MADP: Multi-modal Sequence Learning for Alzheimer’s Disease Prediction with Missing Data

Alzheimer’s disease prediction is essential for enabling early diagnosis and timely intervention. These proactive measures are critical in slowing disease progression and improving the quality of life for affected individuals. A significant challenge in this context is the substantial amount of missing data, which arises due to the variable health status of subjects or other unpredictable circumstances. Moreover, existing methods struggle to accurately model the disease progression and fail to capture the effects of interactions among various factors on disease changes. To bridge this gap, we propose an end-to-end multi-modal sequence learning framework for Alzheimer’s disease prediction with missing data (MADP). Rather than employing fixed-value padding or interpolation, MADP introduces a learnable mechanism designed to address data missing and guide the model to learn the multi-modal features for subsequent disease modeling. Furthermore, we model the static composition and dynamic evolution information in the process of disease progression respectively to decompose complex disease information. An extensive group of experiments on Tadpole dataset demonstrates that the proposed MADP achieves more favorable performance.