International Research Conference on Smart Computing and Systems Engineering · Mar 26, 2026
Multimodal Federated Learning for Undiagnosed Diabetes Detection with Explainable AI
By Perera TR, Ekanayake H, Vibodha O et al.
The rising prevalence of diabetes requires robust and automated screening mechanisms to prevent complications like Diabetic Retinopathy (DR). However, current Artificial Intelligence (AI) solutions often rely on centralized data aggregation, which raises privacy concerns under regulations such as GDPR and HIPAA. Furthermore, most systems are unimodal, utilizing either Electronic Health Records (EHR) or retinal imaging in isolation, thereby limiting diagnostic comprehensiveness. This paper proposes a novel Multimodal Federated Learning (FL) framework that integrates structured clinical data with retinal fundus images to detect undiagnosed diabetes while preserving data privacy. We employ a Late Fusion architecture utilizing a Multilayer Perceptron (MLP) for tabular features and an EfficientNet-B2 backbone for imaging. To address the “black box” nature of deep learning, we integrate Explainable AI (XAI) modules, specifically SHAP for EHR features and Grad-CAM for retinal images, ensuring clinical interpretability. Experimental results demonstrate that this federated multimodal approach achieves superior performance compared to unimodal baselines while maintaining strict data localization.