Alzheimer’s disease is an irreversible neurodegenerative disorder. Current research does not fully elucidate the pathological progression of Alzheimer’s disease across various regions of the brain. Furthermore, traditional methods focus on identifying characteristics correlated with disease states. However, these correlation-based approaches often lack interpretability, offering limited support for biomarker discovery. In contrast, causal feature selection methods allow for more accurate identification and understanding of disease biomarkers. Therefore, we propose a novel method, Interpretable Causal Feature Selection with Graph Convolutional Networks (CFGCN), for early Alzheimer’s diagnosis, grounded in rigorous causal discovery principles. Unlike traditional methods that may overlook the underlying causal structures between brain regions, CFGCN constructs a comprehensive causal graph to elucidate these relationships specifically associated with Alzheimer’s disease. By estimating causal effects, CFGCN extracts features that are not only predictive but also interpretable, providing deeper insights into how different brain regions contribute to disease progression. Experimental results indicate that our method outperforms current advanced feature selection methods, achieving superior performance across six datasets and four evaluation metrics.