Machine learning-based early identification of heart failure/congestion complicating acute coronary syndrome using body composition analysis: a retrospective cohort study

Abstract

Background Heart failure (HF) is a major complication of acute coronary syndrome (ACS) and is associated with poor outcomes. Aims To develop and internally evaluate a machine learning (ML) model for early identification/risk classification of HF or congestion among patients with ACS using bioelectrical impedance analysis (BIA)-derived body composition parameters. Methods This single-center retrospective cohort study included 527 patients with ACS. Body composition was assessed within 48 h of admission using multi-frequency direct segmental BIA. The dataset was first randomly partitioned into training and locked test cohorts. Imputation, standardization, feature selection, and model selection were performed using training data only. A total of 120 model pipelines were compared by training-stage cross-validation, and the locked test set was reserved for one final evaluation. Results HF was identified in 198 patients (37.6%). The final plsRglm + glmBoost model used six BIA/body-composition variables: Protein, Muscle—Right arm, Muscle—Trunk, ECW/TBW ratio—Trunk, ECW/TBW ratio—Left leg, and arm muscle circumference. In the locked test set, the model achieved an area under the receiver operating characteristic curve (AUC) of 0.793 (95% CI 0.719–0.859) and a Brier score of 0.186 (95% CI 0.154–0.221). Calibration assessment showed an intercept of 0.266 (95% CI −0.145 to 0.709) and slope of 0.918 (95% CI 0.645–1.287). SHAP analysis indicated that ECW/TBW ratio—Trunk and ECW/TBW ratio—Left leg were the most influential predictors. Conclusions A leakage-free ML workflow using bedside BIA-derived body composition parameters showed moderate discrimination for early identification/risk classification of HF/congestion among patients with ACS. Because of the retrospective single-center design, limited timing information, and absence of external validation, the model should not be interpreted as proving prospective prediction of future HF events.