Machine learning-enabled cardiometabolic phenotyping for heart failure detection in metabolic syndrome
S Ghay, B Kaur, M B Kassab, S C Bollepalli, H F Bahmad, T K ElajamiAbstract
Introduction
Metabolic syndrome (MetS) is associated with increased risk for heart failure (HF). The underlying pathophysiology is complex and involves adipocyte-derived inflammation, insulin resistance, dyslipidemia, and neurohormonal activation. Together, these factors promote endothelial dysfunction, myocardial injury, and fibrosis, leading to adverse cardiac remodeling and HF. Composite cardiometabolic indices capture interactions among adiposity, lipid metabolism, and body composition. We evaluated whether these indices discriminate prevalent HF among adults with MetS using seven machine learning (ML) models capable of modeling complex and nonlinear cardiometabolic interactions.
Methods
We analyzed data from the National Health and Nutrition Examination Survey (NHANES) 1999-2016 and included adults aged ≥20 years, with ≥3 of the 5 MetS criteria, HbA1c <6.5%, and with known HF status. HF was defined by self-reported physician diagnosis. Four cardiometabolic indices derived from anthropometric measures and laboratory values were evaluated: Atherogenic Index of Plasma (AIP), A Body Shape Index (ABSI), Body Fat Percentage (BFP), and Body Roundness Index (BRI). Missing data were imputed using random forest methods, class imbalance was addressed using SMOTE/SMOTEENN, predictors were standardized, and data were randomly split 70:30 into training and test sets. Seven ML models were compared with logistic regression. Performance was assessed using area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, and calibration, including Brier score and recalibration.
Results
Among 9,738 adults with MetS, 486 had HF (5.0%). The mean age was 58.5±16.3 years, and 51.5% were female. Random forest achieved the highest discrimination (AUC 0.95) with favorable calibration after recalibration and improved Brier score, followed by XGBoost (AUC 0.88) and gradient boosting (AUC 0.86) (Figure 1). K-nearest neighbors showed high sensitivity (0.87) but lower specificity (0.64) (Table 1). Support vector machine demonstrated moderate discrimination, while the neural network showed lower performance. Traditional models showed modest discrimination, including logistic regression (AUC 0.64) and naïve Bayes (AUC 0.67) (Table 1). Calibration intercepts approached zero after recalibration, although Hosmer-Lemeshow tests remained significant (p<0.05), reflecting large sample size effects. Receiver operating characteristic curves for all models are shown in Figure 1.
Conclusion
ML models incorporating composite cardiometabolic indices demonstrated superior discrimination of prevalent HF among adults with MetS compared with traditional models. These findings highlight the ability of ML to capture complex nonlinear cardiometabolic interactions and support the development of scalable screening tools for early HF phenotyping and risk stratification in high-risk MetS populations.Table 1.ML model performance for HF in MetSFigure 1.ML models ROC analysis