Development and Validation of an XGBoost ‐ SHAP Model for Predicting Adverse Outcomes in Elderly Cardiovascular Patients With Polypharmacy: A Retrospective Cohort Study

Figures

• 

  FIGURE 1. Study Flowchart. CONSORT‐style flowchart depicting patient selection and study cohort composition. Of 1856 patients aged ≥ 65 years with cardiovascular disease screened between January 2021 and December 2024, 656 were excluded based on predefined criteria: Terminal illness (n = 112), transfer within 48 h (n = 87), palliative care only (n = 43), missing data > 20% (n = 186), and declined follow‐up (n = 228). The final cohort of 1200 patients with polypharmacy (≥ 5 medications) was divided into training (60%, n = 720), validation (20%, n = 240), and test (20%, n = 240) sets using stratified random sampling. Outcome incidence in the full cohort: Adverse drug events 18.92% (n = 227), 30‐day readmission 17.25% (n = 207), 90‐day readmission 24.83% (n = 298), and 1‐year mortality 12.58% (n = 151).

• 

  FIGURE 2. Comprehensive Performance Evaluation of XGBoost Model for Predicting Adverse Outcomes in Elderly Polypharmacy Patients. Multi‐panel receiver operating characteristic (ROC) analysis with prediction distributions and calibration assessment across four clinical outcomes in the test cohort (n = 240). Left panels: ROC curves comparing XGBoost (solid blue line) versus logistic regression (dashed magenta line). Shaded regions represent 95% confidence intervals from 500 bootstrap iterations. Red stars mark optimal operating points determined by Youden's index. Middle panels: Density distributions of predicted probabilities stratified by actual outcome status (red = event occurred; green = no event). Vertical dashed lines indicate optimal classification thresholds. Right panels: Calibration plots showing agreement between predicted probabilities and observed frequencies. Dotted diagonal lines represent perfect calibration. (A) Adverse Drug Events: XGBoost AUC = 0.857 (95% CI: 0.790–0.910) vs. Logistic Regression AUC = 0.795 (DeLong test p < 0.001). Optimal threshold = 0.500 (sensitivity 80.00%, specificity 75.90%). Brier score = 0.159; (B) 30‐day readmission: XGBoost AUC = 0.843 (95% CI:…

• 

  FIGURE 3. SHAP Summary Plot for Adverse Drug Event Prediction. Beeswarm plot displaying global feature importance and effect directionality for all 12 predictors in the XGBoost ADE prediction model (test set, n = 240). Features ranked by mean absolute SHAP value (descending order, left axis): Age (0.196), eGFR (0.181), medication count (0.159), albumin (0.147), Charlson Comorbidity Index (0.138), hemoglobin (0.089), chronic kidney disease (0.067), heart failure (0.054), diabetes mellitus (0.046), psychotropic use (0.041), sex (0.024), and BMI (0.019). Each point represents one patient's SHAP value for that feature. Color gradient indicates feature value magnitude: Red (high), purple (medium), blue (low). X‐axis shows SHAP value (impact on model output log‐odds), with positive values increasing ADE risk and negative values decreasing risk. Point density at each Y‐coordinate reflects the distribution of SHAP values across patients. Key patterns: Higher age consistently increases risk (red points at positive SHAP values), higher eGFR decreases risk (red points at negative SHAP values), higher medication count increases risk particularly at ≥ 10 medications, and lower albumin (< 3.5 g/dL)…

• 

  FIGURE 4. SHAP Dependence Plots for Top Five Predictive Features. Five‐panel figure showing individual SHAP dependence plots illustrating feature‐outcome relationships and interaction effects (test set, n = 240). Each plot displays feature value (x‐axis) versus SHAP value for that feature (y‐axis), with color indicating interaction feature values. (A) Age: Progressive increase in SHAP value with age, with accelerated slope after age 80 (threshold effect). Color gradient by medication count reveals amplified age effect in patients with more medications (yellow‐red points show higher SHAP values at same age). (B) eGFR: Strong non‐linear inverse relationship with sharp risk increase (negative SHAP values) when eGFR < 45 mL/min/1.73 m2 (CKD stage 3b threshold). Color by age shows stronger effect in older patients (red points). (C) Medication Count: Near‐linear increase from 5 to 9 medications, with accelerated risk at ≥ 10 medications (hyperpolypharmacy threshold). Color by Charlson Comorbidity Index demonstrates amplification in patients with higher comorbidity burden (red‐orange points). (D) Albumin: Inverse relationship with threshold effect at albumin < 3.5 g/dL, where SHAP values sharply…