AEGIS — AI-Enabled Genomic and Integrated Stratification for Multi-Cancer Risk: Deep-Learning Development and Internal Validation in the PLCO Cohort
Principal Investigator
Name
Aneel Paulus
Degrees
M.D., M.S.
Institution
West Eastern Health
Position Title
CEO
About this CDAS Project
Study
PLCO
(Learn more about this study)
Project ID
PLCO-2064
Initial CDAS Request Approval
Jul 14, 2026
Title
AEGIS — AI-Enabled Genomic and Integrated Stratification for Multi-Cancer Risk: Deep-Learning Development and Internal Validation in the PLCO Cohort
Summary
Caring for individuals across the cancer continuum has shown our team that outcomes are shaped long before diagnosis- by the accumulation of demographic, familial, behavioral, and biopsychosocial risk. This has driven our expansion into preventive medicine, where, with our partners, we are developing an integrated risk-stratification model combining biopsychosocial history, laboratory markers, hereditary and genomic risk factors, and imaging into a per-cancer risk-level profile that moves detection upstream. Building and validating such a model requires a large, well-characterized population with longitudinal cancer outcomes. The PLCO Trial, with prospective baseline risk-factor data and adjudicated incident-cancer and mortality outcomes across prostate, lung, colorectal, and ovarian cancers- is an ideal training resource.
Objectives.
1. Develop and internally validate deep-learning models for incident prostate, lung, colorectal, and ovarian cancer risk, evaluating discrimination and calibration.
2. Quantify the incremental value of multi-domain predictors over established single-domain risk factors.
3. Assess performance across demographic and screening-arm subgroups to detect calibration drift and equity gaps.
Study design.
Retrospective cohort study using PLCO as a labeled training corpus for deep-learning risk models. Baseline questionnaire, dietary/supplemental questionnaire, and screening-arm biomarker fields serve as inputs; incident-cancer and mortality endpoints, structured as time-to-event labels, serve as training targets. The four cancers will be modeled separately and within a shared multi-task representation, with the cohort partitioned into training, validation, and held-out test sets to prevent leakage.
Analysis plan.
We will train time-to-event deep neural networks (survival architectures with competing-risk handling) for each cancer, tuned by cross-validation and benchmarked against established risk models and classical survival baselines (Cox, gradient-boosted survival). Performance will be assessed on held-out data using time-dependent AUC / Harrell's C-index and calibration (calibration curves, observed-to-expected ratios) over 5- and 10-year horizons, with recalibration where indicated. Incremental value over conventional factors will be tested via ΔC-index, net reclassification improvement, and integrated discrimination improvement. Subgroup analyses (age, race/ethnicity, screening arm, and sex where cancer-appropriate) will assess fairness and calibration drift; missing data will be handled with multiple imputation; competing mortality with Fine–Gray and competing-risk models. Screening biomarkers (PSA, CA-125) are modeled as optional covariates, with sensitivity analyses excluding them to guard against target leakage. De-identified data only; no re-identification will be attempted.
Expected outcomes and significance.
This work will show whether a multi-domain model improves risk prediction beyond conventional factors and calibrate it for safe, equitable use. The objective is to develop and internally validate these models so that, in future preventive-care use, high-risk individuals can be prioritized for early detection and interventions addressing modifiable factors- smoking, physical activity, diet, body weight- alongside the psychological support that is West Eastern Health's core expertise, advancing our mission of reducing late cancer diagnosis.
Aims
Aim 1: Assemble and harmonize an ML-ready training corpus.
• Map PLCO baseline questionnaire, dietary/supplemental questionnaire, and screening arm biomarker fields (e.g., PSA, CA-125; available only for screening-arm participants) to the model’s multi-domain input schema (biopsychosocial, clinical/laboratory, familial/hereditary).
• Define time-to-event labels for incident prostate, lung, colorectal, and ovarian cancers and for mortality.
• Engineer features, encode missingness explicitly, and partition participants into training, validation, and held-out test sets with no cross-partition leakage.
Aim 2: Train deep-learning per-cancer risk models.
• Develop time-to-event deep neural networks (e.g., DeepSurv/DeepHit-style survival architectures with native competing-risk handling) for each of the four cancers, plus a shared multi-task representation that exploits cross-cancer signal.
• Optimize via cross-validation; tune regularization and class-imbalance handling to control overfitting given rare incident-event rates.
• Benchmark against established published risk models and classical baselines (Cox proportional hazards, gradient-boosted survival).
Aim 3: Evaluate discrimination, calibration, and subgroup fairness.
• Assess discrimination on held-out data via time-dependent AUC / Harrell’s C-index over 5- and 10-year horizons.
• Assess calibration (calibration curves, observed-to-expected ratios) and apply recalibration (isotonic/Platt) where indicated.
• Examine performance across age, race/ethnicity, screening arm, and sex where cancer appropriate (lung, colorectal, and the pooled multi-task model) to detect bias and calibration drift.
Aim 4: Quantify incremental value and produce a deployment-ready
specification.
• Quantify the gain from multi-domain features over conventional single-domain risk factors (change in C-index, net reclassification improvement, integrated discrimination improvement).
• Run sensitivity analyses for competing mortality (Fine–Gray and competing-risk deep learning models), screening-related detection effects, and models excluding screening biomarkers to guard against target leakage.
• Deliver a calibrated, documented model specification to inform West Eastern Health’s preventive-care risk-stratification workflow, using de-identified data only with no attempted re-identification.
Alignment note. These aims operationalize the Project Summary: Aim 1 builds the training corpus described in the study design; Aims 2–3 execute the deep-learning training and evaluation analysis plan; Aim 4 delivers the calibrated, equity-checked model that advances West Eastern Health’s mission of reducing the burden of late cancer diagnosis.
Collaborators
Aneel Paulus West Eastern Health