AD-LDB: A Modality-Incomplete Learning Model for Alzheimer's Disease Diagnosis

To address the issue of missing modalities in the diagnosis of Alzheimer's disease (AD), we propose the AD-LDB model, which is specifically designed to handle incomplete modalities. This model synthesizes absent modalities via the LDM-GAN latent diffusion generation module and integrates Dual Branch State Modeling (Dual-Mamba) with a Bidirectional Modal Adaptive Attention (BMAA) mechanism. This integration enables the simultaneous optimization of feature generation, temporal modeling, and modal interaction. Experimental results demonstrate that for one-year prediction tasks, the model achieves an accuracy of 95.35 %, an F1 score of 95.65 %, and a Matthews Correlation Coefficient (MCC) value of $\mathbf{9 1. 2 4 \%}$. For three-year predictions, the model attains an accuracy of $\mathbf{9 2. 3 4 \%}$ and an MCC value of 89.01 %. Compared with current mainstream models, ADLDB exhibits remarkable long-term prediction stability and holds significant potential for widespread application.