Multimodal Hybrid Attention-Gating Fusion Deep Learning Framework for Alzheimer's Disease Classification

Accurate diagnosis, staging, and progression assessment of Alzheimer's disease (AD) remain challenging due to heterogeneous clinical manifestations and substantial overlap among disease stages. While deep learning approaches based on structural magnetic resonance imaging (MRI) have demonstrated encouraging performance, imaging-only models often fail to capture important clinical and cognitive factors that are essential for comprehensive disease characterization. This work proposes a multimodal deep learning framework that jointly integrates MRI-derived features with clinical biomarkers for improved AD classification and staging. Unimodal (MRI-only and clinical only) and multimodal models are systematically evaluated using subject-level-cross-validation and an independent test set across multiple binary and multi-class tasks, including AD vs cognitively normal (CN), AD vs mild cognitive impairment (MCI), stable MCI vs progressive MCI (SMCI vs PMCI), and multi-class disease staging. Experimental results show that multimodal fusion consistently outperforms unimodal models, improving MCI progression prediction from 59% to 75% and increasing MCI vs AD diagnostic accuracy from 92.31% to 93.85% achieving higher accuracy, improved generalization, and reduced performance variability across evaluation settings. These findings demonstrate that combining imaging and clinical information provides a more robust and clinically meaningful approach for Alzheimer's disease diagnosis and prognosis.