An Adaptive Multimodal Meta-Fusion Framework for Alzheimer's Disease and Frontotemporal Dementia Classification using EEG, MRI, and Cognitive Biomarkers

A single-modality approach to neurodegenerative disease classification is often limited by heterogeneous clinical manifestations, modality-specific feature distributions, and incomplete data availability. This study proposes an Adaptive Multimodal Meta-Fusion (AMMF) framework integrating cognitive-demographic, structural MRI, and EEG biomarkers for classification of the Alzheimer's disease (AD), Frontotemporal Dementia (FTD), and Cognitively Normal (CN) participants. Three modality-specific cohorts from two public repositories were considered: cognitive-demographic data (n = 679), structural MRI data (n = 595), and EEG data (n = 194; BrainLat n = 106 and OpenNeuro n = 88). Cognitive-demographic features are prioritized using explainable XGBoost with SHAP-based feature analysis, MRI data are processed using ConvNeXt-Small for spatial representation learning, and EEG signals are analyzed using a Transformer-based architecture for temporal and functional representation learning. For the matched BrainLat cohort, the three modality branches generated an 18-dimensional cognitive-demographic representation, a 128-dimensional EEG representation, and a 768-dimensional MRI representation. The proposed framework then applied adaptive late fusion at the modality-probability level: each modality generated an independent three-class probability vector, and a modality-availability mask was used to renormalize the predefined modality weights of 0.20, 0.30, and 0.50 assigned to the cognitive-demographic, EEG, and MRI modalities, respectively. The resulting weighted class-probability vector was supplied to a Random Forest meta-classifier for final three-class classification. The complete multimodal evaluation was performed on 106 subject-level matched participants with all three modalities. The proposed framework achieved an accuracy of 99.06%, demonstrating the utility of combining complementary cognitive, functional, and structural biomarkers through adaptive probability-level meta-fusion for AD, FTD, and CN classification. A continuity-corrected McNemar test yielded χ² = 4.000 (p = 0.0455) for C4 versus C7 and χ² = 4.167 (p = 0.0412) for C5 versus C7, indicating statistically significant paired differences, while C6 versus C7 yielded χ² = 3.200 (p = 0.0736). Integration of cognitive-demographic, MRI, and EEG modalities for multimodal disease classification. The proposed framework integrates cognitive-demographic, MRI, and EEG modalities for multimodal disease classification by combining heterogeneous features in a unified representation space and applying Random Forest-based meta-classification to the fused modality probabilities.