Causally Constrained and Uncertainty-Aware Cardiopulmonary Risk Prediction
Organizations: Department of Computer Science, Emory University, Atlanta, GA, USA. · Department of Computer Science, Johns Hopkins University, Baltimore, MD, USA. · Department of Radiation and Cellular Oncology, The University of Chicago, Chicago, IL, USA.
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.
Figures & tables
Appendix figures & tables26 assets
Supplementary material from the paper’s appendix.
Appendix
| Model / Specification | (95% CI) | -index (95% CI) | Brier | |
|---|---|---|---|---|
| CE-Agent (compiled hard bounds, primary) | 0.723 (0.700–0.744) | 0.731 (0.713–0.749) | 0.0433 | 1.02 |
| CE-Agent (compiled soft bounds, CV ) | 0.723 (0.700–0.744) | 0.731 (0.713–0.749) | 0.0433 | 1.02 |
| CE-Agent (point-anchored bounds) ∗ | 0.721 (0.700–0.742) | 0.729 (0.711–0.747) | 0.0436 | 1.02 |
| CE-Agent (prior-locked) | 0.717 (0.696–0.738) | 0.725 (0.707–0.744) | 0.0435 | 1.02 |
| Logistic Regression | 0.725 (0.702–0.747) | 0.734 (0.713–0.749) | 0.0433 | 1.06 |
| XGBoost Classifier | 0.724 (0.702–0.745) | 0.735 (0.717–0.751) | 0.0434 | 1.04 |
| Model / Specification | (95% CI) | -index (95% CI) | Brier | |
|---|---|---|---|---|
| Cardiovascular mortality ( ; and total deaths) | ||||
| CE-Agent (hard, , primary) | 0.807 (0.758–0.855) | 0.805 (0.773–0.835) | 0.0239 | 1.08 |
| CE-Agent (soft, ) | 0.805 (0.752–0.854) | 0.807 (0.776–0.838) | 0.0240 | 1.08 |
| CE-Agent (hard, binary CAC) | 0.811 (0.763–0.858) | 0.804 (0.772–0.833) | 0.0237 | — † |
| CE-Agent (prior-locked) | 0.794 (0.738–0.847) | 0.801 (0.767–0.832) | 0.0255 | 1.09 |
| Unconstrained Model | 0.789 (0.729–0.842) | 0.795 (0.761–0.829) | 0.0245 | 1.08 |
| Model / Specification | (95% CI) | / CS | Slope ( ) | ( ) |
| CE-Agent (hard, primary) | 0.7225 (0.700–0.744) | 0.7212 / 0.7386 | 0.951 ( 1.02 ) | 6.8 |
| CE-Agent (soft, CV ) | 0.7225 (0.700–0.744) | 0.7212 / 0.7386 | 0.951 ( 1.02 ) | 6.8 |
| CE-Agent (sign-only) | 0.7225 (0.700–0.744) | 0.7213 / 0.7387 | 0.950 ( 1.02 ) | 6.8 |
| CE-Agent (point-anchored) | 0.7208 (0.700–0.742) | 0.7215 / 0.7379 | 0.805 ( 1.02 ) | 6.4 |
| CE-Agent (prior-locked) | 0.7166 (0.696–0.738) | 0.7179 / 0.7350 | 0.773 ( 1.02 ) | 6.6 |
| Random Forest (pooled) | 0.7251 (0.703–0.746) | 0.7236 / 0.7410 | 1.143 (1.05) | 6.4 |
| Model / Specification | (95% CI) | / | Int. | Slope (E/O) |
|---|---|---|---|---|
| CE-Agent (hard, ) † | 0.8068 (0.758–0.855) | 0.8157 / 0.8046 | 0.8423 | 0.777 (1.07) |
| CE-Agent (soft, ) | 0.8051 (0.752–0.854) | 0.8145 / 0.8074 | 0.8411 | 0.833 (1.07) |
| CE-Agent (hard, binary CAC) | 0.8107 (0.763–0.858) | 0.8194 / 0.8035 | 0.8418 § | 0.826 (1.07) |
| CE-Agent (hard, ordinal CAC) | 0.8072 (0.759–0.855) | 0.8162 / 0.8032 | — | 0.788 ( 1.06 ) |
| CE-Agent (prior-locked) | 0.7940 (0.738–0.847) | 0.8032 / 0.8005 | 0.8416 | 0.466 (1.08) |
| Unconstrained Model | 0.7886 (0.729–0.842) | 0.7979 / 0.7952 | 0.8246 | 0.768 (1.07) |
| ML Architecture | (22/58) | (22/58) | Int. | Brier | Slope (22/58) |
|---|---|---|---|---|---|
| Cox Proportional Hazards | 0.7972 / 0.8108 | 0.8059 / 0.8216 | 0.8373 | 0.02367 | 1.401 / 0.937 |
| Random Survival Forest | 0.7832 / 0.7855 | 0.7970 / 0.8080 | 0.8236 | 0.02391 | 1.036 / 1.086 |
| Logistic Regression | 0.7876 / 0.7510 | 0.7962 / 0.7614 | 0.8209 | 0.02481 | 0.732 / 0.520 |
| Cox-Boosted Trees | 0.7793 / 0.8009 | 0.7907 / 0.8121 | 0.8192 | 0.02507 | 0.641 / 0.613 |
| DeepSurv Neural Net | 0.7760 / 0.7601 | 0.7857 / 0.7708 | 0.8067 | 0.02504 | 0.617 / 0.329 |
| XGBoost Classifier | 0.7658 / 0.7411 | 0.7753 / 0.7521 | 0.8017 | 0.02454 | 0.646 / 0.538 |
| Endpoint & Specification | (95% CI) | Int. | Brier | Slope (E/O) | |
| Hard ASCVD ( events; competing non-CVD deaths) | |||||
| CE-Agent (hard bounds) | 0.7740 (0.738–0.811) | 0.7531 | 0.8020 | 0.06820 | 0.706 ( 0.93 ) |
| CE-Agent (soft bounds) | 0.7765 (0.742–0.815) | 0.7561 | 0.8019 | 0.06630 | 0.904 (0.92) |
| ACC/AHA ASCVD-PCE | 0.7503 (0.711–0.785) | 0.7302 | 0.7774 | 0.07241 | 0.799 (1.55) |
| Hard CVD ( events; competing non-CVD deaths) | |||||
| CE-Agent (hard bounds) | 0.7740 (0.738–0.811) | 0.7520 | 0.8012 | 0.06878 | 0.701 ( 0.95 ) |
| Transported Architecture | (95% CI) | Brier | Slope (E/O) | |
|---|---|---|---|---|
| CE-Agent (compiled hard bounds) | 0.8029 (0.750–0.851) | 0.8066 | 0.02372 | 1.126 (1.06) |
| CE-Agent (point-anchored soft bounds) | 0.7961 (0.741–0.850) | 0.8052 | 0.02376 | 0.954 (1.07) |
| CE-Agent (point-anchored hard bounds) | 0.7940 (0.739–0.847) | 0.8043 | 0.02379 | 0.913 (1.07) |
| Unconstrained Logistic Reg. | 0.8029 (0.750–0.852) | 0.8023 | 0.02371 | 1.011 (1.06) |
| DeepSurv Neural Network | 0.8022 (0.748–0.852) | 0.8030 | 0.02371 | 1.026 (1.06) |
| Cox Proportional Hazards | 0.7921 (0.735–0.845) | 0.7977 | 0.02403 | 2.107 ( 1.05 ) |
| Model Architecture | AUC (IQR) | CE-Agent Wins | Excess NB ( ) | |
|---|---|---|---|---|
| CE-Agent (point-anchored bounds) | 0.727 (0.726–0.727) | 264/264 | 1,110 | 0.83 |
| Random Survival Forest | 0.686 (0.667–0.703) | 13,275 | — | |
| Cox Proportional Hazards | 0.683 (0.648–0.703) | 13,692 | 0.22 | |
| Gradient-Boosted Survival | 0.671 (0.655–0.688) | 15,777 | 0.09 | |
| XGBoost Classifier | 0.658 (0.500–0.679) | 18,790 | 0.29 | |
| Cox-Boosted Trees | 0.653 (0.500–0.683) | 19,943 | — |
| Model Architecture | (IQR) | CE-Agent Wins | Excess NB † ( ) | |
|---|---|---|---|---|
| CE-Agent (hard bounds) | 0.801 (0.787–0.807) | 43/48 | 2,237 | 2.36 |
| Random Survival Forest | 0.795 (0.784–0.804) | 4/6 | Undef. | 2.24 |
| Cox Proportional Hazards | 0.796 (0.789–0.799) | 4/6 | 8,380 | 2.25 |
| Gradient-Boosted Survival | 0.776 (0.756–0.784) | 6/6 | Undef. | 1.43 |
| XGBoost Classifier | 0.743 (0.734–0.761) | 6/6 | Undef. | 2.17 |
| Cox-Boosted Trees | 0.794 (0.784–0.797) | 5/6 | Undef. | 2.04 |
| NLST Regime | CE-Agent | Unboxed ∗ /Unc. † | Cox PH | RSF | GBSA | XGB | DSurv |
|---|---|---|---|---|---|---|---|
| External OOD discrimination (mean AUC across centers) | |||||||
| Single Hospital | 0.7246 | 0.6977 | 0.6676 | 0.6761 | 0.6620 | 0.6178 | 0.6007 |
| Pooled Hospitals | 0.7246 | 0.7134 | — | — | — | — | — |
| Pooled Hospitals | 0.7266 | 0.7228 | — | — | — | — | — |
| Pooled Hospitals | 0.7258 | 0.7244 | — | — | — | — | — |
| Pooled Hospitals | 0.7251 | 0.7245 | — | — | — | — | — |
| MESA Regime | CE-Agent | Unc. | Cox PH | RSF | GBSA | XGB-Cox | XGB |
|---|---|---|---|---|---|---|---|
| Finite-sample mortality discrimination (mean , draws) | |||||||
| ( deaths) | 0.7173 | 0.5549 | 0.7526 | 0.7487 | 0.7106 | 0.7055 | 0.5000 |
| ( deaths) | 0.7740 | 0.6789 | 0.7800 | 0.7667 | 0.7319 | 0.7445 | 0.7063 |
| ( deaths) | 0.7950 | 0.7429 | 0.7886 | 0.7672 | 0.7433 | 0.7605 | 0.7299 |
| ( deaths) | 0.8000 | 0.7659 | 0.7951 | 0.7674 | 0.7465 | 0.7663 | 0.7250 |
| Full Train ( ) | 0.8068 | 0.7886 | 0.7972 | 0.7832 | — | 0.7793 | 0.7658 |
| Mediator (Metric) | Test (C / U) | Wins | Confounding Audit | |
|---|---|---|---|---|
| 1. ( ) | 0.2990 / 0.2990 | / | concordant | |
| 2. CKD-EPI eGFR ( ) | 0.2710 / 0.2771 | / | DM: | |
| 3. ( ) | 0.2429 / 0.2467 | / | CRP: | |
| 3b. ( AUC ) | 0.7663 / 0.7677 | / | : | |
| 4. ( ) | 0.1591 / 0.1591 | ( ) | concordant | |
| 5. CT Liver Fat ( ) | 0.0883 / 0.0883 | concordant | concordant |
| Inference Regime | Full | Brier | E/O (95% CI) | Slope | |
|---|---|---|---|---|---|
| 1. Observed CT | 0.8046 | 0.8068 | 0.02390 | 1.07 (0.84–1.47) | 0.777 |
| 2. Population-Mean Fill ( ) | 0.7872 | 0.7954 | 0.02363 | 0.81 (0.63–1.11) | 0.907 |
| 3. Projected CT (rel.-gated) | 0.7884 | 0.7961 | 0.02360 | 0.84 (0.65–1.16) | 0.884 |
| 4. Projected CT (var.-gated) | 0.7903 | 0.7970 | 0.02359 | 0.95 (0.74–1.31) | 0.826 |
| 5. Projected CT (unattenuated) ∗ | 0.7904 | 0.7971 | 0.02360 | 0.96 (0.74–1.32) | 0.823 |
| 6. Retrained No-CT ( features) | 0.7903 | 0.7983 | 0.02356 | 1.06 (0.82–1.45) | 0.926 |
| Experiment | Protocol | CE-Agent | ML Baseline | Effect Size |
|---|---|---|---|---|
| 1. Interval-Bound Sample Scaling | hospitals ( active bounds) | : : | Four bounds freed: ; Cox: | ( , ) ( ) |
| 2. Factorial Bound Attribution ( ) | All bound coalitions across centers ( ) | Joint | Sum of singletons ; Four bounds freed | main-effect share; interaction |
| 3. CAC Provenance Audit | Replace the NLST CAC estimate [ 8 ] with the ITALUNG trial estimate [ 65 ] ( ) | NLST estimate ; ITALUNG estimate | Free CAC ; mean RSF ; Cox ; LR | ; free-CAC model above baselines |
| 4. Sign-Inversion Perturbation | All sign coalitions ( inverted) | flipped: med. | flipped: ; flipped: ; flipped: | ( inverted) |
| 5. Positive Scale Resilience | Global prior multiplier ( span) | Held-out | Sign inversion ( larger) | across span |
| 6. Algebraic Equivalence | Include derived vascular-aging composite ( ) | Screened: | Included: , | (Proposition 1 ) |
| Experiment | Protocol | CE-Agent | ML Baseline | Effect Size |
|---|---|---|---|---|
| 1. Interval-Bound Sample Scaling | ( active bounds) | : Full: | Five bounds freed: ; Unc.: | |
| 2. Factorial Bound Attribution ( ) | All coalitions at ( draws) & full split | : ; Full | : ( ); Full ( ) | additive; interaction |
| 3. CAC Provenance Audit | Binary free/locked vs. Detrano interval [ 37 ] vs. Drop CAC | CVD bin. free (locked ); | Unc. ; RSF ; Drop CAC: ASCVD ( CI to ) | on ASCVD at ( draws) ∗ |
| 4. Sign-Inversion Perturbation | inverted signs (CVD death & Hard ASCVD) | flipped: (CVD), (ASCVD) | flipped: (CVD), (ASCVD); Unc. flipped | (CVD), (ASCVD) |
| 5. Positive Scale Resilience | Global prior multiplier ( ) | across | Zero-locked ridge ( free): ( ) | ( smaller vs. sign flip) |
| 6. Algebraic Equivalence | Include derived composites ( test) | Screened: | Included: , | (Proposition 1 ) |
| Dimension | Metric / Setup | CE-Agent | Baselines | Summary |
|---|---|---|---|---|
| 1. Interval Accuracy | Rel. MAE vs. MC ( ) | 2.30% [2.05–2.57%] Cond. imp. ( ) | MC-4000: [ ]; MC-400: ; single Gaussian: ; : | cells; lung graph: |
| 2. Predictive SD | Rel. MAE vs. MC | 2.00% [1.83–2.14%] Cond. imp. | MC-4000: ; MC-400: ; : ( pooled) | cells ( lower vs. MC-4000) |
| 3. Single-Core Latency | Batch / single-record ms per patient | 0.62 ms quad. ( 1.23 ms end-to-end) / 1.80 ms single | MC-400: ; MC-4000: ; MC- : | quad. speedup vs. MC-4000 |
| 4. Counterfactual Monotonicity | Spurious risk rises when setting CAC/emph | 0 / 3,990 ( ) violations | MC-400: (CAC) & (emphysema) rises | ( ) sign violations |
| 5. Reliability Calibration Slope | Slope across MCAR/MAR cells ( ) | 0.959–1.047 (reliability score); ECE | Untuned gate: ; Pop.-mean: ; Plug-in: | Lowest log-loss in cells ( CI ) |
| 6. Missingness Retention | Smaller AUC loss ( ) vs. baselines | Higher retention vs. all 6 trained baselines at MCAR/MAR | Retention gain at MCAR : Cox , XGB-Cox , XGB , RSF , Tree , LR | trained baselines ( CI above zero) ∗ |
| Dimension | Metric / Setup | CE-Agent | Baselines | Summary |
|---|---|---|---|---|
| 1. Interval Accuracy | Rel. MAE vs. MC ( ) | 0.50% med. ( 0.52% mean) [0.43–0.64%] ( ) | MC-4000: ; MC-400: ; : | cells ( lower vs. MC-4000) |
| 2. Predictive SD | Rel. MAE vs. MC | 1.28% med. ( 1.95% mean) [0.85–4.45%] | MC-4000: ( mean); MC-400: | cells ( lower vs. MC-4000) |
| 3. Single-Core Latency | Quadrature / end-to-end ms per patient | 0.72 ms quad. / 1.13 ms end-to-end | MC-400: ; MC-4000: | quad. ( end-to-end) vs. MC-4000 |
| 4. Counterfactual Monotonicity | Spurious risk rises across non-zero edges | 0 / 2,720 ( ) violations | Monte Carlo sampling produces rank reversals | ( ) sign violations |
| 5. Reliability Calibration Slope | Slope across cells (CVD death / Hard ASCVD) | 0.881–1.083 (CVD) / 0.92–1.17 (ASCVD) | Untuned: ; Pop.-mean: ; Plug-in: | cells within |
| 6. Missingness Retention | Smaller loss ( ) vs. baselines | Higher retention vs. all 6 main baselines; on par or better vs. under MAR | Gain (MCAR/MAR): RSF / ; Cox / ; LR / | Point estimate higher: main; MAR, MCAR Ext.; on par or better, MCAR: ASCVD, CVD |
| Benchmark | Protocol | CE-Agent | Baseline | Effect Size |
| inverse-sample-size scaling, adaptive weight selection, and separation control | ||||
| 1. NLST Weight Decay | centers (up to training participants) | : | Subsample grid : | Slope rises as |
| 2. NLST Sample-Size Selector | ( ) if | Oracle gap across all | Pure inner CV ( ): ( ) | AUC vs. fixed hard at |
| 3. NLST Prior-Shift Resilience | Synthetic shift SD ( cells) | Inner CV (row 2 rule at ): oracle gap | Fixed hard bounds: up to | Inner CV detects shift in cells |
| 4. MESA Free-Column | Unit-normal prior ( ) on unpriced cols | separated; : , : | Unreg. cols: ( ), ( ) | ( ), Chinese fold, scanner fold |
| 5. MESA Adaptive Schedule | ( ) across cells | Oracle gap ( CIs cover ) | Pure inner CV ( ): oracle gap | Hard ASCVD ( ): |
| Benchmark | Protocol | CE-Agent | Baseline | Clinical Effect |
| Counterfactual interventional direction consistency across single-environment and pooled fits | ||||
| 1. NLST Smoking & BMI ( ) | Cessation, , [ 5 , 46 ] | rises ( , subsample rev.) | rev.: quit, pack-yr, BMI ( rises) | LOCO BMI rev. in ML ( rises) ∗ |
| 2. MESA LDL ( ) | Counterfactual ( with recorded lipids) vs. CTT RCT [ 39 ] | discordant rises ( , fits) | ML configs reversed ( rises) | Eliminates statin-confounded lipid inversion |
| 3. MESA SBP ( ) | Counterfactual vs. Ettehad RCT [ 47 ] | discordant rises ( , fits) | RSF , RF , XGB-Cox , DeepSurv rises | Prevents tree/neural local non-monotonicity |
| Two-stage clinical CT reading and statin triage efficiency | ||||
| 4. NLST Opportunistic LDCT Triage | Primary-prev. LDCT train ( , non-zero-event hospital fits, ) [ 48 , 49 ] | LOCO ( CT); | Intermediate : ( CT); ML | reads ( , CI , ); vs. ML † |
| Model / Specification | LOCO LDCT | Test | Slope ( ) | 10-yr Death § |
| CE-Agent (hard, primary) | 0.727 (0.708–0.745) ∗ | 0.708 (0.689–0.728) | 0.99 (0.99) | 0.714 (0.693–0.735) |
| CE-Agent (soft, ) | 0.728 (0.709–0.746) ∗ | 0.709 (0.691–0.729) | 0.99 (0.99) | 0.716 (0.696–0.738) |
| CE-Agent (sign-only) † | 0.728 (0.708–0.746) ∗ | 0.710 (0.691–0.729) | 0.98 (0.99) | 0.715 (0.695–0.737) |
| Cox PH (CIF, 11 pred.) | 0.727 (0.707–0.746) | 0.711 (0.691–0.731) | 2.76 ( 1.00 ) | 0.713 (0.693–0.735) |
| Logistic Reg. (11 pred.) | 0.727 (0.708–0.745) | 0.710 (0.690–0.729) | 0.92 (1.10) | 0.712 (0.692–0.734) |
| XGBoost (11 pred.) | 0.726 (0.707–0.744) | 0.709 (0.688–0.729) | 0.87 (1.08) | 0.706 (0.685–0.727) |
| Evaluation Dimension | Protocol / Cohort | CE-Agent | Comparators | Key Finding |
|---|---|---|---|---|
| 1. Single-Hospital OOD ( ) | single-site fits ML baselines ( cells) | 246 / 264 ( 93.2% ) wins; med. per-site ( pooled external) | Sign-only: ( , med. per-site ); RSF , Cox , GBSA , LR , XGB , DeepSurv | med. across external sites ( ) |
| 2. Multi-Center Scaling | hospitals ( subsets/level) | : | Unconstrained : ; XGB: | decays ; slope within in |
| 3. Domain-Gen. & Riley | external sites vs. DRO / Anchor / updating | Med. ; Riley ( ) | Anchor: ; : ; CVaR-DRO: ; Offset LR/XGB: ; Unconstr. : ; LR: | lower ( at matched ); LOCO NB in sites |
| 4. Progressive LDCT Integration | Clinical only + LDCT nodule + Visual emphysema | Held-out ( ): ( sign-only); LOCO: | Held-out full LDCT: LR ( ), XGB , PLCOm2012 ; LOCO full LDCT: PLCOm2012 ∗ | held-out CT gain ( ; nodule step , ; full cohort , ) |
| 5. Smoking Counterfactuals | Quit smoking ( ), pack-yr ( ) | 0 / 33 ( ) hospital fits with any risk inversion (quit ) | Sign-only: ( ); Unconstr. LR/Cox: ; XGB: ; RSF: ; DeepSurv: | Exact smoking monotonicity across of single-site fits † |
| 6. Missing-Data Uncertainty | MCAR/MAR cells ( , , Stage I lung graph) | rel. MAE 0.64–0.79% ( ); 1.21 ms /pt; gated slope 0.906–1.071 | MC-4000: ( , slower); Plug-in slope: ; Pop.-mean: | ECE reduced ( ); full-CT risk inside the no-CT PI under unattenuated projection: |
| Pipeline stage | Extracted output | Hardware | Mean (median) | Range / mode |
|---|---|---|---|---|
| DICOM to NIfTI conversion | 3D NIfTI volume | CPU | ( ) | Shared input |
| Sybil [ 80 ] | 3D lung-cancer CNN (runtime ref.) | RTX A6000 | ( ) | Concurrent |
| lungmask [ 69 ] | 3D emphysema ( ) | RTX A6000 | ( ) | Concurrent |
| TotalSegmentator ( total ) | Thoracic aortic calcium ( ) | RTX A6000 | ( ) | Cumulative |
| TotalSegmentator ( heartchambers_highres ) | Left ventricular chamber volume | RTX A6000 | ( ) | Cumulative |
| TotalSegmentator ( pulmonary_report ) | Pulmonary artery diameter ratio | RTX A6000 | ( ) |
Clinical Risk Prediction
Clinical risk prediction estimates patients’ future health outcomes from medical records, physiological signals, imaging, and other health data.
See all 238 papers