An Explainable Federated Temporal Multimodal Ensemble Deep Learning Model for Early Diagnosis and Progression Prediction of Parkinson’s and Alzheimer’s Diseases Using X-FedTME-Net

Abstract

Early and accurate diagnosis of neurodegenerative diseases such as Parkinson’s disease (PD) and Alzheimer’s disease (AD) remains challenging due to overlapping symptoms, variable progression rates, and differences in imaging data across healthcare centers. To address these issues, this study proposes X-FedTME-Net, an interpretable federated temporal multimodal ensemble deep learning framework for diagnosis and longitudinal progression prediction using MRI data. The model integrates multimodal biomarkers from structural MRI, diffusion tensor imaging, and resting-state functional MRI to capture complementary anatomical, microstructural, and functional brain characteristics. A temporal transformer-based encoder models longitudinal changes in key biomarkers, while a 3D convolutional neural network extracts deep spatial features. A support vector machine further refines classification boundaries. These components are combined through a stacking ensemble strategy to improve stability and generalization. Federated learning enables privacy-preserving multi-center collaboration without sharing raw patient data. To enhance transparency, explainable AI techniques such as Gradient-weighted Class Activation Mapping and SHapley Additive Explanations identify disease-relevant regions and quantify biomarker contributions. Experimental results on multi-center public MRI datasets demonstrate superior diagnostic accuracy, reduced false positives, and improved generalizability compared to standalone models, supporting its clinical potential.