NeuroAI-Lung: Longitudinal Risk Stratification Using NLST Low-Dose CT 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
NLST (Learn more about this study)

Project ID
NLST-1519

Initial CDAS Request Approval
Jul 14, 2026

Title
NeuroAI-Lung: Longitudinal Risk Stratification Using NLST Low-Dose CT Data

Summary
This project will use National Lung Screening Trial low-dose CT screening images and corresponding participant-level structured data to develop and evaluate NeuroAI-Lung, a brain-inspired longitudinal deep learning architecture for subsequent lung cancer risk stratification. The study is retrospective, observational, and computational. No participant contact, intervention, or clinical decision-making will occur.

NeuroAI-Lung is designed as a modular research architecture inspired by functional principles of human diagnostic reasoning. Rather than interpreting each CT screening exam as an isolated snapshot, the model will analyze serial screening history, participant risk factors, screening findings, and outcome timing to estimate future lung cancer risk. The brain-inspired terminology will be used as a functional engineering analogy, not as a claim of literal biological equivalence.

NLST will serve as the primary low-dose CT development and internal validation cohort for this doctoral dissertation research. The broader research program may later evaluate generalizability using additional screening cohorts such as PLCO and UK Biobank, but this application focuses specifically on NLST imaging and structured trial data. This aligns with the dissertation roadmap, which prioritizes longitudinal screening cohorts, time-series memory, calibrated uncertainty, and rigorous evaluation for early cancer prediction.

The requested NLST data will be used to construct patient-level longitudinal records from available low-dose CT screening exams and structured variables, including demographics, smoking-related risk factors when available, screening results, diagnostic follow-up, lung cancer diagnosis timing, stage/pathology variables when available, and mortality outcomes. For each index screening timepoint, predictors will be restricted to information available at or before that timepoint to prevent temporal data leakage.

The modeling approach will compare NeuroAI-Lung against baseline approaches, including single-timepoint CT models, clinical-variable-only models, and traditional statistical or survival-analysis models. 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. Additional ablation studies will test the contribution of longitudinal history, multimodal fusion, uncertainty estimation, calibration, and iterative re-evaluation. All models will be used for retrospective research and dissertation analysis only, not for clinical deployment.

Aims

Aim 1: Construct a temporally valid longitudinal NLST cohort for lung cancer risk prediction. Link low-dose CT screening exams with available demographic, smoking-related, screening, diagnostic, lung cancer outcome, stage/pathology, and mortality variables. Define index screening visits and future outcome windows while ensuring predictors are limited to information available at or before each index date.

Aim 2: Develop NeuroAI-Lung, a brain-inspired longitudinal architecture for screening-based risk stratification. Implement modular components for CT feature extraction, multimodal fusion, longitudinal memory, uncertainty estimation, and probability calibration. The architecture will treat brain-inspired terminology as a functional design analogy, not a literal biological model.

Aim 3: Compare NeuroAI-Lung against standard baseline models. Evaluate whether the proposed longitudinal architecture improves subsequent lung cancer risk prediction compared with single-timepoint CT models, clinical-variable-only models, and traditional statistical or survival-analysis baselines.

Aim 4: Quantify the value of longitudinal history, uncertainty estimation, and model re-evaluation. Conduct ablation studies to assess whether prior screening history, time intervals, multimodal fusion, uncertainty feedback, calibration, and iterative re-evaluation improve discrimination, calibration, and sensitivity to early or subtle risk signals.

Aim 5: Generate interpretable and auditable research outputs. Produce visual attribution analyses, longitudinal trajectory summaries, feature-importance analyses, uncertainty summaries, and audit logs to support transparent retrospective evaluation. The resulting models will be used only for research and dissertation analysis, not clinical care.

Collaborators

Richard Tran Trove-AI
Srinivas Reddy Trove-AI