Diabetic Patient Readmission Prediction Using BiLSTM with XGBoost

Predicting diabetic patient hospital readmissions is a crucial healthcare task, as unplanned readmissions result in increased medical expenses and may lead to treatment gaps. Proper prediction of readmissions enables better patient care and resource allocation. Whereas conventional machine learning algorithms tend to fail at retaining sequential dependencies in patient data, deep learning algorithms such as LSTM suffer when handling imbalanced datasets. A hybrid BiLSTM-XGBoost model is presented in this study to enhance readmission prediction in patients with diabetes. While XGBoost improves feature importance-based classification and gradient boosting, the BiLSTM component records temporal trends in patient histories. A highly unbalanced dataset of 50,000 diabetes patient histories—which includes class imbalance and missing value handling from preprocessing operations—is used to train and evaluate the model. The proposed BiLSTM-XGBoost model is also superior, yielding an AUROC of 0.9973, a recall of 0.9983, and an F1-score of 0.9900. The suggested model surpasses conventional machine learning algorithms like random forest, XGBoost, LightGBM, naïve Bayes, and CatBoost, as well as deep learning algorithms like LSTM, in forecasting patient readmission hazards. The findings emphasize the usefulness of integrating deep learning with boosting methods for more stable decision-making in healthcare analysis.