Transformer-Based Multimodal Learning for Early Detection of Mild Cognitive Impairment in Alzheimer's Disease: A Cross-Modal Attention Framework on the ADNI Cohort

Detecting Mild Cognitive Impairment (MCI) is a central problem of computational neurology owing to the heterogeneous nature of MCI presentations and the highly individualized nature of conversion trajectories. A mild cognitive impairment is a pre-stage of Alzheimer’s disease (AD). The current study proposes a Multimodal Transformer framework (MM-TF) that jointly assesses 3D structural MRI volumes, longitudinal cognitive scores (MMSE, ADAS-Cog, CDR), and genetic-demographic covariates (APOE ε4 status, age and sex) within a single cross-modal attention backbone. Each modality is tokenized by a modality-specific encoder a shared fusion transformer with multi-head co-attention learns cross-modal interactions which are routed to a four-way diagnostic head that is CN/EMCI/LMCI/AD and a conversion-risk head at 36-months. Assessed on 2,214 subjects from the public Alzheimer’s Disease Neuroimaging Initiative (ADNI), the framework achieves an illustrative four-way accuracy of 92.3% (macro-F1 0.921) and sMCI-vs-pMCI AUC of 0.902, outperforming CNN-based and early-fusion baselines in all tested configurations. A broad ablation reveals the contribution of cross-modal attention (+4.2 points relative to early concatenation) and of genetic covariates (+2.5). Numeric findings presented must be replaced with the authors' actual results from experimental runs prior to submission.