External Validation and Site-Specific Retraining of CXR-Age for Cancer Outcome Prediction in an Academic Medical Center Cohort

Principal Investigator

Name
Judy Gichoya

Institution
Emory University

Position Title
William and Kay Casarella Professor of Radiology

Email
jgichoy@emory.edu

About this CDAS Project

Study
PLCO (Learn more about this study)

Project ID
PLCOI-2050

Initial CDAS Request Approval
Jun 1, 2026

Title
External Validation and Site-Specific Retraining of CXR-Age for Cancer Outcome Prediction in an Academic Medical Center Cohort

Summary
Background and Rationale. CXR-Age (Raghu, Weiss, Lu et al., JACC: Cardiovascular Imaging 2021) is an open-source deep learning model that estimates biological age from a single posterior-anterior chest radiograph. The model was developed using PLCO chest radiographs as the phenotypic-age training cohort and has been externally validated for all-cause, cardiovascular, and respiratory mortality in NLST and Korean health-checkup populations. However, the model's transportability to academic medical center populations with different demographic composition, scanner vendor mix, and clinical indications remains incompletely characterized, and its performance as a biomarker for cancer outcomes across cancer types in such populations has not been independently established.

Specific Aims. (1) Externally validate the pretrained CXR-Age model on chest radiographs from the Emory cohort, characterizing accuracy and bias relative to chronological age across demographic subgroups. (2) Fine-tune the model on Emory-cohort radiographs to assess whether site-specific retraining improves prediction. (3) Evaluate the predictive value of the CXR-Age gap for cancer incidence and cancer-related mortality across cancer types represented in the Emory cohort.

Study Design. This is a retrospective cohort study. We will use PLCO chest radiographs and linked mortality and cancer-incidence data as a reference benchmarking cohort, allowing us to reproduce the published CXR-Age performance characteristics, establish baseline bias-correction parameters on the original development population, and provide a direct comparison anchor for our Emory cohort results. The Emory cohort comprises patients with screening or routine chest radiographs and linked cancer registry follow-up. CXR-Age predictions will be generated using the publicly released model weights; site-specific fine-tuning will use a held-out Emory training partition.

Analysis Plan.
External validation: Mean absolute error and Pearson correlation between predicted and chronological age, computed in PLCO and Emory separately and stratified by age decade, sex, self-reported race/ethnicity, and scanner vendor.

Bias correction: Linear bias correction (Cole method) will be fit on a healthy reference subset within each cohort. Corrected CXR-Age gap will be used in all downstream analyses, with chronological age retained as a covariate.

Fine-tuning: The pretrained model will be fine-tuned on Emory training data using parameter-efficient methods. Performance will be compared head-to-head against the un-tuned model on a held-out Emory test set.

Cancer outcome prediction: Cox proportional hazards regression for time to cancer incidence and cancer-specific mortality, with corrected CXR-Age gap as the exposure, adjusted for chronological age, sex, smoking status, race/ethnicity, and comorbidities. Analyses will be stratified by cancer type. Discrimination will be reported as Harrell's C-index with bootstrap confidence intervals.

Subgroup analysis: Performance and predictive value will be reported stratified by demographic subgroups, with non-transportability of correction parameters treated as a substantive finding.

Anticipated Impact. This work will establish whether CXR-Age generalizes to a demographically distinct academic medical center population and quantify its value as a generic cancer risk biomarker derived from a single routine radiograph.

Aims

Aim 1: Externally validate the pretrained CXR-Age model on chest radiographs from the Emory cohort, characterizing accuracy and bias relative to chronological age across demographic subgroups.

We will apply the publicly released CXR-Age model weights to chest radiographs from the Emory cohort without any modification, generating biological age predictions for each radiograph. Predictive accuracy will be quantified using mean absolute error and Pearson correlation between predicted and chronological age. Performance will be reported in aggregate and stratified by age decade, sex, self-reported race and ethnicity, scanner vendor, and clinical indication for imaging. Linear bias-correction parameters (Cole method) will be fit on a healthy reference subset within the Emory cohort and compared to corresponding parameters derived from PLCO, treating non-transportability of correction parameters as a substantive finding worth characterizing.

Aim 2: Fine-tune the CXR-Age model on Emory-cohort radiographs to assess whether site-specific retraining improves prediction relative to the pretrained model.

The pretrained CXR-Age model will be fine-tuned on a held-out Emory training partition using parameter-efficient transfer learning methods (low-rank adaptation), which retain the prognostic representations learned from PLCO while adapting to Emory-specific scanner, demographic, and clinical-indication distributions. Two fine-tuning targets will be explored: chronological age and a phenotypic-age target derived from observed outcomes in the Emory cohort. Performance of the fine-tuned model will be compared head-to-head against the un-tuned model on a held-out Emory test set, using mean absolute error for age prediction accuracy and Harrell's C-index for outcome discrimination, with bootstrap confidence intervals for inference.

Aim 3: Evaluate the predictive value of the CXR-Age gap for cancer incidence and cancer-related mortality across cancer types represented in the Emory cohort.

Using the bias-corrected CXR-Age gap as the primary exposure, we will fit Cox proportional hazards models for time to cancer incidence and cancer-specific mortality in the Emory cohort, adjusted for chronological age, sex, smoking status, race and ethnicity, and comorbidities. Analyses will be stratified by cancer type to preserve interpretability across heterogeneous baseline hazards. Discrimination will be reported as Harrell's C-index with bootstrap confidence intervals, and incremental value beyond chronological age and clinical risk factors will be assessed via change in C-index and net reclassification improvement. Results will be reported overall and stratified by demographic subgroup to surface differential performance.

Collaborators

Judy Gichoya Emory University
Frank Li Emory University