Calibrated Source-Aware Stacking Ensemble for Early Diabetes Risk Prediction Across Diverse Medical Datasets

Abstract

Diabetes mellitus is a prevalent disease that can cause severe complications if undetected or poorly managed, affecting overall health. Diabetes prediction using machine learning is the second most prominent research focus in the field. Early prediction is important for preventive interventions and personalized healthcare. Traditional predictive models typically rely on a single data source, such as clinical, imaging, or genomics, limiting their ability to capture the multifactorial nature of disease while existing ensemble approaches assume homogeneous patient populations. This paper presents a source-aware stacking ensemble pipeline for early diabetes risk prediction that effectively integrates heterogeneous medical datasets and addresses missing patient-level data across modalities. The meta-model incorporates five base models, trained on clinical and genetic factors, retinal images, thermal foot images, and wearable-based stress predictions, one-hot encoded source identifiers and combining calibrated probabilities for final prediction. This enables it to learn from base model outputs while adapting to data source ensuring interpretability and robustness. Among five evaluated meta-classifiers, Logistic Regression performs best. This modular approach allows each model to be optimally trained on its respective dataset while enabling integrated, multimodal predictions, explicitly accounting for dataset heterogeneity. Beyond diabetes, this framework offers scalable solutions for combining diverse factors across unrelated datasets.