Comparative Study of Deep and Ensemble Models for Risk Prediction in Multiple Chronic Diseases

Chronic non-communicable diseases (NCDs), including type 2 diabetes mellitus (T2DM), chronic kidney disease (CKD), and cardiovascular disease (CVD), frequently co-occur as comorbidities and share overlapping risk factors. Since most models detect a single disease at a time, integrated screening is rarely achievable. This paper presents a unified, interpretable deep learning framework for structured clinical data-based multidisease risk prediction. For schema alignment and consistency, the framework uses standardized preprocessing to integrate three benchmark datasets: PIMA Indians Diabetes, UCI Heart Disease, and UCI Chronic Kidney Disease. AUC-based early stopping and stratified cross-validation are used in a single standardized pipeline to evaluate four models: Random Forest (RF), XGBoost (XGB), Multi-Layer Perceptron (MLP), and TabTransformer. Feature importance and attention analysis improve model interpretability by connecting clinical insights with predictive outcomes. According to the experimental results, MLP had the highest AUC (0.837) for diabetes, TabTransformer was the best at heart disease (AUC = 0.975), and tree-based models were the best at classifying CKD (AUC = 1.000). In addition to providing a scalable basis for upcoming multimorbidity prediction and clinical decision support systems, the suggested framework creates a repeatable, disease-agnostic architecture for integrated risk assessment.