A Multimodal Deep Learning System for Alzheimer’s Disease Detection with Automated Clinical Feature Selection

Alzheimer’s disease (AD) is a key factor leading to dementia, which illustrates complex diagnostic challenges due to its varied exposure across imaging and clinical dimensions. Existing artificial intelligence (AI) models largely rely on a single data modality, which limits the diagnostic generalizability and diverges from real-world clinical practice that utilizes multimodal data. To address this limitation, we propose a novel multimodal deep learning framework that combines structured clinical features with magnetic resonance imaging (MRI) data for early and accurate classification of AD stages. Our approach adopts a dual-branch architecture that automatically fuses image embeddings with clinically relevant features selected through SHapley Additive exPlanations (SHAP) analysis and automated cumulative contribution thresholding techniques. This unified feature representation enables reliable classification of subjects into distinct categories: Cognitively Normal (NC), Early Mild Cognitive Impairment (EMCI), Late Mild Cognitive Impairment (LMCI), and Alzheimer’s Disease (AD). The framework has been systematically trained and evaluated on the OASIS dataset, achieving superior performance metrics compared to currently available state-of-the-art convolutional neural network architectures. Specifically, the model attained a precision of 90.89%, a recall of 90.98%, and an F1 score of 91.69%, alongside an overall accuracy of 92.52%. Moreover, the framework enhances interpretability through a transparent feature selection process, thereby contributing to a better understanding of the deeper insights involved in AD diagnosis and progression.