Transformer-Based Multimodal Fusion for Alzheimer’s Disease: A Systematic Review of Neuroimaging-Genomics Integration

Early detection of Alzheimer’s disease (AD) requires integrating neuroimaging and genomic data, yet multimodal fusion remains methodologically challenging. Transformer models, particularly those employing cross-attention, offer a promising approach to model inter-modal dependencies for improved prediction and interpretability. This PRISMA 2020-compliant systematic review synthesized studies (2019–2025) on transformer-based fusion of neuroimaging (sMRI/PET) and genomics (SNPs/pathways) for AD. Ten studies met inclusion criteria after screening 1,444 records. Data were extracted on architectures, evaluation targets, genomic modeling, interpretability, and confounding control. Risk of bias was assessed using PROBAST-AI. Four studies implemented explicit cross-attention, with symmetric designs (e.g., cross-transformers) showing bidirectional genotype-phenotype insights, while asymmetric frameworks prioritized directional interactions. Diagnostic accuracy reached 96.88% for CN/MCI/AD classification, but early-detection endpoints (e.g., MCI conversion) were underrepresented. Genomic tokenization was often shallow (e.g., APOE-centric SNPs), and ancestry confounding was rarely addressed. Interpretability relied on attention maps or SHAP but lacked validation (e.g., LD-aware enrichment). Only one study combined cross-attention with progression prediction, and external validation was absent. Transformer-based fusion advances AD research but faces gaps in generalizability, genomic depth, and clinical translation. Future work should integrate LD-aware tokenization, survival analysis, and multi-cohort validation to bridge these gaps.