AHA Scientific Sessions 2026 Abstract · Nov 2, 2026
Machine Learning-Derived Cardiometabolic Phenotypes: Insights from the National Health and Nutrition Examination Survey, 2009-2018
By Tahavvori A, Mathew R, Ebrahimi N et al.
Introduction: The American Heart Association Cardiovascular-Kidney-Metabolic (CKM) framework was developed to characterize the interconnected burden of cardiometabolic, kidney, and cardiovascular disease (CVD). It remains unclear whether these stages fully capture the heterogeneity of cardiometabolic risk in the population. Hypothesis: We hypothesized that unsupervised machine learning (UML) could identify distinct cardiometabolic phenotypes that transcend conventional CKM staging and provide additional insight into chronic kidney disease (CKD) and CVD risk. Methods: An analysis of 30,352 adults aged≥18 was conducted using pooled NHANES data from 2009-2018. UML using K-means clustering was applied to age, BMI, blood pressure, HbA1c, triglycerides, and HDL cholesterol. Five clusters (Cs) emerged, identified without using CKD, CVD, or CKM stage data. The distribution of CKM stages was assessed, and then CKD and CVD outcomes across Cs. Logistic regression, weighted for the NHANES design and adjusted for age, sex, and race/ethnicity, tested associations between Cs and CKD or CVD. Results: Analysis showed a highly significant association was observed between Cs and CKM stage distribution (χ ² =9055.9, P<0.001). Although CKM Stage 2 represented the dominant category across most Cs, substantial heterogeneity was identified within this stage. Specifically, 82.2% of individuals in C3, 77.3% in C5, and 67.8% in C2 were classified as CKM Stage 2 despite markedly different cardiometabolic profiles. CKD and CVD were subsequently evaluated as outcomes across Cs. CKD prevalence ranged from 5% in C1 to 46.5% in C2, whereas CVD prevalence ranged from 1.7% in C1 to 24.3% in C4. C2 had the highest CKD burden, while C4 had the highest CVD burden and adverse outcomes. Compared with C1, C2 had the highest odds of CKD (OR 6.15, 95% CI 5.18-7.30), followed by C5 (OR 1.81, 95% CI 1.56-2.10), C4 (OR 1.77, 95% CI 1.51-2.07), and C3 (OR 1.60, 95% CI 1.40-1.82). Similar findings were observed for CVD, with the strongest association in C2 (OR 3.96, 95% CI 3.11-5.05), followed by C3 (OR 2.46, 95% CI 2.02-3.00), C4 (OR 2.14, 95% CI 1.72-2.68), and C5 (OR 1.53, 95% CI 1.23-1.91). Results were consistent across the sensitivity analysis when age was excluded from clustering, showing that Cs and outcomes were not driven solely by age. Conclusion: Although these phenotypes showed significant overlap with CKM stages, substantial heterogeneity within individual CKM stages, particularly stage 2, was observed. ML-derived phenotypes identified clinically meaningful subgroups with markedly different CKD and CVD risks despite being classified in the same CKM stage. These findings suggest that data-driven phenotyping approaches may complement existing CKM staging frameworks and improve risk stratification of cardiometabolic disease beyond conventional categorical classification.