Speech and sMRI as multimodal biomarkers for early detection of Alzheimer’s disease using machine learning, deep learning, and transformer-based models

Alzheimer’s disease (AD) and its prodromal stage, MCI are typically diagnosed using cerebrospinal fluid analysis and amyloid positron emission tomography procedures that are invasive, costly, and poorly suited to population-scale screening. Speech and structural magnetic resonance imaging (sMRI) have emerged as complementary non-invasive alternatives: speech captures the real-time functional consequences of neurodegeneration on lexical retrieval, semantic organization, and prosody, while structural MRI reveals the anatomical substrate of cortical and subcortical atrophy that underlies these impairments. This Mini Review examines how machine learning, deep learning, graph based and transformer based models have been applied to these two modalities, individually and in combination, for classifying cognitively normal individuals and those with MCI or AD. We trace the progression from classical machine learning on hand-engineered acoustic and lexical features, through convolutional and recurrent neural networks that learn volumetric atrophy patterns and discourse-level speech representations, to vision transformers and graph neural networks that capture long-range spatial and connectomic dependencies in brain structure. We then examine multimodal fusion strategies early, late, intermediate and the cross-modal attention architectures that link speech and MRI within a single, interpretable framework. While intermediate fusion shows the strongest diagnostic performance in early evidence, this advantage narrows in the three-way classification task most relevant to clinical screening. We close by identifying dataset diversity, language dependency, and the lack of longitudinal multimodal studies as the principal barriers to clinical translation.