A Cross-Modal Attention-Based Multimodal Fusion Model for Alzheimer's Disease Classification: A Focused Systematic Review and Reproducible Framework

Abstract

Alzheimer's disease (AD) classification increasingly combines structural magnetic resonance imaging (MRI), positron emission tomography (PET), cognitive and clinical measures, genetics, fluid biomarkers, speech, and other digital phenotypes. Yet, simple concatenation or late averaging cannot explicitly model how one modality changes the interpretation of another, and many reported results are vulnerable to small cohorts, incomplete modalities, data leakage, weak external validation, and poor calibration. This paper presents a PRISMA informed focused systematic synthesis of 50 DOI-linked publications from 2021-2026 and proposes CMA MFNet, a quality-aware cross-modal attention architecture for AD classification and mild cognitive impairment (MCI) conversion prediction. Modality-specific encoders create token sequences; bidirectional cross-modal attention learns conditional interactions; availability- and quality-aware gates control each modality's contribution; and modality dropout trains the system to remain useful when PET, genetics, or other inputs are absent. The framework includes class-balanced learning, cross-modal contrastive alignment, consistency regularization, calibrated uncertainty, and multimodal explanations. A reproducible validation protocol is specified using ADNI for development and OASIS-3, AIBL, and NACC for external testing, with subject-level splits, nested model selection, ablation tests, bootstrap confidence intervals, and fairness analyses. Recent evidence indicates strong average ADNI performance but materially lower performance in population-based, cross-lingual, and externally validated settings. Therefore, the proposed framework treats generalization, missingness, calibration, and interpretability as first-class objectives rather than post-hoc additions. The paper provides a high-quality research blueprint suitable for implementation and prospective clinical evaluation without fabricating unobserved experimental results.