Multimodal Alzheimer’s Disease Classification From Heterogeneous Datasets Using Multihead Cross-Attention CNN–BiLSTM With BAM

Alzheimer’s disease (AD) is a progressive neurodegenerative disorder that requires reliable computational approaches to interpret complex and heterogeneous medical data. Many existing diagnostic methods rely on single-modality analysis or multimodal approaches that require strictly paired datasets, which limits their applicability when correspondences across modalities are unavailable. To address this limitation, this study proposes a multimodal deep learning framework for binary AD classification that integrates MRI, PET, clinical tabular data, and speech-derived audio features. A label-consistent feature-level fusion strategy enables learning from heterogeneous datasets while maintaining class consistency across modalities. Modality-specific preprocessing is applied to preserve domain characteristics and improve feature quality across data sources. High-level representations are extracted from imaging data using ResNet-50, clinical variables are projected into a shared latent representation through fully connected layers, and speech recordings are converted into discriminative audio feature representations. The multimodal features are fused using a multihead cross-attention mechanism to capture complementary intermodal relationships and produce a shared multimodal representation. The fused features are then processed using a CNN–BiLSTM architecture with a bottleneck attention module (BAM) for refined feature learning. Experimental evaluation achieves 92.57% accuracy, 93.17% precision, 91.88% recall, 92.47% F1-score, and 0.9732 AUC, demonstrating effective multimodal AD classification.