Personalized Chest X-ray Report Templates with Longitudinal Lung Cancer Risk Stratification Using PLCO Data

Principal Investigator

Name
Eranga Ukwatta

Degrees
PhD

Institution
University of Guelph

Position Title
Professor

Email
eukwatta@uoguelph.ca

About this CDAS Project

Study
PLCO (Learn more about this study)

Project ID
PLCOI-1909

Initial CDAS Request Approval
Jul 28, 2025

Title
Personalized Chest X-ray Report Templates with Longitudinal Lung Cancer Risk Stratification Using PLCO Data

Summary
Chest X-rays (CXRs) remain a cornerstone in lung cancer screening and follow-up care, yet radiology reports often lack individualized insights that incorporate a patient’s broader clinical profile. While AI models can identify radiologic abnormalities, they typically fail to contextualize findings with relevant longitudinal risk factors such as smoking history or medication use. The Prostate, Lung, Colorectal, and Ovarian (PLCO) Cancer Screening Trial dataset presents a valuable opportunity to address this gap by integrating imaging and patient-reported data.

This project aims to develop an AI-assisted, structured chest X-ray report template that combines:
● Radiologic findings from baseline CXRs,
● Patient-reported data from the Baseline Questionnaire (BQ) and Medication Use Questionnaire (MUQ), and
● A personalized 5-year lung cancer risk prediction model.
The resulting reports will include:
● Imaging-based findings (e.g., pulmonary nodules, emphysematous changes)
● Quantitative lung cancer risk scores (e.g., “Estimated 5-year lung cancer risk: 7.3%),
● Risk factor flags (e.g., “Heavy smoker”, “Chronic NSAID use”), and
● Follow-up recommendations based on risk stratification (e.g. “Re-screen in 6 months”).

Aims

1. Develop a Multimodal AI Pipeline
Integrate chest X-ray imaging features with tabular data from the BQ and MUQ to train a predictive model for 5-year lung cancer risk.

2. Design a Structured Report Template
Populate a fixed-format template with model predictions, risk tier labels, and key clinical flags based on questionnaire inputs (e.g., smoking, medication use).

3. Evaluate Model and Report Utility
Assess model performance using metrics such as AUC and calibration, and evaluate the interpretability and clinical utility of the generated reports compared to traditional radiology

Collaborators

Zinah Ghulam, MASc Student, Biomedical Engineering
Dr. Richa Mittal, Radiologist, Guelph General Hospital