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

Figures

• 

  Figure 1. Study workflow diagram. The workflow summarizes cohort construction using the study cohort (n = 527), early in-hospital BIA assessment within 48 h of admission, HF/congestion outcome labeling during the index hospitalization, random 70:30 splitting before preprocessing (training n = 368; locked test n = 159), training-only imputation and standardization, training-only feature selection, training-stage comparison of 120 model pipelines, final-model selection by cross-validation, locked test-set evaluation, training-derived risk stratification, and SHAP-based interpretation. ACS, acute coronary syndrome; BIA, bioelectrical impedance analysis; HF, heart failure; CV, cross-validation; AUC, area under the receiver operating characteristic curve; SHAP, SHapley Additive exPlanations.

• 

  Figure 2. Machine learning model-pipeline comparison and final-model selection. (A) A total of 120 model pipelines were attempted during training-stage cross-validation; the heatmap displays pipelines with valid performance estimates after excluding unsuccessful/NA configurations, with the selected final model highlighted. (B) Receiver operating characteristic curves for the top six training-CV-ranked model pipelines in the training and locked test cohorts. Test-set AUCs were reported only after model selection and were not used to choose the final model. (C) Calibration curves for the top six model pipelines. AUC, area under the receiver operating characteristic curve; CV, cross-validation; RF, random forest.

• 

  Figure 3. Clinical utility and prediction-error assessment. (A) Decision curve analysis for the top six model pipelines, interpreted in the context of early HF/congestion risk classification and prioritization for closer monitoring or further assessment rather than automatic treatment decisions. (B) Brier score comparison across the top six model pipelines in the training and locked test cohorts. DCA, decision curve analysis; HF, heart failure.

• 

  Figure 4. Training-derived risk stratification using the final plsRglm + glmBoost model. (A) Predicted probability distribution in the locked test set using training-derived tertile thresholds (T1/T_low = 0.183 and T2/T_high = 0.450). (B) Observed HF/congestion incidence across low-, intermediate-, and high-risk strata in the locked test set with 95% confidence intervals.

• 

  Figure 5. SHAP analysis for interpretation of the final plsRglm + glmBoost model. (A) Feature importance ranking based on mean absolute SHAP values. SHAP values were computed in the training cohort (n = 368) using the training data as the background distribution and model outputs on the probability scale. (B) SHAP summary plot for the final six model variables. (C) SHAP dependence plot for ECW/TBW ratio—Trunk using original feature values, colored by ECW/TBW ratio—Left leg. (D) SHAP dependence plot for ECW/TBW ratio—Left leg using original feature values, colored by ECW/TBW ratio—Trunk. (E) Case-level SHAP explanation for a representative positive/high-probability prediction with raw feature values. (F) Case-level SHAP explanation for a representative negative/low-probability prediction with raw feature values. SHAP values indicate model contribution and do not establish causality. SHAP, SHapley Additive exPlanations; ECW/TBW, extracellular water to total body water ratio; AMC, arm muscle circumference; HF, heart failure.