Identifying Alzheimer's Disease Progression Subphenotypes Via a Graph-based Framework Using Electronic Health Records

Understanding the heterogeneity of neurodegeneration in Alzheimer's disease (AD) and identifying distinct progression pathways is critical for improving diagnosis, treatment, prognosis, and prevention. Motivated by this need, this study aimed to identify disease progression subphenotypes among patients with mild cognitive impairment (MCI) and AD using electronic health records (EHRs). We developed a novel approach that combines a graph neural network (GNN)-based framework with time series clustering to characterize progression subphenotypes from MCI to AD. We applied the proposed framework to a real-world cohort of 2,525 patients (61.66% female; mean age 76 years), of whom 64.83% were Non-Hispanic White, 16.48% Non-Hispanic Black, 2.53% were of other races, and 10.85% were Hispanic. Our model identified four distinct progression subphenotypes, each exhibiting characteristic clinical patterns, with average MCI-to-AD progression times ranging from 805 to 1,236 days. These findings indicate that AD does not follow a uniform progression trajectory but instead manifests heterogeneous pathways, and the proposed framework provides an explainable, data-driven approach for delineating AD progression subphenotypes, offering actionable insights for healthcare informatics research and the clinical management of patients at risk for AD.