Preliminary Evaluation and Comparative Analysis of Lung Cancer Risk Prediction Approaches Using NLST Low-Dose CT and Clinical Data

Principal Investigator

Name
Doyeon Lee

Institution
Hecto

Position Title
Principal Researcher

Email
dylee@hecto.co.kr

About this CDAS Project

Study
NLST (Learn more about this study)

Project ID
NLST-1526

Initial CDAS Request Approval
Sep 28, 2026

Title
Preliminary Evaluation and Comparative Analysis of Lung Cancer Risk Prediction Approaches Using NLST Low-Dose CT and Clinical Data

Summary
This study aims to conduct a preliminary evaluation and comparative analysis of lung cancer risk prediction approaches using baseline low-dose computed tomography (LDCT) images and corresponding clinical, demographic, smoking, questionnaire, and screening data from the National Lung Screening Trial (NLST).

The NLST data will be used as a retrospective research dataset for feasibility assessment, cohort construction, variable evaluation, and preliminary performance comparison. The purpose of this study is not to develop or establish a final clinical prediction model solely from the NLST dataset. Rather, the study will examine whether imaging and structured clinical information from the NLST are suitable for evaluating multimodal lung cancer risk prediction methods and for supporting the design of subsequent model development studies.

The study cohort will include participants in the NLST LDCT screening arm with an available baseline T0 LDCT examination, corresponding participant-level data, and sufficient follow-up information to determine lung cancer outcomes. Participants diagnosed with lung cancer before or at baseline will be excluded. Lung cancer outcomes will be defined for multiple follow-up periods, including 1- through 6-year prediction horizons.

Preliminary analyses may include imaging-only, structured-data-only, and combined imaging-clinical approaches. Their performance will be compared using metrics such as AUROC, AUPRC, sensitivity, specificity, positive and negative predictive values, Brier score, and calibration measures. Additional analyses may evaluate class imbalance, missing data, subgroup performance, baseline screening findings, and the feasibility of identifying participants who later developed lung cancer despite limited or non-suspicious findings at baseline.

The results will be used to assess data suitability, identify important variables, establish preliminary benchmarks, and inform the design of future multimodal lung cancer risk prediction research. Only de-identified NLST data will be used, and all findings will be reported in aggregate form.

Aims

• Construct a retrospective NLST research cohort with baseline T0 LDCT images, corresponding structured data, and longitudinal lung cancer outcomes.

• Evaluate the availability, completeness, and suitability of NLST imaging, clinical, demographic, smoking, questionnaire, and screening variables for multimodal lung cancer risk research.

• Define preliminary lung cancer outcome labels for multiple follow-up periods from 1 through 6 years after baseline screening.

• Conduct preliminary performance comparisons among imaging-only, structured-data-only, and combined imaging-clinical approaches.

• Establish preliminary benchmark results using AUROC, AUPRC, sensitivity, specificity, PPV, NPV, Brier score, and calibration measures.

• Assess the effects of class imbalance, missing data, cohort selection criteria, and baseline screening findings on preliminary model performance.

• Explore performance in clinically relevant subgroups and among participants with limited or non-suspicious findings at baseline who subsequently developed lung cancer.

• Use the findings to determine data suitability and inform the design of future multimodal lung cancer risk prediction studies.

Collaborators

Doyeon Lee Hecto
YONGWOON EOM Hecto