Interpretable Prediction of Alzheimer’s Disease via Neural Granger Causality Discovery

Alzheimer’s disease (AD) is a common neurodegenerative disorder, and clinical scales play a critical role in its early diagnosis and in monitoring disease progression. Because most existing prediction models rely primarily on feature correlations and lack the capacity to uncover the causal mechanisms underlying cognitive decline, we propose a unified framework that integrates the Dynamic Causal Conditional Diffusion Imputation model (DCCDI) and the Neural Granger Causal Discovery and Prediction model (NGCDP). DCCDI employs a diffusion architecture with dynamic causal feature selection and multi-conditional modeling to effectively address the challenges of high dimensionality and high missingness in AD data. NGCDP introduces Neural Granger Causal Discovery into AD modeling to identify potential causal pathways linking multi-modal features to cognitive scores (MMSE, ADAS13). Through causal discovery and the estimation of Individual Treatment Effects for features such as age and APOE4, the framework substantially enhances interpretability of the model. Experiments on the TADPOLE dataset demonstrate that the proposed method outperforms existing state-of-the-art approaches in both imputation and prediction tasks.