NeuroAI-PLCO Images: Longitudinal Risk Stratification Using PLCO Imaging and Structured Data

Principal Investigator

Name
Richard Tran

Degrees
M.Eng.

Institution
Trove-AI

Position Title
Senior Machine Learning Researcher

Email
richard@trove-ai.com

About this CDAS Project

Study
PLCO (Learn more about this study)

Project ID
PLCOI-2056

Initial CDAS Request Approval
Jul 14, 2026

Title
NeuroAI-PLCO Images: Longitudinal Risk Stratification Using PLCO Imaging and Structured Data

Summary
This project will use PLCO imaging data and corresponding structured trial data from the Prostate, Lung, Colorectal, and Ovarian Cancer Screening Trial to evaluate longitudinal imaging-based cancer risk stratification. The study is retrospective, observational, and computational. No participant contact, intervention, biospecimen analysis, or clinical decision-making will occur.

This PLCO Images project is part of a broader doctoral research program on NeuroAI, a modular longitudinal architecture for trustworthy cancer risk modeling. The primary dissertation imaging cohort will use NLST low-dose CT data. PLCO imaging data will serve as a secondary imaging cohort to evaluate whether longitudinal imaging-risk modeling concepts generalize across screening datasets and imaging modalities.

The requested image data and associated structured PLCO data will be used to construct participant-level longitudinal imaging records. Relevant structured data include participant characteristics, screening records, screening results, abnormality records, diagnostic follow-up, cancer diagnosis timing, cancer characteristics when available, treatment variables when needed for outcome characterization, and mortality outcomes. Image-linkage data are required to connect PLCO images with participants and screening timepoints.

The proposed analysis will focus on screening-based risk stratification using imaging data available before or at the index screening timepoint. For each index screening exam, predictors will be restricted to information available at or before that screening timepoint to prevent temporal data leakage. Diagnostic, treatment, cancer diagnosis, and mortality variables will be used primarily for outcome definition, censoring, follow-up characterization, and retrospective evaluation rather than as pre-diagnosis predictors.

The modeling approach will compare single-image models, structured-data-only models, and longitudinal image-plus-structured-data models. The proposed NeuroAI imaging model will extract image features from serial PLCO screening images, combine them with structured participant and screening variables, incorporate time intervals between screening exams, and generate calibrated risk estimates for subsequent cancer outcomes where supported by the available PLCO image and outcome data.

Performance will be evaluated using participant-level train, validation, and test splits. Metrics will include AUROC, AUPRC, sensitivity, specificity, positive predictive value, calibration, Brier score, expected calibration error, uncertainty coverage, and lead-time analysis. Ablation studies will evaluate the contribution of imaging history, structured covariates, longitudinal modeling, uncertainty estimation, and calibration. All models will be used only for retrospective dissertation research and will not be used for clinical care or deployment.

Aims

Aim 1: Construct linked longitudinal PLCO imaging records. Link PLCO imaging data with participant-level characteristics, screening records, screening results, abnormality records, diagnostic follow-up, cancer outcomes, cancer characteristics when available, treatment variables when needed for outcome characterization, and mortality data.

Aim 2: Develop NeuroAI-PLCO Images for longitudinal imaging-based risk modeling. Implement an imaging-clinical architecture that extracts features from PLCO screening images, incorporates prior screening history, encodes time intervals between screening exams, and estimates subsequent cancer risk where supported by the available image and outcome data.

Aim 3: Compare image-only, structured-only, and multimodal longitudinal models. Evaluate whether combining PLCO imaging data with structured participant and screening data improves risk stratification compared with single-modality and single-timepoint baselines.

Aim 4: Assess generalization beyond NLST low-dose CT. Use PLCO imaging data as a secondary imaging cohort to evaluate whether NeuroAI-style longitudinal risk modeling remains useful beyond the primary NLST CT development cohort.

Aim 5: Produce interpretable and auditable imaging research outputs. Generate visual attribution analyses, longitudinal imaging summaries, feature-importance analyses, uncertainty summaries, calibration reports, and audit logs for transparent retrospective dissertation research.

Collaborators

Richard Tran Trove-AI
Srinivas Reddy Trove-AI