Machine learning-based identification of heart failure with reduced ejection fraction using routine laboratory indicators

Figures

• 

  Figure 1. Study flowchart for patient screening, data preprocessing, model development, model evaluation, interpretation, and deployment. The workflow shows patient screening based on inclusion/exclusion criteria, baseline laboratory extraction within 24 h of admission, leakage-free preprocessing after data splitting, RFECV feature selection, development of six machine-learning models (LR, RF, XGB, MLP, SVM, and KNN), model evaluation using discrimination, calibration, threshold analysis, and decision-curve analysis, SHAP interpretation, and research-use web deployment. HFrEF, heart failure with reduced ejection fraction; RFECV, recursive feature elimination with cross-validation; SHAP, SHapley Additive exPlanations.

• 

  Figure 2. Cross-validated recursive feature elimination for routine laboratory predictors. Mean cross-validated AUC was used to evaluate candidate feature subsets. The selected 13-indicator panel balances parsimony and predictive performance. The shaded region represents variability across cross-validation folds. AUC, area under the receiver operating characteristic curve.

• 

  Figure 3. Receiver operating characteristic curves of the six machine-learning models in the training and test sets. (A) Training set. (B) Test set. Each curve shows sensitivity vs. 1-specificity. The dashed diagonal line represents a random classifier (AUC = 0.5). AUC values are shown in the legends. ROC, receiver operating characteristic; AUC, area under the ROC curve.

• 

  Figure 4. Calibration curves of the six machine-learning models in the training and test sets. (A) Training set. (B) Test set. The diagonal line indicates perfect agreement between predicted probabilities and observed event rates. Calibration was summarized using Brier score, calibration intercept, and calibration slope.

• 

  Figure 5. Decision-curve analysis of the six machine-learning models in the training and test sets. (A) Training set. (B) Test set. Net benefit is plotted against threshold probability. Treat-all and treat-none strategies are shown as reference lines. For the present triage scenario, positive predictions are intended to trigger prioritized echocardiography or specialist referral, not treatment initiation before imaging confirmation.

• 

  Figure 6. Precision-recall curves of the final random forest model in the training and test sets. (A) Training set. (B) Test set. Precision is plotted against recall across thresholds. The curves illustrate performance under class imbalance, where HFrEF accounted for 25.5% of the cohort.

• 

  Figure 7. SHAP summary plots for the random forest model. (A) Feature importance ranking by mean absolute SHAP values. (B) SHAP beeswarm plot. Each dot represents one patient; color indicates feature value; horizontal position indicates the direction and magnitude of contribution to HFrEF probability. SHAP, SHapley Additive exPlanations.

• 

  Figure 8. SHAP dependence plots (A-M) for the 13 selected predictors Each panel shows the SHAP value as a function of the corresponding predictor value. LOWESS-smoothed curves illustrate nonlinear predictor effects on HFrEF probability. The 13 selected predictors were proBNP, HCT, TBil, DBil, BUN, β2-MG, CRP, UA, Glb, GGT, MCH, LDL-C, and FIB.

• 

  Figure 9. Representative interface of the web-based calculator for individualized HFrEF prediction. The interface allows entry of the 13 selected laboratory indicators and returns an individualized HFrEF probability with feature-level explanations. The calculator is intended for academic research and requires external validation before clinical use.