Improving calibration and subgroup equity in diabetes readmission prediction using causal reweighting of LSTM models

Abstract

Hospital readmission among people with diabetes is a persistent quality and cost challenge, and risk models optimized solely for predictive accuracy may amplify disparities across demographic groups. This study introduces a heuristic sample-weighting strategy informed by confounder adjustment and integrated with a long short-term memory (LSTM) model for readmission prediction using electronic health record data. Using the UCI Diabetes Readmission dataset, we estimated the confounder-adjusted association between length of stay, a proxy for clinical complexity, and readmission via regression adjustment, and derived encounter-level importance weights from the residuals for use in LSTM training. Compared with a standard LSTM, the reweighted model achieved similar discrimination (AUROC) while improving probability calibration and reducing age-related disparity gaps following threshold optimization, with gender disparities already negligible at baseline. These findings suggest that confounder-informed sample weighting in temporal deep learning can improve the reliability and subgroup robustness of readmission risk estimates without materially affecting overall predictive performance, supporting the potential for more equitable risk stratification in diabetes care.