Characterize Disease Progression Subphenotypes in Real World Populations with Overweight and Obesity using a Graph-based Neural Network Framework

BackgroundObesity is a chronic, heterogeneous condition, with risks, trajectories, and treatment responses that vary widely among individuals. However, research characterizing the heterogeneity of long-term obesity progression--and its impact on the development of obesity-associated outcomes and treatment responses--is scarce. ObjectivesWe aimed to identify progression subphenotypes in a real-world population with overweight or obesity over a period of up to 10 years using electronic health records (EHRs) and evaluate the heterogeneity of treatment effects (HTE) of glucagon-like peptide-1 receptor agonists (GLP-1RAs) across these subphenotypes for major obesity-related disease outcomes. MethodsWe conducted a retrospective cohort study of adults from OneFlorida+ who were eligible for anti-obesity medication, defined by the presence of a documented diagnosis of obesity (BMI [&ge;] 30 kg/m{superscript 2}) or a BMI in the range of 27.0-29.9 kg/m{superscript 2} accompanied by at least one weight-related comorbidity. We developed an outcome-oriented graph neural network (GNN)-based model to identify progression subphenotypes of obesity. Within each subphenotype, we emulated a target trial of GLP-1RA users vs non-users, using propensity score matching and Cox proportional hazards models to evaluate obesity-related disease outcomes, including metabolic dysfunction-associated steatotic liver disease (MASLD), major atherosclerotic cardiovascular disease (ASCVD), stroke, heart failure (HF), and chronic kidney disease (CKD). ResultsAmong 237,103 adults with overweight or obesity, 58.1% were female, with a mean age of 51.8 years. The study population included 48.4% non-Hispanic White (NHW), 27.9% non-Hispanic Black (NHB), and 13.0% Hispanic individuals. Our GNN model identified three distinct progression subphenotypes: a progressive group (43.6%; "substantial-increase" BMI trajectory), an intermediate group (33.4%; "moderate-increase"), and a stable group (64.4%; "steady"). The mean follow-up durations were comparable across the subphenotypes (4.9-5.0 years). The progressive group exhibited greater baseline multimorbidity, higher opioid exposure, and an initially favorable neighborhood socioeconomic status that declined rapidly over time. The subphenotypes differed significantly in their risks of developing obesity-associated comorbidities and in all-cause mortality during follow-up (p < 0.001). GLP-1RA use was associated with a lower HF risk in two subphenotypes (HR (95% CI): 0.75 (0.61-0.91) and 0.74 (0.61-0.91), respectively) and with a possible increase in CKD risk in the progressive group (1.20 (1.02-1.41)). ConclusionsObesity progression subphenotypes derived from routine EHRs reveal markedly heterogeneity in trajectories, clinical risks, and treatment responses. Distinct BMI trajectory groups highlight the potential of data-driven, phenotype-guided care and expose meaningful variation in GLP-1RA treatment effects. These findings underscore the potential of longitudinal EHR analysis to advance precision obesity medicine and warrant prospective validation of HTE.