Leakage-Aware Model Predicts Diabetic Hospital Readmission Using 100,000 Encounters

The author built a model to predict hospital readmission for diabetic patients using roughly 100,000 hospital encounters from 1999 to 2008. Records were grouped by patient so that no individual appeared in both training and test sets, guarding against data leakage. The model reached a ROC-AUC of about 0.69, clearly beating a majority-class baseline that identified no readmitted patients, though gradient boosting offered little gain over logistic regression. The decision threshold was chosen on training data, and calibration showed predicted risk closely matched observed risk. The author stresses that the model was trained on US data and would need local validation before any use in Kenya or other settings.