A Multimodal Time-Series Model with Adaptive Attention Mechanism for Alzheimer's Disease Classification

Alzheimer's disease (AD) is a progressive neurodegenerative disorder characterized by multi-domain symptom deterioration and distinct-stage progression patterns. Dynamic investigations using time-series data enable effective tracking of AD progression trajectories, revealing cognitive function and biomarker variations at different time points, which holds significant implications for early diagnosis. This study proposes a multimodal deep learning framework integrating Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and an adaptive attention mechanism. By fusing temporal features from functional magnetic resonance imaging (fMRI) and structural magnetic resonance imaging (sMRI) with single nucleotide polymorphism (SNP) genetic data, the framework achieves dynamic AD monitoring and early warning. The novel adaptive attention mechanism facilitates cross-modal feature interaction, balancing contribution weights between static and dynamic features while uncovering latent biological correlations across modalities. Validation on the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset demonstrates the model's superior performance in multimodal fusion and disease diagnosis, providing an effective approach for early AD detection.