anchored bounds)0.727 (0.726–0.727)264/2641,1100.83
Random Survival Forest0.686 (0.667–0.703)33/33***13,275—
Cox Proportional Hazards0.683 (0.648–0.703)33/33***13,6920.22
Gradient-Boosted Survival0.671 (0.655–0.688)33/33***15,7770.09
XGBoost Classifier0.658 (0.500–0.679)33/33***18,7900.29
Cox-Boosted Trees0.653 (0.500–0.683)33/33***19,943—
Logistic Regression0.643 (0.602–0.689)33/33***23,0290.37
Decision Tree0.625 (0.594–0.641)33/33***30,3320.23
DeepSurv Neural Network0.606 (0.556–0.642)33/33***42,3000.18

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 (n=6,795n = 6,795). Training groups are the six racial/ethnic and CT scanner subgroups with ≥20\ge20 training events (34–81 per fold); finite-sample scaling is in Figure 4e and Table A11.

Model ArchitectureAUC10_{10} (IQR)CE-Agent WinsNmin⁡N_{\min}Excess NB† (10−310^{-3})
CE-Agent (hard bounds)0.801 (0.787–0.807)43/482,2372.36
Random Survival Forest0.795 (0.784–0.804)4/6Undef.2.24
Cox Proportional Hazards0.796 (0.789–0.799)4/68,3802.25
Gradient-Boosted Survival0.776 (0.756–0.784)6/6Undef.1.43
XGBoost Classifier0.743 (0.734–0.761)6/6Undef.2.17
Cox-Boosted Trees0.794 (0.784–0.797)5/6Undef.2.04
Logistic Regression0.762 (0.719–0.779)6/68,3802.30
Decision Tree0.677 (0.646–0.690)6/6Undef.1.33
DeepSurv Neural Network0.756 (0.716–0.783)6/6Undef.1.81

Table A9.

BenchmarkProtocolCE-AgentBaselineEffect Size
(N × λ) inverse-sample-size scaling, adaptive weight selection, and separation control
1. NLST 1/nn Weight DecayK∈{1,2,4,8,16,32}K \in\{1, 2, 4, 8, 16, 32\} centers (up to 36.4k training participants)λ∗(K):1→0.1→10−2→10−3→10−4\lambda^*(K): 1 \to0.1 \to10^{-2} \to10^{-3} \to10^{-4}Subsample grid n=250→37.5kn = 250 \to37.5\mathrm{k}: λ∗=1→0\lambda^* = 1 \to0Slope rises 0.81→1.000.81 \to1.00 as λ∗∝1/n\lambda^* \propto1/n
2. NLST Sample-Size Selectorλlit(n)=Λ0/n\lambda_{\mathrm{lit}}(n) = \Lambda_0/n (Λ0=1,576.5\Lambda_0 = 1{,}576.5) if n<Nmin⁡n < N_{\min}Oracle gap ≤+0.19×10−3\le+0.19 \times10^{-3} across all K=1…32K = 1 \ldots32Pure inner CV (K=1K = 1): +1.06×10−3+1.06 \times10^{-3} (AUC 0.720)+0.0049+0.0049 AUC vs. fixed hard at K=32K = 32
3. NLST Prior-Shift ResilienceSynthetic shift Δ∈{0.10,0.25,0.50}\Delta\in\{0.10, 0.25, 0.50\} SD (28 cells)Inner CV (row 2 rule at n≥Nmin⁡n \ge N_{\min}): oracle gap ≤+7.2×10−3\le+7.2 \times10^{-3}Fixed hard bounds: up to +102.5×10−3+102.5 \times10^{-3}Inner CV detects shift in 28/28 cells
4. MESA Free-Column N(0,1)\mathcal{N}(0,1)Unit-normal prior (λfree=1/n\lambda_{\mathrm{free}} = 1/n) on unpriced cols0/50 separated; n=150n = 150: 0.7407, n=250n = 250: 0.7607Unreg. cols: 29/50 (0.6512), 13/50 (0.7171)+0.0895 (n=150n = 150), +0.093 Chinese fold, +0.101 scanner fold
5. MESA Adaptive 1/n1/n ScheduleΛ0,fr/n\Lambda_{0,\mathrm{fr}}/n (Λ0,fr=31.6\Lambda_{0,\mathrm{fr}} = 31.6) across 12 cellsOracle gap [−0.9,+0.7]×10−3[-0.9, +0.7] \times10^{-3} (12/12 CIs cover 0)Pure inner CV (n=250n = 250): +521×10−3+521 \times10^{-3} oracle gapHard ASCVD (λ∗=0.01\lambda^* = 0.01): ∣slope−1∣|\mathrm{slope} - 1| 0.294 →\to 0.096
Causal invariance, DRO, Steyerberg updating, and Gaussian MAP controls
6. Causal Anchor Regression [42]g∈[0,103]g \in[0, 10^3] across 33 NLST & 16 MESA OOD foldsNLST med. 0.7268; MESA med. 0.7752Anchor best gg: NLST 0.6619; MESA 0.746633/33 NLST (+0.0649 in medians); 15/16 MESA (med. paired +0.0345)
7. Label-Free χ2\chi^2 & CVaR DRO [43]Oracle across 7 χ2\chi^2 radii & 7 CVaR tail levelsNLST med. 0.7268; MESA med. 0.7752Oracle DRO: NLST 0.6460; MESA 0.7446 / 0.747533/33 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-oddsPrevalence-only envelope: 0.7268 / 0.7752Offset LR / XGB: NLST 0.6465 / 0.6929; MESA 0.7406Outperforms LR / XGB in 33/33 & 28/33 NLST, 16/16 MESA
9. Gaussian Point-Prior MAPWidth-matched N(cjpub,σj2)\mathcal{N}(c_j^{\mathrm{pub}}, \sigma_j^2), σj=(uj−ℓj)/3.92\sigma_j = (u_j - \ell_j)/3.92KKT exact c^j=0\hat{c}_j = 0: 95.1%–98.6% of fitsGaussian MAP exact c^j=0\hat{c}_j = 0: 0.0% (68–115/200 TAC flip)Retains full AUC (∣Δ∣≤0.001|\Delta| \le0.001), 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 (K=1K = 1 hospital) NLST external evaluation (a) and held-out MESA hard ASCVD (d; NB, net benefit; ptp_t, decision threshold; shading, ptp_t of 5% to 10%). b,e, Minimum sample size Nmin⁡N_{\min} under the criteria of Riley et al. [23] as constraints are added, in NLST (b; pp, free slope parameters; soft, λ≥1\lambda\ge1; Site LR, site-level logistic regression; Max site, largest single hospital) and for MESA hard ASCVD and cardiovascular mortality (e; peffp_{\mathrm{eff}}, 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 RegimeCE-Agent Unboxed* /Unc.{sup}†`Cox PHRSFGBSAXGBDSurv
External OOD discrimination (mean AUC across 33 − KK centers)
K=1K = 1 Single Hospital0.72460.69770.66760.67610.66200.61780.6007
K=2K = 2 Pooled Hospitals0.72460.7134—————
K=4K = 4 Pooled Hospitals0.72660.7228—————
K=8K = 8 Pooled Hospitals0.72580.7244—————
K=16K = 16 Pooled Hospitals0.72510.7245—————
K=32K = 32 LOCO (32 → 1)0.72560.72510.72560.7260—0.72800.7271
Calibration slope error (∣median⁡ slope⁡−1∣\lvert\operatorname{median}\ \operatorname{slope} - 1\rvert), sample-size criteria, and DCA
K=1K = 1 |Med. slope − 1|0.1780.512{sup}†`———0.400—
K=2K = 2 |Med. slope − 1|0.1730.183{sup}†`———0.468—
K=4K = 4 |Med. slope − 1|0.1600.072{sup}†`———0.388—
K=8K = 8 |Med. slope − 1|0.1660.035{sup}†`———0.260—
K=16K = 16 |Med. slope − 1|0.1660.021{sup}†`———0.163—
K=32K = 32 |Med. slope − 1|0.1530.017{sup}†`———0.068—
Riley Free Params / Cols. pp4.012.0{sup}†`23–3423–3423–3423–3423–34
Riley Required Nmin⁡N_{\min}1,1106,043{sup}†`13,69213,27515,77718,79042,300
DCA Int. Excess NB (×10−310^{-3})0.830.50{sup}†`0.22—0.090.290.18
DCA Mean Excess NB (×10−310^{-3})4.362.42{sup}†`1.03—0.974.162.17
DCA Positive Cells (/330)289/330196/330{sup}†`213/330—125/330132/330131/330

Table A10. NLST models pooled across 1 to 32 training hospitals: external discrimination, calibration, minimum sample size, and net benefit (n=37,521n = 37,521).

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 ArchitectureAUC10_{10} (95% CI)CfullC_{\mathrm{full}}BrierSlope (E/O)
CE-Agent (compiled hard bounds)0.8029 (0.750–0.851)0.80660.023721.126 (1.06)
CE-Agent (point-anchored soft bounds)0.7961 (0.741–0.850)0.80520.023760.954 (1.07)
CE-Agent (point-anchored hard bounds)0.7940 (0.739–0.847)0.80430.023790.913 (1.07)
Unconstrained Logistic Reg.0.8029 (0.750–0.852)0.80230.023711.011 (1.06)
DeepSurv Neural Network0.8022 (0.748–0.852)0.80300.023711.026 (1.06)
Cox Proportional Hazards0.7921 (0.735–0.845)0.79770.024032.107 (1.05)
Cox-Boosted Trees (XGBoost)0.7761 (0.718–0.832)0.78660.024080.871 (1.05)
Random Survival Forest (RSF)0.7713 (0.710–0.827)0.7817——
XGBoost Classifier0.7699 (0.715–0.823)0.78070.024250.927 (1.05)
Classification Decision Tree0.7674 (0.720–0.814)0.75540.024201.286 (1.05)

Table A7. NLST-trained models applied to MESA cardiovascular mortality on 12 shared predictors (n=1,974n = 1,974). Models are trained exclusively on NLST (n=37,521n = 37,521) 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: AUC10_{10} 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 (R2R^2) for predicting MESA’s measured intermediate biomarkers from their parent variables with n=100n = 100 training participants (bars) and with the full training cohort (n=4,821n = 4{,}821; 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 (PM2.5_{2.5}, fine particulate matter; SBP, systolic blood pressure). d, Expected-to-observed ratio (E/O; band, 0.90 to 1.10) and 10-year concordance (C10C_{10}) 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 (n=5,269n = 5{,}269; 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 (K=1K=1 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 (n=5,269n=5{,}269) or any (n=6,467n=6{,}467) 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.

ExperimentProtocolCE-AgentML BaselineEffect Size
1. Interval-Bound Sample ScalingK∈{1,2,4,8,16,32}K \in\{1, 2, 4, 8, 16, 32\} hospitals (4 active bounds)K=1K = 1: 0.7246 → K=32K = 32: 0.7256Four bounds freed: 0.6977 → 0.7251; Cox: 0.6676 → 0.7256+0.0268 (K=1K = 1, P<0.001P < 0.001) → +0.0005 (K=32K = 32)
2. Factorial Bound Attribution (242^4)All 16 bound coalitions across 33 centers (K=1K = 1)Joint 4-bound ΔAUC=+0.0268\Delta\mathrm{AUC} = +0.0268Sum of singletons ∑Δj=+0.0323\sum\Delta_j = +0.0323; Four bounds freed 0.697783.0% main-effect share; interaction max⁡∣IS∣=0.0036\max|I_S| = 0.0036
3. CAC Provenance AuditReplace the NLST CAC estimate [8] with the ITALUNG trial estimate [65] (K=1K = 1)NLST estimate 0.7246; ITALUNG estimate 0.7248Free CAC 0.7134; mean RSF 0.6761; Cox 0.6676; LR 0.6376ΔAUC=+0.0002\Delta\mathrm{AUC} = +0.0002; free-CAC model above 8/8 baselines
4. Sign-Inversion PerturbationAll 29=5122^9 = 512 sign coalitions (0 → 9 inverted)0 flipped: med. AUC = 0.72463 flipped: 0.596; 6 flipped: 0.431; 9 flipped: 0.290ΔAUC=−0.434\Delta\mathrm{AUC} = -0.434 (0 → 9 inverted)
5. Positive Scale ResilienceGlobal prior multiplier s∈[0.2×,5.0×]s \in[0.2\times, 5.0\times] (25× span)Held-out C∈C \in [0.7297, 0.7306]Sign inversion ΔAUC=−0.434\Delta\mathrm{AUC} = -0.434 (480× larger)∣ΔC∣≤0.0009|\Delta C| \le0.0009 across 25× span
6. Algebraic EquivalenceInclude derived vascular-aging composite (n=15,931n = 15{,}931)Screened: C=0.7306C = \mathbf{0.7306}Included: r=0.9999999999r = 0.9999999999, max⁡∣Δη∣=1.78×10−15\max|\Delta\eta| = 1.78 \times10^{-15}Δrank=0\Delta\mathrm{rank} = 0 (Proposition 1)
7. Composite RelaxationScreened vs. progressive relaxation at K=1K = 1 (33 centers)Screened, published weights: 0.7251 (κ\kappa-ratio 1.0×)Relax BMI: 0.7190; Shared-parent: 0.6985; All 4 free: 0.6947 (κ\kappa-ratio 321×)ΔAUC=−0.0304\Delta\mathrm{AUC} = -0.0304 (30/33 centers)
8. 3D LDCT ExpansionTraining LDCT subcohorts (n=5,269n = 5{,}269; 6,467; 1,243), 5-fold CV and cross-site: +3 LDCTAUC: 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 CC-index: 0.6926ΔAUC=+0.011\Delta\mathrm{AUC} = +0.011 to +0.017 (6/6 comparisons)
9. Zeroing of Reversed WeightsCOPD-confounded PA diameter ratio (1,040 CV fits)c^PA=0.000\hat{c}_{\mathrm{PA}} = \mathbf{0.000} in 928/1,040 (89.2%)Direction constraint lifted: c^unc∈[−0.25,−0.12]\hat{c}_{\mathrm{unc}} \in[-0.25, -0.12]89.2% (928/1,040) zeroed*

Table A14. Mechanistic ablations of the Causal Envelope in NLST (n=37,521n = 37{,}521 training; rows 5–6, n=15,931n = 15{,}931 held-out test).

ExperimentProtocolCE-AgentML BaselineEffect Size
1. Interval-Bound Sample Scalingn∈{250,500,1,000,2,000,4,821}n \in\{250, 500, 1{,}000, 2{,}000, 4{,}821\} (5 active bounds)n=250n = 250: 0.7171 → Full: 0.8068Five 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 (252^5)All 32 coalitions at n=500n = 500 (50 draws) & full splitn=500n = 500: Δall=+0.0569\Delta_{\mathrm{all}} = +0.0569; Full +0.0086n=500n = 500: ∑Δj=+0.0542\sum\Delta_j = +0.0542 (95.3%); Full ∑Δj=+0.0087\sum\Delta_j = +0.0087 (101%)95.3%–101% additive; interaction ∣Ijk∣≤0.0044\lvert I_{jk}\rvert\le0.0044
3. CAC Provenance AuditBinary free/locked vs. Detrano interval [37] vs. Drop CACCVD bin. free 0.8107 (locked 0.7983); ln⁡(CAC+1)\ln(\mathrm{CAC}+1) 0.8068Unc. 0.7886; RSF 0.7832; Drop CAC: ASCVD ΔC=−0.0106\Delta C = -0.0106 (95% CI −0.017 to −0.005)+0.0043 on ASCVD at n=500n = 500 (40/50 draws)*
4. Sign-Inversion Perturbation0 → 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 flippedΔAUC10=−0.329\Delta\mathrm{AUC}_{10} = -0.329 (CVD), −0.214 (ASCVD)
5. Positive Scale ResilienceGlobal prior multiplier s∈[0.5×,2.0×]s \in[0.5\times, 2.0\times] (0.2 ×–5×)AUC10\mathrm{AUC}_{10} 0.787–0.810 across 0.5 ×–2×Zero-locked ridge (8 free): AUC10=0.7130\mathrm{AUC}_{10} = 0.7130 (−0.094)∣ΔAUC10∣≤0.020\lvert\Delta\mathrm{AUC}_{10}\rvert\le0.020 (16× smaller vs. sign flip)
6. Algebraic EquivalenceInclude derived composites (n=1,974n = 1{,}974 test)Screened: AUC10=0.8068\mathrm{AUC}_{10} = \mathbf{0.8068}Included: max⁡∣Δη∣=6.08×10−7\max\lvert\Delta\eta\rvert= 6.08 \times10^{-7}, ΔAUC10=0.0000\Delta\mathrm{AUC}_{10} = 0.0000ΔAUC10=0.0000\Delta\mathrm{AUC}_{10} = 0.0000 (Proposition 1)
7. Composite RelaxationScreened vs. progressive relaxation at n=250n = 250 & n=500n = 500 (200 event-stratified draws)Screened: 0.7150 (n=250n = 250), 0.7662 (n=500n = 500)All 10 free: 0.6299 (n=250n = 250, κ=2.2×1010\kappa= 2.2 \times10^{10}), 0.7277 (n=500n = 500)ΔAUC10=−0.0851\Delta\mathrm{AUC}_{10} = -0.0851 (n=250n = 250)
8. 5-Phenotype CT ProgressionNo CT → +CAC → +TAC → +Chamber → +Ao. diam. → +Liver fatASCVD 0.7576 → 0.7744; 0.7983 → 0.8069Unc. LR: 0.7468 → 0.7725 (ASCVD), 0.7920 → 0.7876 (CVD)ASCVD ΔAUC10=+0.0168\Delta\mathrm{AUC}_{10} = +0.0168 (95% CI +0.006 to +0.027)
9. Zeroing of Reversed Weights213 fits (200 bootstrap, 5 cross-validation, and 8 single-group)max⁡∣νjc^j∣≤4.4×10−9\max\lvert\nu_j \hat{c}_j\rvert\le4.4 \times10^{-9}; ΔAUC10=+0.0123\Delta\mathrm{AUC}_{10} = +0.0123Unconstrained flips 6/14 (42.9%) signs (AUC10=0.7886\mathrm{AUC}_{10} = 0.7886)63.4%–84.5% zeroed across 6 predictors

Table A15. Ablations of the Causal Envelope in MESA (n=6,795n = 6{,}795).

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; PM2.5_{2.5}, 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 (∣ΔC∣|\Delta C| index∣≤0.0009| \le0.0009) across a 25× span in NLST, d; AUC10_{10} 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 1/n1/n penalty (adaptive 1/n1/n-to-CV rule in a; pure Λ0/n\Lambda_0/n in e), or cross-validated (CV) soft bounds against the unconstrained model (left axis), with the AUC-optimal penalty weight λ∗\lambda^{*} (amber, right axis), across pooled NLST hospitals (a) and MESA training subsamples (e; shaded band, small-sample gain from the 1/n1/n penalty). b,f, Single-environment external discrimination of CE against Anchor Regression [42], χ2\chi^2 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).

BenchmarkProtocolCE-AgentBaselineClinical Effect
Counterfactual interventional direction consistency across single-environment and pooled fits
1. NLST Smoking & BMI (K=1K = 1)Cessation, −10/−20-10/-20 pack-yr, BMI −5-5 [5, 46]0.0% rises (0/33 K=1K = 1, 0/50 subsample rev.)K=1K = 1 rev.: 9% quit, 21% pack-yr, 41% BMI (51.6% rises)LOCO BMI rev. in 4/8 ML (55%–62% rises)*
2. MESA LDL (−1-1 mmol/L)Counterfactual Δ\DeltaLDL (n=1,968n = 1{,}968 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 (−10-10 mmHg)Counterfactual Δ\DeltaSBP 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% risesPrevents tree/neural local non-monotonicity
Two-stage clinical CT reading and statin triage efficiency
4. NLST Opportunistic LDCT TriagePrimary-prev. LDCT train (n=15,727n = 15{,}727, 31 non-zero-event hospital fits, pt=5.0%p_t = 5.0\%) [48, 49]LOCO +2.52/1k (18.3% CT); K=1K = 1 +1.48/1kIntermediate PCE →\to CAC: +1.38/1k (40.5% CT); K=1K = 1 ML < 0−55% reads (+1.14/1k, 95% CI 0.23–2.13, P=0.019P = 0.019); 27/31 vs. K=1K = 1 ML†
5. MESA 2018 ACC/AHA CAC TriageLipid-therapy-free test ages 45–75 (n=1,465n = 1{,}465, pt=7.5%p_t = 7.5\%) [17]Equal-budget (28.5% CT): +16.0/1k; Two-stage selective CT (7.1% CT, 24.9% statin): +13.0/1kGuideline PCE →\to 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 (W^90\widehat{W}_{90}) 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 (σinput\sigma_{\mathrm{input}}), 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 r≥0.9998r \ge0.9998 between risk estimates. ∣E∣|E|, 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 (n=15,727n = 15{,}727), 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 (n=1,465n = 1,465), 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 (n=15,931n = 15,931 held-out participants; Figure 8 and Table A20–Table A22), CE-Agent achieved accurate 6-year calibration (E/Ocr_{\mathrm{cr}} = 0.99, versus 0.86 for published PLCOm2012) and matched recalibrated PLCOm2012 in competing-risk AUC (AUCcr_{\mathrm{cr}} = 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 (AUCcr_{\mathrm{cr}} = 0.710), and maintained accurate calibration (E/Ocr_{\mathrm{cr}} = 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 (AUCcr\mathrm{AUC}_{\mathrm{cr}}; 95% CI) and E/Ocr_{\mathrm{cr}} 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 (n=7,786n = 7{,}786) 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 / SpecificationLOCO LDCTTest AUCcr_{\mathrm{cr}}Slope (E/Ocr_{\mathrm{cr}})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, Λ0/n\Lambda_0/n)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 / QuadrantCT / FlagAll (1.5%/5%)LDCT (1.5%/5%)All (2%/5%)LDCT (2%/5%)
Joint cardiopulmonary triage net benefit NBsumNB_{\mathrm{sum}} 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 (n=15,931n = 15{,}931)
Dual high (Lung+^{+}CVD+^{+})17.8% (2,832)170 (29.5%) 6-yr lung cancers; 226 (42.9%) 10-yr CVD deaths
Lung-only (Lung+^{+}CVD−^{-})65.3% (10,401)385 (67.0%) lung cancers (CVD-missed); 243 (46.5%) CVD deaths
CVD-only (Lung−^{-}CVD+^{+})0.8% (135)2 (0.3%) lung cancers; 12 (2.3%) CVD deaths (45.2% CVD+^{+} total)
Dual low (Lung−^{-}CVD−^{-})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 (n=15,931n = 15{,}931; LDCT arm n=7,786n = 7{,}786).

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 (n=7,786n = 7,786), adding radiologist-reported nodules and visual emphysema [7] raised AUCcr_{\mathrm{cr}} from 0.685 to 0.708 (P=0.036P = 0.036; Figure 8c). With up to 60% of predictors missing in the training cohort (n=37,521n = 37,521, 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, P<0.001P < 0.001; 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 (n=7,786n = 7{,}786), 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, P=0.024P = 0.024; 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 D={(xi,yi)}i=1n\mathcal{D} = \{(x_i, y_i)\}_{i=1}^{n} of nn participants with baseline clinical and chest CT records xix_i across qq candidate variables and adjudicated binary cardiopulmonary outcomes yi∈{0,1}y_i \in\{0,1\}. 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 dd admissible baseline predictors Vadm⊆{1,…,q}V_{\mathrm{adm}} \subseteq\{1,\ldots,q\} (re-indexed as {1,…,d}\{1,\ldots,d\}). Each observed entry XijX_{ij} is standardized to Zij=(Xij−μ^j)/σ^jZ_{ij} = (X_{ij} - \hat{\mu}_j)/\hat{\sigma}_j using center μ^j\hat{\mu}_j and positive scale σ^j>0\hat{\sigma}_j > 0 (empirical training moments in MESA and pre-specified cohort scales in NLST; Appendix C), forming Z∈Rn×dZ \in\mathbb{R}^{n \times d}. Verified published effect ranges are then compiled into a convex parameter set, the Causal Envelope ΩCE⊆R1+d\Omega_{\mathrm{CE}} \subseteq\mathbb{R}^{1+d}, before viewing outcome labels. In Stage II (Figure 1b,Figure 1c), CE-Agent estimates the unpenalized baseline intercept c^0∈R\hat{c}_0 \in\mathbb{R} and standardized log-odds weights c^∈Rd\hat{\boldsymbol{c}} \in\mathbb{R}^{d} within ΩCE\Omega_{\mathrm{CE}}. At deployment, analytic variance propagation outputs calibrated risk p^i∈(0,1)\hat{p}_i \in(0,1), 90% predictive interval width W^90,i\hat{W}_{90,i}, and missing-data log-odds uncertainty σinput,i≥0\sigma_{\mathrm{input},i} \geq0.

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 Cℓ=∑p∈pa(Cℓ)wpℓZpC_{\ell} = \sum_{p \in\mathrm{pa}(C_{\ell})} w_{p\ell} Z_p formed with fixed weights wpℓ∈Rw_{p\ell} \in\mathbb{R} from retained standardized parent columns Zp∈RnZ_p \in\mathbb{R}^{n} (pa(Cℓ)⊆Vadm\mathrm{pa}(C_{\ell}) \subseteq V_{\mathrm{adm}}) lies in the column space of ZZ. Excluding CℓC_{\ell} 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 MmM_m with standardized column Zm∈RnZ_m \in\mathbb{R}^{n}—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 MmM_m is unmeasured at deployment, CE-Agent projects its conditional mean μim\mu_{im} and variance vimv_{im} recursively from its parents, using sign-constrained training weights (b^m,w^pm)(\hat{b}_m,\hat{w}_{pm}), parent moments (μip,vip)(\mu_{ip},v_{ip}) equal to (Zip,0)(Z_{ip}, 0) when observed, and the training residual variance σ^m2\hat{\sigma}_m^2:

μim=b^m+∑p∈pa⁡(Mm)w^pmμip,vim=σ^m2+∑p∈pa⁡(Mm)w^pm2vip,σ^m2=Var⁡^(Zm−μm).(1)\begin{aligned} \mu_{im} &= \hat{b}_m + \sum_{p\in\operatorname{pa}(M_m)} \hat{w}_{pm}\mu_{ip},\\ v_{im} &= \hat{\sigma}_m^2 + \sum_{p\in\operatorname{pa}(M_m)} \hat{w}_{pm}^{2}v_{ip}, \qquad\hat{\sigma}_m^2 = \widehat{\operatorname{Var}}(Z_m-\mu_m). \tag*{(1)} \end{aligned}

Here μm=(μ1m,…,μnm)⊤\mu_m=(\mu_{1m},\ldots,\mu_{nm})^{\top}, and the variance recursion treats parent projection errors as independent. Binary intermediate biomarkers Mm∈{0,1}M_m\in\{0,1\} 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 μim\mu_{im}, 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 RegimeFull CCAUC10_{10}BrierE/O (95% CI)Slope
1. Observed CT0.80460.80680.023901.07 (0.84–1.47)0.777
2. Population-Mean Fill (ZCT=0Z_{\mathrm{CT}} = 0)0.78720.79540.023630.81 (0.63–1.11)0.907
3. Projected CT (rel.-gated)0.78840.79610.023600.84 (0.65–1.16)0.884
4. Projected CT (var.-gated)0.79030.79700.023590.95 (0.74–1.31)0.826
5. Projected CT (unattenuated)∗^{*}0.79040.79710.023600.96 (0.74–1.32)0.823
6. Retrained No-CT (18 features)0.79030.79830.023561.06 (0.82–1.45)0.926
7. External ML (84 configs)0.673–0.8000.676–0.8060.0235–0.02580.59–1.190.43–3.09†^{\dagger}

Table A13. MESA cardiovascular mortality when all five CT phenotypes are unavailable at prediction (n=1,974n = 1{,}974). 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 Δxjpub>0\Delta x_j^{\mathrm{pub}}>0, 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 [βj,loraw,βj,hiraw][\beta_{j,\mathrm{lo}}^{\mathrm{raw}},\beta_{j,\mathrm{hi}}^{\mathrm{raw}}] and point estimate βjraw\beta_j^{\mathrm{raw}} 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 κj=σ^j/Δxjpub\kappa_j=\hat{\sigma}_j/\Delta x_j^{\mathrm{pub}} and records the established biological direction δj=sign⁡(κjβjraw)∈{−1,+1}\delta_j=\operatorname{sign}(\kappa_j\beta_j^{\mathrm{raw}})\in\{-1,+1\}. 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 Csign⊆VadmC_{\mathrm{sign}}\subseteq V_{\mathrm{adm}}, 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 Cfree=Vadm∖CsignC_{\mathrm{free}}=V_{\mathrm{adm}}\setminus C_{\mathrm{sign}}. 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 CsignC_{\mathrm{sign}}, predictors passing all four checks form the interval-bounded subset CboxC_{\mathrm{box}} with rescaled bounds [ℓj,uj][\ell_j,u_j]. Established clinical risk factors fixed at published point estimates form Cpoint⊆Csign∖CboxC_{\mathrm{point}}\subseteq C_{\mathrm{sign}}\setminus C_{\mathrm{box}}, and predictors flagged on any compatibility check retain directional sign constraints in Csign∖(Cbox∪Cpoint)\mathcal{C}_{\mathrm{sign}} \setminus(\mathcal{C}_{\mathrm{box}} \cup\mathcal{C}_{\mathrm{point}}). A fixed weight is either cjpub=κjβjrawc_j^{\mathrm{pub}} = \kappa_j\beta_j^{\mathrm{raw}} or, in single-hospital NLST fits and in MESA, a weight obtained by numerically differentiating published risk equations [15, 16, 19]. For intercept c0∈Rc_0 \in\mathbb{R} and weight vector c=(c1,…,cd)T∈Rd\mathbf{c} = (c_1, \ldots, c_d)^{\mathsf{T}} \in\mathbb{R}^d, the compiled Causal Envelope is (Figure 1a and Figure 2a):

[ℓj,uj]=κj[βj,loraw,βj,hiraw],ΩCE={(c0,c)∈R1+d | cj∈[ℓj,uj] ∀j∈Cbox,cj=cjpub ∀j∈Cpoint, δjcj≥0 ∀j∈Csign}.(2)\begin{aligned} [\ell_j,u_j] &= \kappa_j[\beta_{j,\mathrm{lo}}^{\mathrm{raw}},\beta_{j,\mathrm{hi}}^{\mathrm{raw}}],\\ \Omega_{\mathrm{CE}} &= \left\{(c_0,c)\in\mathbb{R}^{1+d}\ \middle|\ c_j\in[\ell_j,u_j]\ \forall j\in\mathcal{C}_{\mathrm{box}},\right.\\ &\qquad\left. c_j=c_j^{\mathrm{pub}}\ \forall j\in\mathcal{C}_{\mathrm{point}},\ \delta_jc_j\ge0\ \forall j\in\mathcal{C}_{\mathrm{sign}}\right\}. \tag*{(2)} \end{aligned}

In primary pooled multi-center NLST estimation of cardiovascular (d=13d=13, three interval bounds) and lung cancer (d=11d=11, five interval bounds) risk, event counts far exceed dd, so predictors with compatible published intervals enter Cbox\mathcal{C}_{\mathrm{box}} and Cpoint\mathcal{C}_{\mathrm{point}} 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 (d=22d=22), core clinical risk factors are fixed in Cpoint\mathcal{C}_{\mathrm{point}}, 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 Cbox\mathcal{C}_{\mathrm{box}}. 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 Ωsign={(c0,c)∈R1+d∣δjcj≥0 ∀j∈Csign}\Omega_{\mathrm{sign}}=\{(c_0,c)\in\mathbb{R}^{1+d}\mid\delta_jc_j\ge0\ \forall j\in\mathcal{C}_{\mathrm{sign}}\}, linear log-odds ηi=c0+∑j=1dcjZij\eta_i=c_0+\sum_{j=1}^{d}c_jZ_{ij}, and empirical logistic negative log-likelihood Lemp(c0,c;D)=−1n∑i=1n[yiηi−ln⁡(1+eηi)]\mathcal{L}_{\mathrm{emp}}(c_0,c;\mathcal{D})=-\frac{1}{n}\sum_{i=1}^{n}[y_i\eta_i-\ln(1+e^{\eta_i})]. CE-Agent computes (c^0,c^)(\hat{c}_0,\hat{c}) by solving the convex optimization problem:

min⁡(c0,c)∈Ωsign Jλ,λfree(c0,c;D)=Lemp(c0,c;D)+λ∑j∈Cboxϕbox(cj;ℓj,uj)+λ∑j∈Cpoint(cj−cjpub)2+λfree2∑j∉Cbox∪Cpointcj2.(3)\min_{(c_0,c)\in\Omega_{\mathrm{sign}}}\ \mathcal{J}_{\lambda,\lambda_{\mathrm{free}}}(c_0,c;\mathcal{D})=\mathcal{L}_{\mathrm{emp}}(c_0,c;\mathcal{D})+\lambda\sum_{j\in\mathcal{C}_{\mathrm{box}}}\phi_{\mathrm{box}}(c_j;\ell_j,u_j)+\lambda\sum_{j\in\mathcal{C}_{\mathrm{point}}}(c_j-c_j^{\mathrm{pub}})^2+\frac{\lambda_{\mathrm{free}}}{2}\sum_{j\notin\mathcal{C}_{\mathrm{box}}\cup\mathcal{C}_{\mathrm{point}}}c_j^2. \tag*{(3)}

Here ϕbox(cj;ℓj,uj)=max⁡(0,ℓj−cj,cj−uj)2\phi_{\mathrm{box}}(c_j;\ell_j,u_j)=\max(0,\ell_j-c_j,c_j-u_j)^2 is a continuously differentiable squared-hinge penalty outside [ℓj,uj][\ell_j,u_j], and c0c_0 remains unpenalized (in NLST competing-risk evaluations, c^0\hat{c}_0 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 λfree≥0\lambda_{\mathrm{free}}\ge0 provides optional weak per-SD ridge stabilization solely on predictors without interval or point priors (λfree=0\lambda_{\mathrm{free}}=0 in primary fits and λfree(n)=1/n\lambda_{\mathrm{free}}(n)=1/n in small-sample MESA subcohorts; Appendices B–C). Taking λ→∞\lambda\to\infty yields the primary hard-bound estimator with exact membership (c^0,c^)∈ΩCE(\hat{c}_0,\hat{c}) \in\Omega_{\mathrm{CE}} and zero hyperparameter tuning, whereas finite λ=Λ0/n\lambda= \Lambda_0/n keeps the total prior precision Λ0\Lambda_0 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 λ=0\lambda= 0 is the otherwise unconstrained maximum-likelihood fit. By Karush–Kuhn–Tucker (KKT) optimality conditions, directional sign constraints preserve concordant empirical associations and set c^j=0\hat{c}_j = 0 on sign-only predictors j∈Csign∖(Cbox∪Cpoint)j \in\mathcal{C}_{\mathrm{sign}} \setminus(\mathcal{C}_{\mathrm{box}} \cup\mathcal{C}_{\mathrm{point}}) whenever the empirical gradient favors the sign opposite to δj\delta_j, as when confounding reverses the association (Figure 2b and Proposition 2). Restricting parameters to ΩCE\Omega_{\mathrm{CE}} lowers the minimum development sample size Nmin⁡N_{\min} of Riley et al. [23] through two routes: fewer effective degrees of freedom (peff≤dp_{\mathrm{eff}} \leq d) and higher anticipated discrimination of the constrained model, with the anticipated Cox–Snell R2R^2 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 j∈Vadmj \in\mathcal{V}_{\mathrm{adm}} in patient record xix_i carries mean μij\mu_{ij}, missing-data variance vijv_{ij}, and missingness indicator mij∈{0,1}m_{ij} \in\{0,1\}, taking (Zij,0,0)(Z_{ij},0,0) when observed and imputed or projected moments (μij,vij,1)(\mu_{ij},v_{ij},1) 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 expit⁡(u)=(1+e−u)−1\operatorname{expit}(u) = (1 + e^{-u})^{-1} denoting the logistic sigmoid, a single reverse-topological pass with out-of-fold gate parameters αg=(α0,αm,αv)∈R×R≥02\alpha_g = (\alpha_0,\alpha_m,\alpha_v) \in\mathbb{R} \times\mathbb{R}_{\geq0}^{2} computes each predictor’s reliability gate gij(αg)∈(0,1)g_{ij}(\alpha_g) \in(0,1), reported as the reliability score in Results. The same pass returns the total pathway sensitivities TijT_{ij}, the summed effect of predictor jj on the log-odds over all directed paths to the outcome. Each link weight entering TijT_{ij} is multiplied by the gate of its source predictor (Appendix B), so missing or uncertain predictors contribute less to the propagated log-odds mean MiM_i and missing-data variance ViV_i:

gij(αg)=expit⁡(α0−αmmij−αvvij),g_{ij}(\alpha_g) = \operatorname{expit}(\alpha_0 - \alpha_m m_{ij} - \alpha_v v_{ij}),
Mi=∑j∈VadmTijμij,Vi=∑j∈VadmTij2vij+Vlat,i.(4)M_i = \sum_{j \in\mathcal{V}_{\mathrm{adm}}} T_{ij}\mu_{ij}, \qquad V_i = \sum_{j \in\mathcal{V}_{\mathrm{adm}}} T_{ij}^{2}v_{ij} + V_{\mathrm{lat},i}. \tag*{(4)}

where Vlat,i≥0V_{\mathrm{lat},i} \geq0 captures latent-pathway residual variance when latent nodes are modeled (Appendix B). With training-prevalence intercept variance σb2=(nyˉ(1−yˉ))−1\sigma_b^2 = \left(n\bar{y}(1-\bar{y})\right)^{-1}, fixed-size Gauss–Hermite quadrature over ξ∼N(0,1)\xi\sim\mathcal{N}(0,1) and safeguarded Newton–bisection inversion of the log-odds mixture cumulative distribution Fi(t)F_i(t) (t∈Rt \in\mathbb{R}; full mixture quadrature in equation (9), Appendix B) yield calibrated risk p^i\hat{p}_i, missing-data log-odds uncertainty σinput,i\sigma_{\mathrm{input},i}, and 90% predictive interval width W^90,i\widehat{W}_{90,i}:

p^i=Eξ[expit⁡(c^0+Mi+(σb2+Vi)1/2ξ)],σinput,i=Vi1/2,\hat{p}_i = \mathbb{E}_{\xi}\left[\operatorname{expit}\left(\hat{c}_0 + M_i + \left(\sigma_b^2 + V_i\right)^{1/2}\xi\right)\right], \qquad\sigma_{\mathrm{input},i} = V_i^{1/2},
W^90,i=expit⁡(Fi−1(0.95))−expit⁡(Fi−1(0.05)).(5)\widehat{W}_{90,i} = \operatorname{expit}\left(F_i^{-1}(0.95)\right) - \operatorname{expit}\left(F_i^{-1}(0.05)\right). \tag*{(5)}

By first-order delta expansion, W^90,i≈3.29 SD⁡(ηi)p^i(1−p^i)\widehat{W}_{90,i} \approx3.29\,\operatorname{SD}(\eta_i)\widehat{p}_i(1-\widehat{p}_i), where SD⁡(ηi)\operatorname{SD}(\eta_i) is the standard deviation of the predicted log-odds, combining baseline, missing-data, and parameter variance (equation (B5)). This width is largest near p^i=0.50\widehat{p}_i = 0.50 whether or not a record is complete. Uncertainty-guided chart review therefore ranks records by σinput,i\sigma_{\mathrm{input},i}—or its cohort-measured missing-predictor component σmiss,i\sigma_{\mathrm{miss},i} (Appendix B)—to isolate sensitivity to unmeasured features, and two-stage CT triage reads CT features only when the 90% CT-uncertainty interval, logit⁡(p^i)±1.645 σinput,i\operatorname{logit}(\widehat{p}_i) \pm1.645\,\sigma_{\mathrm{input},i} 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 ∣V∣|V| and directed links ∣E∣|E|. Monte Carlo propagation with S=4,000S = 4{,}000 draws redraws every input node and re-evaluates every link for each draw and then sorts the outcomes, in Θ(S(∣V∣+∣E∣)+Slog⁡S)\Theta(S(|V| + |E|) + S\log S) 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 ∣V∣|V| or ∣E∣|E|, giving O(∣V∣+∣E∣)O(|V|+|E|) 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 stageExtracted outputHardwareMean (median)Range / mode
DICOM to NIfTI conversion3D NIfTI volumeCPU2.5 s (2.2 s)Shared input
Sybil [80]3D lung-cancer CNN (runtime ref.)1× RTX A60003.2 s (3.2 s)Concurrent
lungmask [69]3D emphysema (LAA<−950< -950)1× RTX A60007.6 s (7.2 s)Concurrent
TotalSegmentator (total)Thoracic aortic calcium (≥130\ge130 HU)1× RTX A600026.8 s (25.8 s)Cumulative
TotalSegmentator (heartchambers_highres)Left ventricular chamber volume1× RTX A600030.6 s (29.4 s)Cumulative
TotalSegmentator (pulmonary_report)Pulmonary artery diameter ratio1× RTX A600045.6 s (45.7 s)39.4–51.7 s∗^{*}
CE-Agent scoring & uncertaintyCalibrated p^i\hat{p}_i, σinput,i\sigma_{\mathrm{input},i}, W90,iW_{90,i}1 CPU core1.10–1.23 msBatch (2.1–2.6 ms single)†^{\dagger}

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 4545–8484 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 (n=15,727n = 15{,}727), 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 4545–7575 years not on lipid-lowering therapy at baseline (n=1,465n = 1{,}465), with the no-CT risk recalibrated in the large to the training event rate; 178 participants with diabetes or LDL ≥190\ge190 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 m2^2 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], χ2\chi^2 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 AUCcr\mathrm{AUC}_{\mathrm{cr}} [40]. Calibration, overall accuracy, and clinical utility were summarized by competing-risk Briercr\mathrm{Brier}_{\mathrm{cr}} and Aalen–Johansen E/Ocr\mathrm{E/O}_{\mathrm{cr}}, 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 AUCcr\mathrm{AUC}_{\mathrm{cr}} 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; tt-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; W90W_{90}, 90% interval width; σinput\sigma_{\mathrm{input}}, 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/Ocr−1∣_{\mathrm{cr}} - 1| closest to zero for calibration). 95% CIs are percentile bootstrap intervals over held-out participants (CC-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 λ\lambda, 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 [ℓj,uj][\ell_j, u_j]; 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): ΔC=+0.0016\Delta C = +0.0016 (−0.0039 to +0.0072; bounds ±\pm0.01) and Δ\DeltaBrier =−3.2×10−4= -3.2 \times10^{-4} (−5.16 ×10−4\times10^{-4} to −1.30 ×10−4\times10^{-4}; bounds ±\pm0.01); both intervals lie within their bounds. †^{\dagger} The 10-yr competing-risk Brier score (Briercr_{\mathrm{cr}}) 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, CC-index and Brier come from models trained on full-follow-up labels, and AUCcr_{\mathrm{cr}} and E/Ocr_{\mathrm{cr}} 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. ‡^{\ddagger} They target composite endpoints that include nonfatal events, so Brier score and E/Ocr_{\mathrm{cr}} are not scored against cardiovascular death (matched-endpoint calibration in Table A2 and Table A6); after training-prevalence logit recalibration all three reach E/Ocr=1.02_{\mathrm{cr}} = 1.02 (Methods).

Note: Bold marks the best value in each column per endpoint (∣E/Ocr−1∣|\mathrm{E/O}_{\mathrm{cr}} - 1| 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/Ocr_{\mathrm{cr}} was evaluated on the primary continuous ln⁡(CAC+1)\ln(\mathrm{CAC} + 1) 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 (P>0.40P > 0.40). ‡The published equations target composite cardiovascular endpoints that include nonfatal events, so their Brier score and E/Ocr_{\mathrm{cr}} are reported on the matched endpoint (PCE on hard ASCVD, lower block, where published PREVENT has E/Ocr=0.87_{\mathrm{cr}} = 0.87; Framingham in Table A6) and not against cardiovascular death; after train prevalence-logit recalibration the equations reach E/Ocr=1.05_{\mathrm{cr}} = 1.05–1.06 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. tt-AUCcr_{\mathrm{cr}} / CS, integrated competing-risk and cause-specific time-dependent AUC across 1–13 yr; E/Ocr_{\mathrm{cr}}, 10-yr Aalen–Johansen competing-risk expected-to-observed ratio (E/O = 0.99 on n=14,971n = 14{,}971 uncensored 10-yr status across all CE-Agent models); CIF, competing-risk cumulative incidence function (unadjusted cause-specific Cox PH: AUCcr_{\mathrm{cr}} = 0.7192, slope 1.797, E/Ocr_{\mathrm{cr}} = 1.11, NB5%_{5\%} = 6.2); NB5%_{5\%}, net benefit (×10−3\times10^{-3}) at risk threshold pt=5%p_t = 5\% (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 (P≥0.30P \ge0.30 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/Ocr_{\mathrm{cr}} against cardiovascular death is not reported for them (1.02 after training-prevalence recalibration; Table A1), and their NB5%_{5\%} and ICI are given after recalibration.

Note: C10C_{10}, 10-yr Harrell’s CC-index; CfullC_{\mathrm{full}}, full-follow-up Harrell’s CC-index; {sup}†primary specification; {sup}§evaluated under the locked binary CAC prior (Cfull=0.805C_{\mathrm{full}} = 0.805; 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 tt-AUC spans 1–15 yr, weighted by Kaplan–Meier survival increments (competing-risk integrated tt-AUCcr=0.8186_{\mathrm{cr}} = 0.8186 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 tt-AUC not computed.

Note: Bold values denote column-best baseline performance (∣slope−1∣|\mathrm{slope} - 1| closest to zero for calibration). Abbreviations: C10C_{10}, 10-yr Harrell’s CC-index; Int. tt-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 tt-AUC of CE-Agent (hard bounds) exceeds that of PCE by 0.025 (P = 0.010). 10-yr concordance C10C_{10} 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/Ocr_{\mathrm{cr}} 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 tt-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); CfullC_{\mathrm{full}} denotes full-follow-up Harrell’s CC-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 RegimeCE-AgentUnc.Cox PHRSFGBSAXGB-CoxXGB
Finite-sample mortality discrimination (mean AUC10_{10}, 20 draws)
n=250n = 250 (6 deaths)0.71730.55490.75260.74870.71060.70550.5000
n=500n = 500 (13 deaths)0.77400.67890.78000.76670.73190.74450.7063
n=1,000n = 1{,}000 (26 deaths)0.79500.74290.78860.76720.74330.76050.7299
n=2,000n = 2{,}000 (52 deaths)0.80000.76590.79510.76740.74650.76630.7250
Full Train (n=4,821n = 4{,}821)0.80680.78860.79720.7832—0.77930.7658
Cross-group OOD AUC10_{10}, slope error (∣slope10−1∣|\mathrm{slope}_{10} - 1|), sample-size criteria, and DCA
Hispanic OOD (E=36E = 36)0.77760.71840.78350.77170.75350.79190.7326
Scanner OOD (E=34E = 34)0.78480.72380.70790.78230.73520.77690.7381
LOGO Mean (CVD Death)0.80510.80460.80540.79650.79460.82150.7827
LOGO Mean (Hard ASCVD)0.77650.78460.77730.77340.76190.77730.7672
∣slope10−1∣|\mathrm{slope}_{10} - 1| (n=250n = 250)0.7200.9730.160————
∣slope10−1∣|\mathrm{slope}_{10} - 1| (n=500n = 500)0.4180.7750.125———0.378
∣slope10−1∣|\mathrm{slope}_{10} - 1| (n=1,000n = 1{,}000)0.3440.5160.272———0.448
∣slope10−1∣|\mathrm{slope}_{10} - 1| (Hispanic)0.3300.5130.2190.0510.7270.3910.459
∣slope10−1∣|\mathrm{slope}_{10} - 1| (Scanner)0.4000.6770.6200.3970.0900.4970.546
Riley Nmin⁡N_{\min} (CVD Death)2{,}2378,3808,380Undef.Undef.Undef.Undef.
Riley Nmin⁡N_{\min} (Hard ASCVD)1{,}2023,0553,055Undef.Undef.Undef.Undef.
Excess NB (×10−4\times10^{-4}, CVD)4.673.685.374.432.673.812.62
Excess NB (×10−3\times10^{-3}, ASCVD)2.362.352.252.241.432.042.17

Table A11. MESA models by training size (n=250n = 250 to 4,821) and by racial/ethnic and scanner group: CE-Agent versus 31 baseline configurations (n=6,795n = 6,795).

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; ***P<0.001P < 0.001, two-sided paired Wilcoxon signed-rank test); Nmin⁡N_{\min}, minimum required sample size under Riley et al. [23]; Excess NB (10−310^{-3}), integrated excess net benefit (×10−3\times10^{-3}) over max(treat-all, 0) across 330 institution-threshold cells. Point-anchored CE-Agent (p=4p = 4 free slope parameters out of 12 active slopes; the thirteenth weight, on the upstream inflammation score, is fixed in every specification) requires Nmin⁡=1,110N_{\min} = 1,110, 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 AUC10_{10} 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, P<0.001P < 0.001; 19/24 demographic shifts, P<0.001P < 0.001; two-sided paired Wilcoxon signed-rank tests); Nmin⁡N_{\min}, minimum required sample size on cardiovascular mortality (peff=8.0p_{\mathrm{eff}} = 8.0 vs. 22.0 unconstrained; hard ASCVD Nmin⁡=1,202N_{\min} = 1,202 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 (×10−3\times10^{-3}) 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 K=1K = 1, 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 Nmin⁡N_{\min} across the 33 hospitals using 23–34 encoded predictor columns (Table A8). Slope rows hold the single-hospital (K=1K = 1) point-anchored specification fixed across KK; the sample-size-adaptive 1/n1/n relaxation converges to unit slope (1.00) as KK grows (Table A18). At K=1K = 1, 28/33 centers (84.8%; 29/33 at K=2K = 2) have calibration slopes within [0.80, 1.25], vs. 8/33 (24.2%) unconstrained, and 12/33 (36.4%) meet Nmin⁡N_{\min} on their own data, vs. 0/33 unconstrained. XGB-Cox (not tabulated) reaches 0.6211 at K=1K = 1 and 0.7298 at K=32K = 32.

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; EE, training events (approximate mean training deaths in parentheses for n=250…2,000n = 250 \ldots2{,}000); 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 AUC10_{10} is 0.7171, 0.7794, 0.7949, 0.8009 at n=250…2,000n = 250 \ldots2{,}000 (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 n=250n = 250, 464/620 at n=500n = 500, 510/620 at n=1,000n = 1{,}000, and 522/620 at n=2,000n = 2{,}000), 74.7%–89.7% on hard ASCVD (463/620, 556/620, and 550/620 at n=250n = 250, 500, and 1,000), and 66.9%–91.4% of cross-group OOD and LOGO comparisons. With peff=8.0p_{\mathrm{eff}} = 8.0 (CVD) and 9.0 (ASCVD) vs. 22 unconstrained, CE-Agent meets Nmin⁡N_{\min} in 3/4 racial/ethnic and 2/4 scanner cohorts on hard ASCVD. Excess net benefit: CVD death, 4.67×10−44.67 \times10^{-4} vs. 0.53×10−40.53 \times10^{-4} for ASCVD-PCE (P=0.001P = 0.001); hard ASCVD, 2.36×10−32.36 \times10^{-3} vs. 1.95×10−31.95 \times10^{-3} for PREVENT (P=0.038P = 0.038) and 1.79×10−31.79 \times10^{-3} for ASCVD-PCE (P=0.068P = 0.068).

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)Δ100\Delta_{100}n=100n = 100 WinsConfounding Audit
1. ln(TAC + 1) (R2R^2)0.2990 / 0.2990+0.004743.5% / 56–100%2/2 concordant
2. CKD-EPI eGFR (R2R^2)0.2710 / 0.2771+0.007048.0% / 63–100%DM: +0.093 → +0.087
3. ln(CAC + 1) (R2R^2)0.2429 / 0.2467+0.020475.5% / 44–99%CRP: −0.060 → +0.015
3b. CAC > 0 (AUC)0.7663 / 0.7677+0.012777.5% / 52–100%PM2.5_{2.5}: −0.035 → −0.006
4. ln(CRP + 1) (R2R^2)0.1591 / 0.1591+0.006527.5% (100% ≥ U)2/2 concordant
5. CT Liver Fat (R2R^2)0.0883 / 0.0883+0.0000100% concordant1/1 concordant
6. Ao. Diameter (R2R^2)0.0680 / 0.0680+0.007027.5% (100% ≥ U)2/2 concordant
7. Chamber Area (R2R^2)0.0296 / 0.0296+0.00084.0% (100% ≥ U)1/1 concordant
8. Systolic BP (R2R^2)0.0171 / 0.0250+0.000923.5% / 38–100%Alc.: −0.088 → −0.021
9. Diabetes (AUC)0.6331 / 0.6331+0.00675.0% (100% ≥ U)1/1 concordant
10. Respiratory Disease (AUC)0.5543 / 0.5543+0.019335.5% (100% ≥ U)1/1 concordant

Table A12. Direction-constrained prediction of 10 intermediate biomarkers from their parent variables in MESA, with a confounding audit (n=4,821n = 4{,}821 training, 1,9741{,}974 test). Constrained fitting gives no parent sign reversals, versus mean reversal rates of 10%–27% across unconstrained regression and seven ML baselines at n=100n = 100.

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. R2R^2, coefficient of determination; C, sign-constrained; U, unconstrained; Δ100\Delta_{100}, mean test gain (C − U) across 200 subsamples of size n=100n = 100 (non-negative across all 10 mediators); n=100n = 100 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; PM2.5_{2.5} slope on continuous ln(CAC + 1); 76% of alcohol-to-SBP and 82% of PM2.5_{2.5}-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; PM2.5_{2.5}, 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, PM2.5_{2.5}); 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 μim=b^m+∑pw^pmμip\mu_{im} = \hat{b}_m + \sum_p \hat{w}_{pm}\mu_{ip} in equation (1) with observed-entry gate weight expit⁡(α0)\operatorname{expit}(\alpha_0); variance-gated (var.-gated) scales μim\mu_{im} by expit⁡(α0−αvvim)/expit⁡(α0)\operatorname{expit}(\alpha_0 - \alpha_v v_{im})/\operatorname{expit}(\alpha_0), and reliability-gated (rel.-gated) scales μim\mu_{im} by expit⁡(α0−αm−αvvim)/expit⁡(α0)\operatorname{expit}(\alpha_0 - \alpha_m - \alpha_v v_{im})/\operatorname{expit}(\alpha_0). 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 ΔC=+0.0031\Delta C = +0.0031 and ΔE/O=+0.153\Delta E/O = +0.153 (95% CI +0.118 to + 0.209) and moves calibration-in-the-large (CITL) from +0.233 to +0.046 (ΔCITL=−0.187\Delta\mathrm{CITL} = -0.187, 95% CI −0.200 to − 0.174); 10-year C10C_{10} 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. †^{\dagger}Row 7 spans 12 external ML architectures across 6 imputers plus retrained no-CT (84 configurations; C10C_{10} range 0.686–0.816; Brier 0.02348–0.02579). Unattenuated CE-Agent has lower ∣E/O−1∣\lvert E/O - 1\rvert and higher CC than 68/84 (81.0%) configurations; reliability-gated and retrained no-CT CE-Agent have lower ∣slope−1∣\lvert\mathrm{slope} - 1\rvert 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 C=0.7747C = 0.7747 (C10=0.7706C_{10} = 0.7706), 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: Δj\Delta_j, marginal single-bound AUC gain; ISI_S, interaction for bound set SS from the inclusion–exclusion (Möbius) decomposition of all 242^4 bound combinations; rr, Pearson correlation; Δη\Delta\eta, log-odds difference; κ\kappa-ratio, design-matrix condition-number ratio relative to screened matrix; c^PA\hat{c}_{\mathrm{PA}} and c^unc\hat{c}_{\mathrm{unc}}, 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 (n=1,243n = 1{,}243) are zeroed in 260/260 (100%).

Note: Bold values highlight primary CE-Agent metrics. Abbreviations: Unc., unconstrained model; Δj\Delta_j, marginal single-bound AUC10\mathrm{AUC}_{10} gain; Δall\Delta_{\mathrm{all}}, gain with all bounds applied; IjkI_{jk}, pairwise interaction from the inclusion–exclusion (Möbius) decomposition of bound combinations; bin., binary CAC coding; Δη\Delta\eta, log-odds difference; κ\kappa, design-matrix condition number; Ao. diam., ascending aortic diameter; νj\nu_j, KKT dual multiplier satisfying complementary slackness νjc^j=0\nu_j\hat{c}_j = 0 (equation (B3)). *In Row 3, the Detrano et al. [37] coronary interval bound improves small-sample Hard ASCVD discrimination at n=500n = 500 (+0.0043, winning 40/50 draws). In full-sample Hard ASCVD (n=4,812n = 4{,}812, 137 test events), Stage 1 compatibility screening assigns CAC a directional sign constraint (or 1/n1/n soft bound) because Detrano et al. report a coronary-only endpoint, which differs from the coronary-plus-stroke composite (soft-bound AUC10=0.7765\mathrm{AUC}_{10} = 0.7765, Table A6). All cell discrimination values denote 10-year AUC10\mathrm{AUC}_{10}; small-sample experiments (n=250,500n = 250, 500) 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 DimensionProtocol / CohortCE-AgentComparatorsKey Finding
1. Single-Hospital OOD (K=1→32K = 1 \to32)33 single-site fits ×8 ML baselines (264 cells)246 / 264 (93.2%) wins; med. per-site AUC6_6 = 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. Δ\DeltaAUC6_6 across external sites (P<0.001P < 0.001)
2. Multi-Center ScalingK=K = 1, 2, 4, 8, 16, 32 hospitals (33 subsets/level)AUC6_6: 0.679 → 0.702 → 0.708 → 0.713 → 0.718 → 0.715Unconstrained 11-pred.: 0.653 → 0.694 → 0.704 → 0.713 → 0.718 → 0.708; XGB: 0.629 → 0.712λlit(n)\lambda_{\mathrm{lit}}(n) decays 0.283 → 0.006; slope within [0.80, 1.25] in 36% → 70% → 88% → 100% → 100% → 70%
3. Domain-Gen. & Riley Nmin⁡N_{\min}K=1K = 1 external sites vs. DRO / Anchor / updatingMed. 0.679; Riley Nmin⁡N_{\min} = 3,296 (peff=6p_{\mathrm{eff}} = 6)Anchor: 0.633; χ2\chi^2-DRO: 0.617; CVaR-DRO: 0.629; Offset LR/XGB: 0.625/0.650; Unconstr. 11-param.: 8,383; 34-col. LR: 37,1832.5–11.3× lower Nmin⁡N_{\min} (5.7× at matched C=0.679C = 0.679); LOCO NB > 0 in 27/33 sites
4. Progressive LDCT IntegrationClinical only → + LDCT nodule → + Visual emphysemaHeld-out (n=7,786n = 7,786): 0.685 → 0.700 → 0.708 (0.710 sign-only); LOCO: 0.700 → 0.726 → 0.729Held-out full LDCT: LR 0.701 (P=0.036P = 0.036), XGB 0.696, PLCOm2012 0.696; LOCO full LDCT: PLCOm2012 0.704*+0.023 held-out CT gain (P=0.036P = 0.036; nodule step +0.015, P=0.102P = 0.102; full cohort +0.009, P=0.098P = 0.098)
5. Smoking CounterfactualsQuit 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: 33/33Exact smoking monotonicity across 100% of single-site fits†
6. Missing-Data Uncertainty11 MCAR/MAR cells (0%–60%, n=37,521n = 37,521, Stage I lung graph)W^90\widehat{W}_{90} rel. MAE 0.64–0.79% (r≥0.9999r \ge0.9999); 1.21 ms/pt; gated slope 0.906–1.071MC-4000: 1.55–2.31% (93.7 ms, 77.5× slower); Plug-in slope: 0.512–0.607; Pop.-mean: 0.519–0.650ECE 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, n=37,521n = 37{,}521).

DimensionMetric / SetupCE-AgentBaselinesSummary
1. W^90\widehat{W}_{90} Interval AccuracyRel. MAE vs. 10510^{5}-draw MC (rr)2.30% [2.05–2.57%] Cond. imp. (r=0.9997r = 0.9997)MC-4000: 3.06% [2.74–3.33%]; MC-400: 9.07%; single Gaussian: 8.03%; v=0v = 0: 64.3–79.3%11/11 cells; lung graph: 0.64–0.79%
2. Predictive SD σ^p\hat{\sigma}_{p}Rel. MAE vs. 10510^{5}-draw MC2.00% [1.83–2.14%] Cond. imp.MC-4000: 5.23%; MC-400: 15.06%; v=0v = 0: 71.4–85.4% (76.5% pooled)11/11 cells (2.6× lower vs. MC-4000)
3. Single-Core LatencyBatch / single-record ms per patient0.62 ms quad. (1.23 ms end-to-end) / 1.80 ms singleMC-400: 20.8 ms; MC-4000: 203.8 ms; MC-10510^{5}: 5.0 s329× quad. speedup vs. MC-4000
4. Counterfactual MonotonicitySpurious risk rises when setting CAC/emph = 00 / 3,990 (0.0%) violationsMC-400: 197 (CAC) & 449 (emphysema) rises0/3,990 (0.0%) sign violations
5. Reliability Calibration SlopeSlope across 11 MCAR/MAR cells (0%–60%)0.959–1.047 (reliability score); ECE 0.0132 → 0.0053Untuned gate: 0.76–0.86; Pop.-mean: 0.78–0.89; Plug-in: 0.71–0.78Lowest log-loss in 11/11 cells (95% CI < 0)
6. 40% Missingness RetentionSmaller AUC loss (0% → 40%) vs. baselinesHigher retention vs. all 6 trained baselines at 40% MCAR/MARRetention gain at MCAR 40%: Cox +0.008, XGB-Cox +0.015, XGB +0.018, RSF +0.022, Tree +0.026, LR +0.1156/6 trained baselines (95% CI above zero)∗^{*}
7. Overconfident False NegativesEvents predicted below half the training event rate, per 1,000 (O/E) in top-20% σinput\sigma_{\text{input}}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 22.3/1,0002.1–9.8× lower vs. five of six baselines
8. Uncertainty-Guided TriageRetained Δ\DeltaAUC across 0%–30% review+0.011 (20%) to +0.013 (30% σinput\sigma_{\text{input}})Total width W^90\widehat{W}_{90} (ρ=0.999\rho= 0.999 with p^\hat{p}): −0.057 (20%) to −0.069 (30%)+0.068 to +0.082 vs. W^90\widehat{W}_{90} 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; rr, 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; ρ\rho, Spearman rank correlation. ∗^{*} Row 6 reports difference-in-degradation (AUC0%−AUC40%\mathrm{AUC}_{0\%} - \mathrm{AUC}_{40\%}) base − (AUC0%−AUC40%\mathrm{AUC}_{0\%} - \mathrm{AUC}_{40\%}) CE, with 95% CI >0> 0 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; rr, 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]; σmiss\sigma_{\mathrm{miss}}, missing-predictor log-odds uncertainty; ρ\rho, Spearman rank correlation. Row 9 ranks participants by the 90% predictive-interval width W^90\widehat{W}_{90} 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 λ(n)\lambda(n) 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: λ∗\lambda^*, best-performing penalty weight (lowest out-of-sample log loss; Figure A4a,Figure A4e plot the weight with the highest AUC); λlit\lambda_{\mathrm{lit}}, penalty weight set from published interval widths; Oracle gap, out-of-sample log-loss excess over λ∗\lambda^*; CV, cross-validation; cols, columns; Unreg., unregularized; Λ0,fr\Lambda_{0,\mathrm{fr}}, development-calibrated prior precision scale (31.6; closed-form Λ0=2,289.5\Lambda_0 = 2{,}289.5 in Figure A4e) under free-column N(0,1)\mathcal{N}(0,1) stabilization; gg, 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; ptp_t, 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 →\to CAC, CAC referral at intermediate risk by recalibrated PCE; Two-stage selective CT, CAC read only when the 90% CT-uncertainty interval crosses ptp_t.

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 →\to CAC referral (P=0.057P = 0.057; +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 (n=15,931n = 15{,}931; LDCT arm n=7,786n = 7{,}786) and 33-hospital out-of-center (n=37,521n = 37{,}521; LDCT arm n=18,519n = 18{,}519) benchmarks for 6-year competing-risk lung cancer incidence and 10-year lung cancer mortality, complementing Figure 8. Evaluations span single-hospital generalization (K=1→32K = 1 \to32), 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 (n=15,931n = 15{,}931) and in leave-one-center-out validation in the LDCT arm (n=18,519n = 18{,}519). Six-year competing-risk AUCcr_{\mathrm{cr}} (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/Ocr_{\mathrm{cr}}, and 10-year competing-risk lung cancer mortality AUCcr_{\mathrm{cr}} (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/Ocr_{\mathrm{cr}} − 1| closest to zero). LOCO, 33-hospital leave-one-center-out development validation (n=18,519n = 18{,}519 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, ΔAUCcr_{\mathrm{cr}} = +0.004 (95% CI −0.008 to +0.017; P=0.542P = 0.542; 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 ΔAUCcr_{\mathrm{cr}} = −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, P=0.040P = 0.040; GBSA, P<0.001P < 0.001. 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/Ocr_{\mathrm{cr}} = 0.99), whereas pooled Cox CIF (AUCcr_{\mathrm{cr}} = 0.724; slope 2.59) and LR (0.715; E/Ocr_{\mathrm{cr}} = 1.10) gain discrimination with poorer calibration. §The same 11-predictor envelope fitted to 10-yr lung cancer mortality (E/Ocr_{\mathrm{cr}} = 0.98 vs. 0.93 for published PLCOm2012; noninferior to recalibrated PLCOm2012: ΔAUCcr_{\mathrm{cr}} = −0.009, 95% CI −0.019 to +0.001, within the preregistered −0.02 margin; two-sided P=0.082P = 0.082). Refitting on the mortality label raised AUCcr_{\mathrm{cr}} by 0.011 (P<0.001P < 0.001) 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 n≈792n \approx792) of mean AUC6_6 at the other 32 hospitals; pooled external values (also rows 2–3) score those hospitals together. CE-Agent meets Riley’s Nmin⁡=3,296N_{\min} = 3,296 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 NBsum=NBlung+NBCVDNB_{\mathrm{sum}} = NB_{\mathrm{lung}} + NB_{\mathrm{CVD}} 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 n=15,931n = 15{,}931 cohort at 1.5%/5.0% (99.4% meet USPSTF 2021 eligibility by NLST enrollment design). Quadrant cells give uncensored 6-year lung cancer (n=575n = 575) and 10-year cardiovascular death (n=525n = 525) counts with inverse-probability-of-censoring weighted competing-risk shares. Two-sided paired bootstrap P<0.05P < 0.05 versus Dual CE-Agent in the same cohort; P<0.001P < 0.001; bootstrap PP values are exploratory. In the n=7,786n = 7{,}786 LDCT arm, where imaging input is read from screening CT, Dual CE-Agent itself exceeds USPSTF + PCE (recal.) by +2.6/1k (P=0.024P = 0.024). 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 GCE=(V,E,τ,ρ)\mathcal{G}_{\mathrm{CE}}=(\mathcal{V},\mathcal{E},\tau,\rho) with node set V\mathcal{V} and directed edge set E\mathcal{E}, 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 v∈Vv\in\mathcal{V} carries a timing index τ(v)∈{T0,T>0}\tau(v)\in\{T_0,T_{>0}\} distinguishing baseline T0T_0 from post-baseline T>0T_{>0} variables and a clinical role ρ(v)∈{exposure,mediator,outcome,decision,unmeasured}\rho(v)\in\{\text{exposure},\text{mediator},\text{outcome},\text{decision},\text{unmeasured}\} (decision nodes are clinical actions, such as starting a drug, and are excluded from baseline predictors). Each derived linear composite CℓC_\ell is formed from its standardized baseline parents pa⁡(Cℓ)\operatorname{pa}(C_\ell) with fixed weights wpℓ∈Rw_{p\ell}\in\mathbb{R} as Cℓ=∑p∈pa⁡(Cℓ)wpℓZpC_\ell=\sum_{p\in\operatorname{pa}(C_\ell)}w_{p\ell}Z_p. Adding γℓCℓ\gamma_\ell C_\ell for γℓ∈R\gamma_\ell\in\mathbb{R} to the baseline linear predictor c0+∑j∈VadmcjZjc_0+\sum_{j\in\mathcal{V}_{\mathrm{adm}}}c_jZ_j shifts the parent weights to c~∈Rd\tilde{c}\in\mathbb{R}^d:

c~p=cp+γℓwpℓ∀p∈pa⁡(Cℓ),c~j=cj∀j∈Vadm∖pa⁡(Cℓ).(6)\tilde{c}_p=c_p+\gamma_\ell w_{p\ell}\quad\forall p\in\operatorname{pa}(C_\ell),\qquad\tilde{c}_j=c_j\quad\forall j\in\mathcal{V}_{\mathrm{adm}}\setminus\operatorname{pa}(C_\ell). \tag*{(6)}

For missing binary intermediate biomarkers Mm∈{0,1}M_m\in\{0,1\} with training prevalence π^m\hat{\pi}_m and standard deviation s^m=π^m(1−π^m)\hat{s}_m=\sqrt{\hat{\pi}_m(1-\hat{\pi}_m)}, sign-constrained logistic regression on standardized parent moments (μip,vip)(\mu_{ip},v_{ip}) yields conditional probability qim=expit⁡(b^m+∑p∈pa⁡(Mm)w^pmμip)q_{im}=\operatorname{expit}\left(\hat{b}_m+\sum_{p\in\operatorname{pa}(M_m)}\hat{w}_{pm}\mu_{ip}\right), standardized conditional mean μim=(qim−π^m)/s^m\mu_{im}=(q_{im}-\hat{\pi}_m)/\hat{s}_m, and delta-method variance vim=[qim(1−qim)+qim2(1−qim)2∑p∈pa⁡(Mm)w^pm2vip]/s^m2v_{im}=\left[q_{im}(1-q_{im})+q_{im}^{2}(1-q_{im})^{2}\sum_{p\in\operatorname{pa}(M_m)}\hat{w}_{pm}^{2}v_{ip}\right]/\hat{s}_m^{2}, complementing continuous linear projection in equation (1). For latent mediators Lr∈LL_r\in\mathcal{L} inferred linearly from standardized upstream parents with residual variance vL=0.5v_L=0.5, the downstream edge into the outcome carries reliability gate giLr(αg)=expit⁡(α0−αm−αvviLr)g_{iL_r}(\alpha_g)=\operatorname{expit}(\alpha_0-\alpha_m-\alpha_vv_{iL_r}).

Proposition 1 (Shared-parent composite rank deficiency) *Let the standardized baseline feature matrix Z∈Rn×dZ\in\mathbb{R}^{n\times d} over Vadm\mathcal{V}_{\mathrm{adm}} have full column rank rank⁡(Z)=d\operatorname{rank}(Z)=d, and let Cℓ=∑p∈pa⁡(Cℓ)wpℓZpC_\ell=\sum_{p\in\operatorname{pa}(C_\ell)}w_{p\ell}Z_p. If pa⁡(Cℓ)⊆Vadm\operatorname{pa}(C_\ell)\subseteq\mathcal{V}_{\mathrm{adm}}, then rank⁡([Z,Cℓ])=rank⁡(Z)=d\operatorname{rank}([Z,C_\ell])=\operatorname{rank}(Z)=d, and for any diagonal weight matrix W⪰0W\succeq0 the unregularized Fisher matrix Z~TWZ~\tilde{Z}^{\mathsf{T}}W\tilde{Z} of Z~=[Z,Cℓ]∈Rn×(d+1)\tilde{Z}=[Z,C_\ell]\in\mathbb{R}^{n\times(d+1)} is singular. Moreover, for all-ones 1∈Rn\mathbf{1}\in\mathbb{R}^n, η=c01+Zc+γℓCℓ\eta=c_0\mathbf{1}+Zc+\gamma_\ell C_\ell equals η=c01+Zc~\eta=c_0\mathbf{1}+Z\tilde{c} on every participant under equation (6). Conversely, rank⁡([Z,Cℓ])=d+1\operatorname{rank}([Z,C_\ell])=d+1 if and only if ∑p∈pa⁡(Cℓ)∖VadmwpℓZp∉col⁡(Z)\sum_{p\in\operatorname{pa}(C_\ell)\setminus\mathcal{V}_{\mathrm{adm}}}w_{p\ell}Z_p\notin\operatorname{col}(Z).

Proof. When pa⁡(Cℓ)⊆Vadm\operatorname{pa}(C_\ell)\subseteq\mathcal{V}_{\mathrm{adm}}, CℓC_\ell lies in the column space col⁡(Z)\operatorname{col}(Z) of the parent columns, giving rank⁡([Z,Cℓ])=rank⁡(Z)=d\operatorname{rank}([Z,C_\ell])=\operatorname{rank}(Z)=d. Defining unull=[wℓT,−1]T∈Rd+1∖{0}u_{\mathrm{null}}=[w_\ell^{\mathsf{T}},-1]^{\mathsf{T}}\in\mathbb{R}^{d+1}\setminus\{0\}, where wℓ∈Rdw_\ell\in\mathbb{R}^d has entries wpℓw_{p\ell} for p∈pa⁡(Cℓ)p\in\operatorname{pa}(C_\ell) and 00 otherwise, gives Z~unull=0\tilde{Z}u_{\mathrm{null}}=0. Hence unullT(Z~TWZ~)unull=∥W1/2Z~unull∥22=0u_{\mathrm{null}}^{\mathsf{T}}(\tilde{Z}^{\mathsf{T}}W\tilde{Z})u_{\mathrm{null}}=\lVert W^{1/2}\tilde{Z}u_{\mathrm{null}}\rVert_2^2=0 for any diagonal weight matrix W⪰0W\succeq0, proving singularity of Z~TWZ~\tilde{Z}^{\mathsf{T}}W\tilde{Z}. Substituting Cℓ=∑p∈pa⁡(Cℓ)wpℓZpC_\ell=\sum_{p\in\operatorname{pa}(C_\ell)}w_{p\ell}Z_p into η=c01+Zc+γℓCℓ=c01+Zc~\eta=c_0\mathbf{1}+Zc+\gamma_\ell C_\ell=c_0\mathbf{1}+Z\tilde{c} gives the result by equation (6). Conversely, writing Cℓ=∑p∈pa⁡(Cℓ)∩VadmwpℓZp+∑p∈pa⁡(Cℓ)∖VadmwpℓZpC_\ell=\sum_{p\in\operatorname{pa}(C_\ell)\cap\mathcal{V}_{\mathrm{adm}}}w_{p\ell}Z_p+\sum_{p\in\operatorname{pa}(C_\ell)\setminus\mathcal{V}_{\mathrm{adm}}}w_{p\ell}Z_p shows Cℓ∉col⁡(Z)C_\ell\notin\operatorname{col}(Z), and hence rank⁡([Z,Cℓ])=d+1\operatorname{rank}([Z,C_\ell])=d+1, if and only if the sum outside Vadm\mathcal{V}_{\mathrm{adm}} lies outside col⁡(Z)\operatorname{col}(Z). ∞

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 jj undergoes a coding transformation ψj:xjnative↦xjcode\psi_j : x_j^{\mathrm{native}} \mapsto x_j^{\mathrm{code}} prior to standardization Zj=(Xj−μ^j)/σ^jZ_j = (X_j - \hat{\mu}_j)/\hat{\sigma}_j, the general conversion factor in equation (2) evaluated at baseline reference level x0x_0 is κj=σ^j/∣ψj(x0+Δxjpub)−ψj(x0)∣\kappa_j = \hat{\sigma}_j/\lvert\psi_j(x_0 + \Delta x_j^{\mathrm{pub}}) - \psi_j(x_0)\rvert, which simplifies to κj=σ^j/Δxjpub\kappa_j = \hat{\sigma}_j/\Delta x_j^{\mathrm{pub}} under identity coding ψj(x)=x\psi_j(x) = x; βjraw\beta_j^{\mathrm{raw}} is first oriented to the direction of the cohort coding. After rescaling, the four-criterion compatibility check assigns predictor jj to a directional sign constraint in Csign∖Cbox\mathcal{C}_{\mathrm{sign}} \setminus\mathcal{C}_{\mathrm{box}} with δj=sign⁡(κjβjraw)∈{−1,+1}\delta_j = \operatorname{sign}(\kappa_j\beta_j^{\mathrm{raw}}) \in\{-1,+1\} rather than an interval bound in Cbox\mathcal{C}_{\mathrm{box}} 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 Cbox\mathcal{C}_{\mathrm{box}}.

Soft bounds and zeroing of reversed weights

For finite penalty weight λ≥0\lambda\ge0 in equation (3) with Cpoint∩Cbox=∅\mathcal{C}_{\mathrm{point}} \cap\mathcal{C}_{\mathrm{box}} = \varnothing, the soft-bound estimator minimizes over (c0,c)∈Ωsign(c_0,c) \in\Omega_{\mathrm{sign}}:

Jλ,λfree(c0,c;D)=Lemp(c0,c;D)+λ∑j∈Cboxϕbox(cj;ℓj,uj)+λ∑j∈Cpoint(cj−cjpub)2+λfree2∑j∉Cbox∪Cpointcj2,(7)\begin{aligned} \mathcal{J}_{\lambda,\lambda_{\mathrm{free}}}(c_0,c;\mathcal{D})={}&\mathcal{L}_{\mathrm{emp}}(c_0,c;\mathcal{D})+\lambda\sum_{j\in\mathcal{C}_{\mathrm{box}}}\phi_{\mathrm{box}}(c_j;\ell_j,u_j)\\ &+\lambda\sum_{j\in\mathcal{C}_{\mathrm{point}}}(c_j-c_j^{\mathrm{pub}})^2+\frac{\lambda_{\mathrm{free}}}{2}\sum_{j\notin\mathcal{C}_{\mathrm{box}}\cup\mathcal{C}_{\mathrm{point}}}c_j^2, \tag*{(7)} \end{aligned}
ϕbox(cj;ℓj,uj)=max⁡(0,ℓj−cj,cj−uj)2,\phi_{\mathrm{box}}(c_j;\ell_j,u_j)=\max(0,\ell_j-c_j,c_j-u_j)^2,

subject to δjcj≥0\delta_jc_j \ge0 for all j∈Csignj \in\mathcal{C}_{\mathrm{sign}}, with λfree=0\lambda_{\mathrm{free}}=0 in primary fits and the sample-size-adaptive per-SD unit-normal prior λfree(η)=1/η\lambda_{\mathrm{free}}(\eta)=1/\eta (weighted by each gated design column’s empirical variance) on unbounded predictors j∉Cbox∪Cpointj \notin\mathcal{C}_{\mathrm{box}}\cup\mathcal{C}_{\mathrm{point}} in small-sample MESA scaling experiments. The continuously differentiable squared-hinge violation ϕbox\phi_{\mathrm{box}} vanishes throughout [ℓj,uj][\ell_j,u_j] and penalizes excursions quadratically outside the interval. Because a published 95% interval [ℓj,uj][\ell_j,u_j] corresponds to Gaussian prior standard deviation σj=(uj−ℓj)/3.92\sigma_j=(u_j-\ell_j)/3.92 and quadratic negative-log-prior coefficient (2σj2)−1(2\sigma_j^2)^{-1} on summed negative log-likelihood nLempn\mathcal{L}_{\mathrm{emp}}, the per-patient penalty weight is set from the published interval widths and scales inversely with cohort size as λlit(n)=Λ0/n\lambda_{\mathrm{lit}}(n)=\Lambda_0/n, where Λ0=median⁡j∈Cbox[(2σj2)−1]\Lambda_0=\operatorname{median}_{j\in\mathcal{C}_{\mathrm{box}}}[(2\sigma_j^2)^{-1}] equals 1,576.5 on NLST cardiovascular mortality and 224.5 on NLST lung cancer. Taking λ→∞\lambda\to\infty (with λfree=0\lambda_{\mathrm{free}}=0) enforces cj∈[ℓj,uj]c_j\in[\ell_j,u_j] for all j∈Cboxj\in\mathcal{C}_{\mathrm{box}} and cj=cjpubc_j=c_j^{\mathrm{pub}} for j∈Cpointj\in\mathcal{C}_{\mathrm{point}}, recovering the hard-bound limit over ΩCE\Omega_{\mathrm{CE}}. For each sign-only predictor j∈Csign∖(Cbox∪Cpoint)j\in\mathcal{C}_{\mathrm{sign}}\setminus(\mathcal{C}_{\mathrm{box}}\cup\mathcal{C}_{\mathrm{point}}) with dual multiplier νj≥0\nu_j\geq0, first-order optimality over Ωsign\Omega_{\mathrm{sign}} at λfree=0\lambda_{\mathrm{free}}=0 satisfies the KKT conditions:

∇cjLemp(c^0,c^;D)−δjνj=0,δjc^j≥0,νj(δjc^j)=0.(8)\nabla_{c_j}\mathcal{L}_{\mathrm{emp}}(\hat{c}_0,\hat{c};D)-\delta_j\nu_j=0,\qquad\delta_j\hat{c}_j\geq0,\qquad\nu_j(\delta_j\hat{c}_j)=0. \tag*{(8)}

Proposition 2 (Zeroing of sign-reversed weights) Because Lemp(c0,c;D)\mathcal{L}_{\mathrm{emp}}(c_0,c;D) and ϕbox\phi_{\mathrm{box}} are convex and continuously differentiable on R1+d\mathbb{R}^{1+d} and both ΩCE\Omega_{\mathrm{CE}} and Ωsign\Omega_{\mathrm{sign}} are non-empty closed convex polyhedra, any local minimizer (c^0,c^)(\hat{c}_0,\hat{c}) of equations (3) and (B2) for λfree≥0\lambda_{\mathrm{free}}\geq0, including the primary hard-bound limit (λ→∞\lambda\to\infty, λfree=0\lambda_{\mathrm{free}}=0), is globally optimal and zeroes sign-reversed coordinates j∈Csign∖(Cbox∪Cpoint)j\in\mathcal{C}_{\mathrm{sign}}\setminus(\mathcal{C}_{\mathrm{box}}\cup\mathcal{C}_{\mathrm{point}}): whenever confounding or competing mortality reverses the empirical gradient against the published causal sign so that δj∇cjLemp(c^0,c^;D)>0\delta_j\nabla_{c_j}\mathcal{L}_{\mathrm{emp}}(\hat{c}_0,\hat{c};D)>0, complementary slackness forces c^j=0\hat{c}_j=0 (whereas an interval-bounded predictor j∈Cboxj\in\mathcal{C}_{\mathrm{box}} with ℓj≤uj\ell_j\leq u_j and min⁡(δjℓj,δjuj)>0\min(\delta_j\ell_j,\delta_ju_j)>0 clamps at its literature boundary nearer zero as λ→∞\lambda\to\infty).

Proof. Convexity of Lemp\mathcal{L}_{\mathrm{emp}}, ϕbox\phi_{\mathrm{box}}, and λfree2cj2\frac{\lambda_{\mathrm{free}}}{2}c_j^2 together with linear constraints δjcj≥0\delta_jc_j\geq0 makes the KKT conditions necessary and sufficient for global optimality. On any sign-only coordinate j∈Csign∖(Cbox∪Cpoint)j\in\mathcal{C}_{\mathrm{sign}}\setminus(\mathcal{C}_{\mathrm{box}}\cup\mathcal{C}_{\mathrm{point}}), stationarity gives ∇cjLemp(c^0,c^;D)+λfreec^j−δjνj=0\nabla_{c_j}\mathcal{L}_{\mathrm{emp}}(\hat{c}_0,\hat{c};D)+\lambda_{\mathrm{free}}\hat{c}_j-\delta_j\nu_j=0 (reducing directly to equation (B3) when λfree=0\lambda_{\mathrm{free}}=0). Because δj2=1\delta_j^2=1 and λfree(δjc^j)≥0\lambda_{\mathrm{free}}(\delta_j\hat{c}_j)\geq0, multiplying by δj\delta_j yields νj=δj∇cjLemp(c^0,c^;D)+λfree(δjc^j)≥δj∇cjLemp(c^0,c^;D)\nu_j=\delta_j\nabla_{c_j}\mathcal{L}_{\mathrm{emp}}(\hat{c}_0,\hat{c};D)+\lambda_{\mathrm{free}}(\delta_j\hat{c}_j)\geq\delta_j\nabla_{c_j}\mathcal{L}_{\mathrm{emp}}(\hat{c}_0,\hat{c};D). Whenever δj∇cjLemp(c^0,c^;D)>0\delta_j\nabla_{c_j}\mathcal{L}_{\mathrm{emp}}(\hat{c}_0,\hat{c};D)>0, positivity νj>0\nu_j>0 and νj(δjc^j)=0\nu_j(\delta_j\hat{c}_j)=0 force c^j=0\hat{c}_j=0. □

Quadrature and variance decomposition

In analytic uncertainty propagation (Methods), each observed-predictor node j∈S=Vadmj\in\mathcal{S}=\mathcal{V}_{\mathrm{adm}} carries moments (Zij,0,0)(Z_{ij},0,0) when observed, (0,1,1)(0,1,1) under population-mean fill (with imputed binary predictors carrying standardized Bernoulli variance clipped at π^ij(1−π^ij)/σ^j2\hat{\pi}_{ij}(1-\hat{\pi}_{ij})/\hat{\sigma}_j^2, where π^ij\hat{\pi}_{ij} is the imputed probability that predictor jj equals 1), or (μ^ijimp,v^ijimp,1)(\hat{\mu}_{ij}^{\mathrm{imp}},\hat{v}_{ij}^{\mathrm{imp}},1) under out-of-fold conditional Gaussian imputation. Each latent mediator L∈LL\in\mathcal{L} carries standardized residual variance vL=0.5v_L=0.5 with miL=1m_{iL}=1. Active directed edges from node jj to direct child k∈ch⁡act(j)k\in\operatorname{ch}_{\mathrm{act}}(j) comprise the outcome YY and any mediator children that are latent or missing for record ii, so observed intermediate biomarkers block indirect upstream paths without double-counting. Each active edge carries the reliability-gated weight ηijk(θ)=gij(αg)sjk(θ)c^jk\eta_{ijk}(\theta)=g_{ij}(\alpha_g)s_{jk}(\theta)\hat{c}_{jk}, where θ\theta collects the gate parameters αg\alpha_g and the edge-scale multipliers. The base edge weight c^jk\hat{c}_{jk} is c^j\hat{c}_j 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 sjk(θ)s_{jk}(\theta) collects unit-mean log-normal multipliers exp⁡(σlog⁡u−12σlog⁡2)\exp(\sigma_{\log}u-\frac{1}{2}\sigma_{\log}^2) (u∼N(0,1)u\sim\mathcal{N}(0,1)) with σlog⁡=0.35\sigma_{\log}=0.35 on the global outcome scale sYs_Y and the latent-edge scale and σlog⁡=0.30\sigma_{\log}=0.30 on the plausible-to-established weight ratio. Out-of-fold fitted gate parameters αg=(α0,αm,αv)\alpha_g=(\alpha_0,\alpha_m,\alpha_v) vary around the fitted center αˉg\bar{\alpha}_g (initialized at (4.0,3.8,1.6)(4.0, 3.8, 1.6)) with normal standard deviations (0.5,0.5,0.3)(0.5, 0.5, 0.3) and (αm,αv)(\alpha_m,\alpha_v) clipped at zero. Reverse-topological recursion from TiY=1T_{iY}=1 via Tij(θ)=∑k∈ch⁡act(j)γijk(θ)Tik(θ)T_{ij}(\theta)=\sum_{k\in\operatorname{ch}_{\mathrm{act}}(j)}\gamma_{ijk}(\theta)T_{ik}(\theta) yields Mi(θ)=∑j∈STij(θ)μijM_i(\theta)=\sum_{j\in S}T_{ij}(\theta)\mu_{ij} and Vi(θ)=∑j∈STij(θ)2vij+∑L∈LTiL(θ)2vLV_i(\theta)=\sum_{j\in S}T_{ij}(\theta)^2v_{ij}+\sum_{L\in\mathcal{L}}T_{iL}(\theta)^2v_L.

At each of the R=33=27R=3^3=27 Gauss–Hermite nodes over the three gate parameters αg\alpha_g (weights ωr\omega_r), the 3×33\times3 nodes over the plausible-to-established weight ratio and the latent-edge scale are moment-matched into a single Gaussian component (Mir,Vir)(M_{ir},V_{ir}) on the logit scale. Incorporating training-prevalence intercept variance σb2=(nyˉ(1−yˉ))−1\sigma_b^2=(n\bar{y}(1-\bar{y}))^{-1}, marginal risk p^i\hat{p}_i integrates sYs_Y over a 16-node Gauss–Hermite rule {(sY,q,ωq)}q=116\{(s_{Y,q},\omega_q)\}_{q=1}^{16} (with 16 inner Gauss–Hermite nodes for EξE_\xi), while the log-odds mixture cumulative distribution function Fi(t)F_i(t) integrates sYs_Y over a 64-point uniform normal-density grid {(s~Y,q′,ω~q′)}q′=164\{(\tilde{s}_{Y,q'},\tilde{\omega}_{q'})\}_{q'=1}^{64} on [−6,6][-6,6] to resolve steep tail transitions:

p^i=∑r=127∑q=116ωrωqEξ[expit⁡(c^0+sY,qMir+(σb2+sY,q2Vir)1/2ξ)],Fi(t)=∑r=127∑q′=164ωrω~q′Φ(t−c^0−s~Y,q′Mir(σb2+s~Y,q′2Vir)1/2),(9)\begin{aligned} \hat{p}_i&=\sum_{r=1}^{27}\sum_{q=1}^{16}\omega_r\omega_q E_\xi\left[\operatorname{expit}\left(\hat{c}_0+s_{Y,q}M_{ir}+\left(\sigma_b^2+s_{Y,q}^2V_{ir}\right)^{1/2}\xi\right)\right],\\ F_i(t)&=\sum_{r=1}^{27}\sum_{q'=1}^{64}\omega_r\tilde{\omega}_{q'}\Phi\left(\frac{t-\hat{c}_0-\tilde{s}_{Y,q'}M_{ir}}{\left(\sigma_b^2+\tilde{s}_{Y,q'}^2V_{ir}\right)^{1/2}}\right), \tag*{(9)} \end{aligned}

where Φ(⋅)\Phi(\cdot) denotes the standard normal cumulative distribution function. Safeguarded Newton–bisection (60 maximum iterations) on logit tt solves Fi(t0.05,i)=0.05F_i(t_{0.05,i})=0.05 and Fi(t0.95,i)=0.95F_i(t_{0.95,i})=0.95 to obtain monotone risk quantiles [q^0.05,i,q^0.95,i]=[expit⁡(t0.05,i),expit⁡(t0.95,i)][\hat{q}_{0.05,i},\hat{q}_{0.95,i}]=[\operatorname{expit}(t_{0.05,i}),\operatorname{expit}(t_{0.95,i})].

By the law of total variance, the marginal log-odds variance Var⁡(ηi)=SD⁡(ηi)2\operatorname{Var}(\eta_i)=\operatorname{SD}(\eta_i)^2 decomposes into baseline, input, and parameter channels:

Var⁡(ηi)=σb2⏟baseline+E[sY2]Eθ[Vi(θ)]⏟input+E[sY2]Var⁡θ(Mi(θ))+Var⁡(sY)Eθ[Mi(θ)]2⏟parameter.(10)\operatorname{Var}(\eta_i)=\underbrace{\sigma_b^2}_{\text{baseline}}+\underbrace{\mathbb{E}[s_Y^2]\mathbb{E}_{\theta}[V_i(\theta)]}_{\text{input}}+\underbrace{\mathbb{E}[s_Y^2]\operatorname{Var}_{\theta}(M_i(\theta))+\operatorname{Var}(s_Y)\mathbb{E}_{\theta}[M_i(\theta)]^2}_{\text{parameter}}. \tag*{(10)}

With standard normal 95th percentile z0.95=Φ−1(0.95)=1.645z_{0.95}=\Phi^{-1}(0.95)=1.645, a first-order Taylor expansion of expit⁡(ηi)\operatorname{expit}(\eta_i) around the predictive mean yields W^90,i≈2z0.95SD⁡(ηi)p^i(1−p^i)=3.29SD⁡(ηi)p^i(1−p^i)\widehat{W}_{90,i}\approx2z_{0.95}\operatorname{SD}(\eta_i)\hat{p}_i(1-\hat{p}_i)=3.29\operatorname{SD}(\eta_i)\hat{p}_i(1-\hat{p}_i) in equation (5). At the central parameter vector θˉ\bar{\theta}, σinput,i2=Vi(θˉ)=(σinputcomplete)2+σmiss,i2\sigma_{\mathrm{input},i}^2=V_i(\bar{\theta})=(\sigma_{\mathrm{input}}^{\mathrm{complete}})^2+\sigma_{\mathrm{miss},i}^2, decomposing total input variance into the complete-cohort baseline σinputcomplete\sigma_{\mathrm{input}}^{\mathrm{complete}} 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 σmiss,i2=∑j∈SmeasTij(θˉ)2vij\sigma_{\mathrm{miss},i}^2=\sum_{j\in S_{\mathrm{meas}}}T_{ij}(\bar{\theta})^2v_{ij} for cohort-measured predictors Smeas⊆SS_{\mathrm{meas}}\subseteq S. Per patient on a graph G=(V,E)\mathcal{G}=(\mathcal{V},\mathcal{E}), assembling the node moments visits ∣V∣|\mathcal{V}| nodes and ∣E∣|\mathcal{E}| links once, and vectorized reverse-topological sensitivity accumulation performs Kgrid=33×32=243K_{\mathrm{grid}}=3^3\times3^2=243 operations per link and node in O(Kgrid(∣V∣+∣E∣))O(K_{\mathrm{grid}}(|\mathcal{V}|+|\mathcal{E}|)) time. Subsequent scalar logit-mixture moment integration (27×16×1627\times16\times16 evaluations) and safeguarded Newton–bisection root-finding (≤60\leq60 iterations over 27×6427\times64 mixture components) perform a fixed number of operations, O(1)O(1), independent of ∣V∣|\mathcal{V}| and ∣E∣|\mathcal{E}|.

SS-draw Monte Carlo simulation instead takes Θ(S(∣V∣+∣E∣)+Slog⁡S)\Theta\left(S(|V|+|E|)+S\log S\right) time. Figure 7 benchmarks both across ∣V∣∈{22,50,100,200,350,500,750,1,000}|V|\in\{22,50,100,200,350,500,750,1{,}000\} with 8.5% latent mediators, 30% binary input nodes, 25% missingness, and ∣E∣=round⁡(1.8∣V∣)|E|=\operatorname{round}(1.8|V|) 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 (c^0,c^)(\hat{c}_0,\hat{c}) with logistic Hessian weights W=diag⁡(p^i(1−p^i))W=\operatorname{diag}(\hat{p}_i(1-\hat{p}_i)) and active prior curvature Λprior⪰0\Lambda_{\mathrm{prior}}\succeq0, effective degrees of freedom compress from dd to peff=tr⁡((ZTWZ+Λprior)−1ZTWZ)≤dp_{\mathrm{eff}}=\operatorname{tr}\left((Z^{\mathsf T}WZ+\Lambda_{\mathrm{prior}})^{-1}Z^{\mathsf T}WZ\right)\le d (equaling the number of fitted parameters, p=4p=4, in the point-anchored single-hospital NLST cardiovascular specification and median peff=6p_{\mathrm{eff}}=6 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 peffp_{\mathrm{eff}}, prevalence yˉ\bar{y}, Cox–Snell RCS2R^2_{\mathrm{CS}}, and maximum RCS,max2=1−(yˉyˉ(1−yˉ)1−yˉ)2R^2_{\mathrm{CS,max}}=1-\left(\bar{y}^{\bar{y}}(1-\bar{y})^{1-\bar{y}}\right)^2:

Nmin⁡=max⁡(⌈NC1⌉,⌈NC2⌉,⌈NC3⌉),NC1=peff(S−1)ln⁡(1−RCS2/S)N_{\min}=\max\left(\lceil N_{\mathrm{C1}}\rceil,\lceil N_{\mathrm{C2}}\rceil,\lceil N_{\mathrm{C3}}\rceil\right),\qquad N_{\mathrm{C1}}=\frac{p_{\mathrm{eff}}}{(S-1)\ln\left(1-R^2_{\mathrm{CS}}/S\right)}
NC2=peff(SC2−1)ln⁡(1−RCS2/SC2),SC2=RCS2RCS2+δRCS,max2(11)N_{\mathrm{C2}}=\frac{p_{\mathrm{eff}}}{(S_{\mathrm{C2}}-1)\ln\left(1-R^2_{\mathrm{CS}}/S_{\mathrm{C2}}\right)},\qquad S_{\mathrm{C2}}=\frac{R^2_{\mathrm{CS}}}{R^2_{\mathrm{CS}}+\delta R^2_{\mathrm{CS,max}}} \tag*{(11)}
NC3=(1.960.05)2yˉ(1−yˉ).N_{\mathrm{C3}}=\left(\frac{1.96}{0.05}\right)^2\bar{y}(1-\bar{y}).

Here NC1N_{\mathrm{C1}} enforces global shrinkage S≥0.90S\ge0.90, NC2N_{\mathrm{C2}} limits expected Nagelkerke R2R^2 optimism to δ=0.05\delta=0.05, and NC3N_{\mathrm{C3}} bounds marginal outcome prevalence error within ±0.05\pm0.05. Criterion C1 binds across cohorts, so Nmin⁡=⌈NC1⌉N_{\min}=\lceil N_{\mathrm{C1}}\rceil rises with peffp_{\mathrm{eff}} and falls with the anticipated RCS2R^2_{\mathrm{CS}} of each specification, converted from its median held-out AUC at the pooled training prevalence [67]. In NLST cardiovascular mortality (yˉ=4.77%\bar{y}=4.77\%), p=4p=4 yields Nmin⁡=1,110N_{\min}=1{,}110 versus 6,043 unconstrained (p=12p=12), 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 RCS2R^2_{\mathrm{CS}} would require 3,329). In NLST 6-year lung cancer (yˉ=4.00%\bar{y}=4.00\% on uncensored single-hospital records), peff=6p_{\mathrm{eff}}=6 yields Nmin⁡=3,296N_{\min}=3{,}296 versus 8,383 unconstrained (11 parameters) and 37,183 (34 expanded columns). In MESA, peff=8.0p_{\mathrm{eff}}=8.0 versus 22.0 yields Nmin⁡=2,237N_{\min}=2{,}237 versus 8,380 on cardiovascular mortality and peff=9.0p_{\mathrm{eff}}=9.0 versus 22.0 yields Nmin⁡=1,202N_{\min}=1{,}202 versus 3,055 on hard ASCVD.

Experimental protocols

Standardization, optimization, and decision curves

In MESA, predictors are standardized using empirical training-fold moments (μ^j,σ^j)(\hat{\mu}_j,\hat{\sigma}_j), whereas NLST predictors use pre-specified cohort reference centers and scales, and each training entry is multiplied by its reliability gate gijg_{ij} (equation (4)), which downweights missing values. Constrained estimation in equation (3) minimizes the summed objective nJλ,λfreen\mathcal{J}_{\lambda,\lambda_{\mathrm{free}}} via bound-constrained limited-memory Broyden–Fletcher–Goldfarb–Shanno (L-BFGS-B) optimization [82] with projected gradient tolerance 10−1210^{-12}, function tolerance 0, and λfree=0\lambda_{\mathrm{free}} = 0 in primary fits (with active-set projected Newton polishing to 10−810^{-8} whenever the summed projected gradient exceeds 10−210^{-2}, certified by summed gradient ≤10−2\le10^{-2} or half-squared Newton decrement ≤10−6\le10^{-6}). Under soft bounds, a sample-size-adaptive selector applies λlit(n)=Λ0/n\lambda_{\mathrm{lit}}(n) = \Lambda_0/n below the Riley Nmin⁡N_{\min} (at anticipated C=0.75C = 0.75) or 10-events-per-free-parameter threshold and five-fold inner cross-validation across 13 grid points (00, 10−610^{-6}–10410^{4}, and ∞\infty) 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 αˉg\bar{\alpha}_g 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 NB⁡(pt)\operatorname{NB}(p_t) from true-positive and false-positive counts TP⁡(pt)\operatorname{TP}(p_t) and FP⁡(pt)\operatorname{FP}(p_t) and excess net benefit ΔNB⁡(pt)\Delta\operatorname{NB}(p_t) over treat-all NB⁡all(pt)\operatorname{NB}_{\mathrm{all}}(p_t) and treat-none at threshold pt∈(0,1)p_t \in(0,1) are

NB⁡(pt)=TP⁡(pt)n−FP⁡(pt)npt1−pt,ΔNB⁡(pt)=NB⁡(pt)−max⁡(NB⁡all(pt),0).(12)\operatorname{NB}(p_t) = \frac{\operatorname{TP}(p_t)}{n} - \frac{\operatorname{FP}(p_t)}{n}\frac{p_t}{1-p_t}, \qquad\Delta\operatorname{NB}(p_t) = \operatorname{NB}(p_t) - \max\left(\operatorname{NB}_{\mathrm{all}}(p_t),0\right). \tag*{(12)}

Across 33 NLST hospitals, excess net benefit ΔNB⁡c→c′(pt)\Delta\operatorname{NB}_{c\to c'}(p_t) from training center cc to external center c′≠cc' \ne c is averaged across all 32 external centers; the thresholds at which this average is positive for every training center, T33={pt:min⁡cΔNB⁡‾c(pt)>0}\mathcal{T}_{33} = \{p_t : \min_c \overline{\Delta\operatorname{NB}}_c(p_t) > 0\}, 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 r∈{5%,10%,20%,40%,60%}r \in\{5\%, 10\%, 20\%, 40\%, 60\%\} and under BMI-dependent missing-at-random (MAR) mechanisms with probability min⁡(2rui,1)\min(2ru_i, 1) for BMI percentile ui∈[0,1]u_i \in[0,1] (mean rate rr 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 ln⁡(CRP+1)\ln(\mathrm{CRP}+1) 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 g∈{0,1,2,4,10,100,1,000}g \in\{0, 1, 2, 4, 10, 100, 1{,}000\}, χ2\chi^2-divergence DRO, which needs no hospital or subgroup labels, across divergence radii ρ∈{0,0.1,0.25,0.5,1,2,5}\rho\in\{0, 0.1, 0.25, 0.5, 1, 2, 5\} and CVaR-DRO across tail fractions α∈{1,0.9,0.75,0.5,0.25,0.1,0.05}\alpha\in\{1, 0.9, 0.75, 0.5, 0.25, 0.1, 0.05\} [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 σj=(uj−ℓj)/3.92\sigma_j = (u_j - \ell_j)/3.92. 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 −5 kg/m2-5\ \mathrm{kg}/\mathrm{m}^2 BMI in NLST (operating through the upstream BMI-to-latent-inflammation pathway in the point-anchored single-hospital model), and −1 mmol/L-1\ \mathrm{mmol}/\mathrm{L} LDL, −10 mmHg-10\ \mathrm{mmHg} 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 ≥190 mg/dL\ge190\ \mathrm{mg}/\mathrm{dL} (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 pt=5.0%p_t = 5.0\% in NLST [48, 49] and against 2018 ACC/AHA intermediate-risk (PCE 7.5%–20%) CAC referral rules at pt=7.5%p_t = 7.5\% in MESA [17].

Lung cancer and joint triage protocols

In NLST lung cancer estimation (d=11d = 11), 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; Λ0=224.5\Lambda_0 = 224.5), four sign-only predictors (smoking intensity/duration ≥0\ge0, baseline T0T_0 radiologist non-calcified nodule ≥4\ge4 mm read ≥0\ge0, occupational exposure ≥0\ge0, and educational attainment ≤0\le0), 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 T0T_0 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 n=15,803n = 15{,}803 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 (n=15,931n = 15{,}931) 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 (n=7,786n = 7{,}786; 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 ∣logit⁡(p^i(1))−logit⁡(pt)∣≤1.645 σinput,i|\operatorname{logit}(\hat{p}^{(1)}_i) - \operatorname{logit}(p_t)| \le1.645\,\sigma_{\mathrm{input},i} on either endpoint, where p^i(1)\hat{p}^{(1)}_i is the Tier 1 risk and σinput,i\sigma_{\mathrm{input},i} 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 PP 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 (≥130\ge130 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, n=1,243n = 1{,}243). Emphysema was quantified within automated 3D lung masks (lungmask [69]). Only baseline (T0T_0) 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<−950< -950 HU [84]) yields 0.663 held-out (n=2,773n = 2{,}773) and 0.707 leave-one-center-out (LOCO; n=6,573n = 6{,}573) AUCcr_{\mathrm{cr}}. 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 (≤2.5×10−3\le2.5 \times10^{-3} 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 ≤42\le42 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<−950< -950, low-attenuation lung area fraction below −950-950 HU.

DimensionMetric / SetupCE-AgentBaselinesSummary
1. W^90\widehat{W}_{90} Interval AccuracyRel. MAE vs. 10510^{5}-draw MC (rr)0.50% med. (0.52% mean) [0.43–0.64%] (r≥0.99998r \ge0.99998)MC-4000: 2.13–2.90%; MC-400: 5.9–10.1%; v=0v = 0: 69.3%11/11 cells (4.8× lower vs. MC-4000)
2. Predictive SD σ^p\hat{\sigma}_{p}Rel. MAE vs. 10510^{5}-draw MC1.28% 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 LatencyQuadrature / end-to-end ms per patient0.72 ms quad. / 1.13 ms end-to-endMC-400: 18.9 ms; MC-4000: 184.8 ms257× quad. (164× end-to-end) vs. MC-4000
4. Counterfactual MonotonicitySpurious risk rises across 17 non-zero edges0 / 2,720 (0.0%) violationsMonte Carlo sampling produces rank reversals0/2,720 (0.0%) sign violations
5. Reliability Calibration SlopeSlope 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: 0.45–0.5111/11 cells within [0.88, 1.17]
6. 40% Missingness RetentionSmaller AUC10_{10} loss (0% → 40%) vs. baselinesHigher retention vs. all 6 main baselines; on par or better vs. 40/40 under MARGain (MCAR/MAR): RSF 0.021/0.044; Cox 0.103/0.051; LR 0.235/0.191Point 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 NegativesEvents predicted below half the training event rate, per 1,0002.07–3.10 / 1,000 (CVD); 1.7–9.6 / 1,000 (ASCVD)Untuned: 8.8–19.7 (CVD), 24.3–50.3 (ASCVD); Ext.: ≤ 24.3/72.42.5–11.7× lower at 40% MCAR
8. Uncertainty-Guided TriageRetained Δ\DeltaAUC10_{10} across 0%–30% review+0.061 (30% σinput\sigma_{\mathrm{input}}); +0.083 / +0.104 (USPSTF σmiss\sigma_{\mathrm{miss}})Triage by W^90\widehat{W}_{90} (ρ≥0.990\rho\ge0.990): −0.046 (15%), −0.036 (20%), −0.004 (30%)0.8% flagged at 0% missingness
9. Targeted CT TriggeringCT budget 0%–100% (Stage I graph)30% CT budget by W^90\widehat{W}_{90} or distance to 7.5% risk → AUC10_{10} = 0.794–0.795No CT: 0.776; Random 30% CT: 0.777 (needs 70% for 0.794)70% CT scans spared (Δ\DeltaAUC10_{10} = 0.000)

Table A17. Uncertainty propagation with missing data in MESA: CE-Agent versus 40 model-and-imputer configurations (n=1,974n = 1,974).

References

  1. [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. [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. [3]US Preventive Services Task Force. Screening for lung cancer: US Preventive Services Task Force recommendation statement. JAMA 325, 962–970 (2021).
  4. [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. [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. [6]Tammemägi, M. C. et al. Selection criteria for lung-cancer screening. New England Journal of Medicine 368, 728–736 (2013).
  7. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [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. [19]Khan, S. S. et al. Development and validation of the American Heart Association’s PREVENT equations. Circulation 149, 430–449 (2024).
  20. [20]Niu, C. et al. Medical multimodal multitask foundation model for lung cancer screening. Nature Communications 16, 1523 (2025).DOI
  21. [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. [22]Demircan, K. et al. Data-driven prioritization of high-risk individuals for weight loss interventions. Nature Medicine 32, 2117–2127 (2026).DOI
  23. [23]Riley, R. D. et al. Calculating the sample size required for developing a clinical prediction model. BMJ 368, m441 (2020).DOI
  24. [24]Jacobs, D. R., Jr. et al. Report of the Conference on Low Blood Cholesterol: mortality associations. Circulation 86, 1046–1060 (1992).
  25. [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. [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. [27]Pearl, J. Causality: Models, Reasoning, and Inference 2nd edn (Cambridge University Press, 2009).
  28. [28]Peters, J., Janzing, D. & Schölkopf, B. Elements of Causal Inference: Foundations and Learning Algorithms (MIT Press, 2017).
  29. [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. [30]Riley, R. D. et al. Uncertainty of risk estimates from clinical prediction models: rationale, challenges, and approaches. BMJ 388, e080749 (2025).DOI
  31. [31]Pauker, S. G. & Kassirer, J. P. The threshold approach to clinical decision making. New England Journal of Medicine 302, 1109–1117 (1980).
  32. [32]Geifman, Y. & El-Yaniv, R. Selective classification for deep neural networks. Advances in Neural Information Processing Systems 30, 4878–4887 (2017).
  33. [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. [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. [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. [36]Bild, D. E. et al. Multi-Ethnic Study of Atherosclerosis: objectives and design. American Journal of Epidemiology 156, 871–881 (2002).
  37. [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. [38]Hoerl, A. E. & Kennard, R. W. Ridge regression: biased estimation for nonorthogonal problems. Technometrics 12, 55–67 (1970).
  39. [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. [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. [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. [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. [43]Duchi, J. C. & Namkoong, H. Learning models with uniform performance via distributionally robust optimization. The Annals of Statistics 49, 1378–1406 (2021).
  44. [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. [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. [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. [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. [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. [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. [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. [51]Tibshirani, R. Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society: Series B (Methodological) 58, 267–288 (1996).
  52. [52]Wasserthal, J. et al. TotalSegmentator: robust segmentation of 104 anatomic structures in CT images. Radiology: Artificial Intelligence 5, e230024 (2023).DOI
  53. [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. [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. [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. [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. [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. [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. [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. [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. [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. [62]Navarese, E. P. et al. Causal effects of inflammation on long-term mortality: a Mendelian randomization study. Aging 18, 45–59 (2026).
  63. [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. [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. [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. [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. [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. [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. [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. [70]Levey, A. S. et al. A new equation to estimate glomerular filtration rate. Annals of Internal Medicine 150, 604–612 (2009).
  71. [71]Cox, D. R. Regression models and life-tables. Journal of the Royal Statistical Society: Series B (Methodological) 34, 187–202 (1972).
  72. [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. [73]Friedman, J. H. Greedy function approximation: a gradient boosting machine. The Annals of Statistics 29, 1189–1232 (2001).DOI
  74. [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. [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. [76]Breiman, L. Random forests. Machine Learning 45, 5–32 (2001).
  77. [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. [78]Brier, G. W. Verification of forecasts expressed in terms of probability. Monthly Weather Review 78, 1–3 (1950).
  79. [79]Kaplan, E. L. & Meier, P. Nonparametric estimation from incomplete observations. Journal of the American Statistical Association 53, 457–481 (1958).
  80. [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. [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. [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. [83]Nelder, J. A. & Mead, R. A simplex method for function minimization. The Computer Journal 7, 308–313 (1965).
  84. [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).

Paper details

vlnc:W_x97svkjj
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

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

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents