Time-Aware Transformer Framework for Diabetes Readmission Prediction

Diabetes is a chronic metabolic disorder that has become one of the major global health challenges due to its high prevalence and severe complications. Effective management of this disease requires proper blood sugar control and regular patient monitoring. Recently, machine learning-based techniques have been widely used for predictive modeling in healthcare, enabling more accurate forecasting and personalized interventions. In this paper, we predict hospital readmission of diabetic patients using both traditional and advanced machine-learning techniques. The traditional models include XGBoost, LightGBM, CatBoost, Decision Tree, and Random Forest. Moreover, we utilize an LSTM neural network, one of the most powerful modern machine learning models, to capture temporal dependencies. To train and test the models, the Diabetes ۱۳۰-US Hospitals dataset, containing ۱۰۱,۷۶۷ records with ۵۰ features is used. Results show that among traditional models, LightGBM performs the best, while the Transformer-based model outperforms all traditional models and LSTM/CNN architectures by capturing long-range temporal dependencies and heterogeneous clinical features more effectively. In this work, we employ Explainable AI (XAI) techniques to enhance model interpretability and ensure decision-making transparency. Specifically, SHAP values are used to identify key factors influencing readmissions, such as the number of lab procedures and discharge disposition. This study demonstrates that model selection, validation, and interpretability are key steps in predictive healthcare modeling. This helps health providers design interventions for improved follow-up adherence and better management of diabetes.

کلیدواژه ها:

Hospital Readmission ، Predictive Modeling ، Machine Learning ، Deep Learning ، Time-Aware Modeling