Racial and ethnic disparities in pancreatic cancer risk assessment using SEER–Medicare claims data.

e16010 Background: Racial and ethnic disparities in pancreatic cancer outcomes may be partially driven by differences in risk assessment using administrative healthcare data, highlighting the need for approaches that improve equity in claims-based risk stratification. Claims-based risk stratification models derived from Medicare data are increasingly used for population-level surveillance and early identification of high-risk patients; however, their performance across racial and ethnic groups has not been well characterized. Methods: Using the SEER–Medicare 2000–2019 linked dataset, we identified 26,849 patients aged 66 years or older with incident pancreatic cancer based on ICD-O-3 site codes, excluding cases diagnosed at autopsy or death certificate, those with inconsistent vital records, missing cause of death, or prior malignancy other than non-melanoma skin cancer. Longitudinal Medicare claims data including diagnoses (ICD-9/10), procedures, and healthcare utilization were used to construct claims-based risk stratification models to estimate 6-month pancreatic cancer risk. Model performance was evaluated overall and across racial and ethnic groups to assess disparities in predictive accuracy. Results: At the 6-month prediction horizon, overall model discrimination was moderate (AUROC 0.743). Substantial variation in predictive performance was observed across racial and ethnic groups, with lower accuracy among minority populations. After applying alternative modeling strategies designed to account for group imbalance during risk estimation, overall discrimination improved (AUROC 0.762), and inter-group performance disparities were substantially reduced (range decreased by 67%; standard deviation decreased by 70%). Group-specific AUROCs improved for White (0.847→0.886), Black (0.853→0.880), Asian (0.641→0.816), and Hispanic (0.845→0.860) beneficiaries. Conclusions: Claims-based pancreatic cancer risk stratification models exhibit meaningful racial and ethnic disparities in predictive performance among older adults. Modeling strategies that account for group imbalance can improve both overall discrimination and equity across racial and ethnic groups, supporting their use to address disparities in claims-based pancreatic cancer risk assessment. These findings underscore the importance of evaluating disparities in claims-based risk assessment tools used for cancer surveillance and population health management.