Unsupervised learning using EHR and census data to identify distinct subphenotypes of newly diagnosed hypertension patients

Background: Hypertension (HTN) is a complex condition with significant heterogeneity in presentation and treatment response. Identifying distinct subphenotypes of HTN may improve our understanding of its underlying mechanisms and guide more precise treatment or public health initiatives. Methods: Using EHR and Medicaid claims data from the OneFlorida+ research consortium (2012–2021), we identified a cohort of adult Floridians with newly diagnosed HTN (first diagnosis following two outpatient blood pressures ≥140/90 mmHg & no prior anti-HTN treatment). We extracted demographic and clinical data from the diagnosis visit and ≤1 year prior. We used hierarchical clustering (unsupervised machine learning) to identify distinct subphenotypes within the OneFlorida+ HTN population. Results: A total of 40,686 patients were included (mean ± SD age, 60.9 ± 17.5 y; 55% women). Five subphenotypes (S1-5) were identified. S1 was characterized by older age, higher Body Mass Index (BMI), and prevalent type 2 diabetes. S2 included over 50% of Black patients who were primarily women, younger, with higher BMI, but living in communities with higher levels of socioeconomic vulnerabilities. S3 contained a higher percentage of Hispanic patients with comparatively lower BMI. S4 is characterized by higher age and co-morbidities. S5 had 94% of patients with chronic kidney disease. Distinctions in social determinants of health factors were also observed. Conclusions: Unsupervised learning identified 5 HTN subphenotypes varying in demographic, socioeconomic, and risk profiles. Further investigation into the biological mechanisms of these subphenotypes and the relationships to social factors may enhance our ability to deliver targeted interventions that consider social policy implications in addition to the traditional behavioral and physiological interventions.