Heart failure risk prediction based on machine learning and interpretability analysis

Figures

• 

  Figure 1. Complete technical workflow of this study. The workflow comprises three main stages: [(Left), blue] data processing and evaluation, including preprocessing and stratified 80:20 train-test split (n=1,735/434) maintaining class balance; [(Middle), orange] feature analysis and model training across 10 algorithms under standardized conditions; [(Right), green] model evaluation via 6 metrics emphasizing imbalance-robust measures (PR-AUC, F1-Score), followed by dual interpretability analysis (SHAP + LIME) with cross-validation. Detailed descriptions of each component are provided in subsequent subsections.

• 

  Figure 2. Correlation heatmap of 15 clinical features. Key findings: (1) LVEF shows significant negative correlation with heart failure risk (r = −0.53), the strongest predictor, consistent with its established role as the gold standard for assessing ventricular systolic function (21–23); (2) Age (r = 0.22), Diabetes (r = 0.23), NT-proBNP (r = 0.16) show positive correlation with heart failure risk; (3) Systolic and diastolic blood pressure are highly correlated (r = 0.83); (4) Most features show weak correlation (|r| < 0.3), indicating minimal multicollinearity and supporting feature independence assumptions in predictive modeling. All correlation coefficients passed two-tailed t-test (p < 0.001).

• 

  Figure 3. SHAP Summary Plot (Beeswarm Plot). Features ranked by importance (top to bottom). Point horizontal position indicates SHAP value (positive = increased risk, negative = decreased risk); color indicates feature value magnitude (red = high, blue = low). LVEF shows widest distribution with clear stratification: low LVEF (blue) increases risk (positive SHAP), high LVEF (red) decreases risk (negative SHAP), validating its pathophysiological role.

• 

  Figure 4. ROC curve comparison of 10 models. Each curve represents a model's trade-off between sensitivity (TPR) and specificity (1-FPR) across all classification thresholds. The black dashed line represents random classifier performance (AUC = 0.5000). Logistic Regression and SVM curves are closest to the top-left corner, indicating optimal discrimination ability.

• 

  Figure 5. Precision-Recall curve comparison of 10 models. PR curves emphasize positive class (high-risk) performance, making them more clinically relevant than ROC curves for imbalanced datasets. The baseline (0.3502) represents random classifier performance. Logistic Regression maintains high precision across all recall levels, demonstrating robust minority class detection.

• 

  Figure 6. Model performance radar charts across six evaluation metrics. Each hexagonal chart displays a model's balanced performance profile. Logistic Regression exhibits the most filled and balanced profile, indicating comprehensive excellence across all metrics without obvious weaknesses.

• 

  Figure 7. Confusion matrix comparison across 10 models. Each 2 × 2 heatmap displays classification outcomes: TN (top-left), FP (top-right), FN (bottom-left), TP (bottom-right). Color intensity indicates count magnitude. Logistic Regression achieves optimal balance with low FN (28) and FP (23), critical for clinical deployment where missed diagnoses carry severe consequences.

• 

  Figure 8. Horizontal comparison of 10 models across six evaluation metrics. Models are ranked by performance (bottom to top) within each metric subplot. The best-performing model in each metric is highlighted. Logistic Regression dominates across most metrics, particularly the imbalance-robust measures (PR-AUC, F1-Score).

• 

  Figure 9. SHAP feature importance ranking. Features ranked by mean |SHAP value| in descending order. LVEF demonstrates dominant importance (≈2.3), approximately twice that of second-ranked Diabetes (≈1.2), quantitatively validating its status as the core diagnostic indicator for heart failure risk assessment, consistent with clinical guidelines.

• 

  Figure 10. LIME single instance explanation (Sample #0). Green bars: features supporting high-risk prediction; red bars: features supporting low-risk prediction. Bar length indicates contribution magnitude. This elderly diabetic patient shows moderate-low overall risk due to protective factors (normal renal function, normal blood pressure, no anemia/hypertension), demonstrating how LIME enables personalized risk profiling.

• 

  Figure 11. LIME feature importance aggregated across 50 random samples. Features ranked by mean |LIME weight| in descending order. Gradient color encoding: yellow (low importance) to purple (high importance). Top-3 ranking (LVEF, Diabetes, Age) shows complete concordance with SHAP analysis, validating cross-method consistency.

• 

  Figure 12. SHAP vs. LIME feature importance comparison. Both methods normalized to [0, 1] range for direct comparison. Features ranked by SHAP importance (bottom to top). Blue bars: SHAP; orange bars: LIME. High bar overlap for top-5 features validates cross-method consistency, with 100% top-3 concordance (LVEF, Diabetes, Age) providing robust evidence for clinical feature prioritization.