Identifying common disease trajectories of Alzheimer’s disease with electronic health records

Abstract

AbstractBackgroundsAlzheimer’s disease (AD), a leading cause of dementia, poses a growing global public health challenge. While recent studies have identified AD risk factors, they often focus on specific comorbidities, neglecting the complex interrelations and temporal dynamics. Our study addresses this by analyzing AD progression through longitudinal trajectories, utilizing clinical diagnoses over time. Using machine learning and network analysis, we created a computational framework to identify common AD progression patterns.MethodsWe analyzed patient diagnoses from UC Health Data Warehouse’s Electronic Health Records, coded with the International Classification of Diseases, version 10 (ICD-10). Using the Fine and Gray model to detect significant temporal risk factors between diagnoses, we examined associations between diagnosis pairs and refined the patients’ diagnostic trajectories, delineating all possible trajectory combinations. These refined trajectories were compared using Dynamic Time Warping and grouped into clusters with hierarchical clustering. We investigated common AD trajectories through network analysis and compared patient demographics, symptoms, and AD manifestations across clusters. The Greedy Equivalence Search algorithm was used to infer causal relationships within these trajectories. We rigorously evaluated these trajectories through association tests and comparison to controls,ResultsOur analysis included 24,473 eligible AD patients, which was filtered to include 5,762 patients with 6,794 unique AD progression trajectories. We identified four trajectory clusters: 1) a mental health cluster (e.g., anxiety disorder → depressive episode) (N_patient = 1,448); 2) an encephalopathy cluster (e.g., hypertension → other disorders of brain) (N_patient = 3,223); 3) a neurodegenerative disease cluster (e.g., transient cerebral ischemic attacks → other degenerative disease of nervous system) (N_patient = 1,502); and 4) a vascular disease cluster (e.g. hypertension → other cerebrovascular diseases) (N_patient = 1,446). Significant differences were observed in demographics, symptoms, and AD features across clusters. Causal analysis indicated that 26.2% of the identified trajectory connections were causal. We also observed patients with risk trajectories faced higher risks of AD compared to those without the trajectory or with only a single risk factor.ConclusionWe uncovered AD diagnosis trajectories, incorporating temporal aspects and causal relationships. These insights improve our understanding of AD development and AD subtypes, and can enhance risk assessment. Our findings can significantly benefit patient care and medical research by moving toward earlier and more accurate diagnoses, along with personalized treatment, such as medical risk factors management and lifestyle modifications.