Development and Interpretation of Machine Learning Models for Short-Term Functional Recovery and 90-Day Readmission in Patients with Cardiovascular Disease

Figures

• 

  Figure 1. Study flow diagram. Flowchart of patient inclusion, cohort allocation, model development, and model evaluation. A total of 553 patients with cardiovascular disease hospitalized between 2022 and 2024 were included after excluding records with missing data, invalid information, repeated admissions, or data errors. Patients were randomly divided into a training cohort of 388 patients and an independent testing cohort of 165 patients. The distributions of 1-week functional recovery and 90-day readmission are shown separately for the training and testing cohorts. The workflow also summarizes LASSO-based variable selection, training of 10 machine learning models with repeated cross-validation, model evaluation using discrimination, calibration, decision curve analysis, SHAP interpretation, and final deployment of the selected Random Forest and XGBoost models in a web calculator.

• 

  Figure 2. LASSO regression for feature selection. LASSO regression was used to select candidate predictors for model development. (A) Coefficient profiles of candidate variables across different values of the regularization parameter λ. As the penalty increased, less informative coefficients were gradually shrunk toward zero. (B) Cross-validation curve showing binomial deviance according to log(λ). The vertical dotted lines indicate the selected λ values based on cross-validation criteria. Variables retained under the selected penalty were used as the final candidate predictors for subsequent machine learning model construction.

• 

  Figure 3. Discrimination performance of machine learning models for 1-week functional recovery in the testing cohort. Receiver operating characteristic and precision–recall curves were used to compare the test-set performance of 10 machine learning algorithms for predicting 1-week functional recovery. (A) ROC curves showing the relationship between sensitivity and 1-specificity across probability thresholds. The Random Forest model achieved the highest test-set AUC, while XGBoost showed comparable discrimination. (B) Precision–recall curves showing the trade-off between precision and recall, providing additional evaluation of model performance for the binary recovery outcome. Together, these curves supported the selection of XGBoost as the final functional recovery model based on overall discrimination, precision–recall performance, and interpretability considerations.

• 

  Figure 4. Discrimination performance of machine learning models for 90-day readmission in the testing cohort. Testing-cohort ROC and precision–recall curves were generated to evaluate the ability of 10 machine learning models to predict 90-day readmission. (A) ROC curves demonstrate modest discrimination across models, with the Random Forest model achieving the highest AUC among the evaluated algorithms. (B) Precision–recall curves illustrate the difficulty of predicting readmission in the setting of a low event rate and class imbalance. The overall results indicate that 90-day readmission was more challenging to predict than 1-week functional recovery.

• 

  Figure 5. Calibration and clinical utility of models for 1-week functional recovery. Calibration curves and decision curve analysis were used to evaluate probability agreement and potential clinical usefulness of the functional recovery prediction models in the testing cohort. (A) Calibration curves compare mean predicted probabilities with observed event rates for each model; the diagonal reference line represents perfect calibration. Models closer to the diagonal show better agreement between predicted and observed recovery probabilities. (B) Decision curve analysis shows the net benefit of each model across different threshold probabilities, with “treat all” and “treat none” strategies included as clinical reference strategies. These analyses provide complementary information beyond discrimination alone.

• 

  Figure 6. Calibration and clinical utility of models for 90-day readmission. Calibration and decision curve analyses were performed for the 90-day readmission models in the testing cohort. (A) Calibration curves show the relationship between predicted readmission probabilities and observed readmission rates. The compressed range of observed event rates reflects the relatively low frequency of 90-day readmission. (B) Decision curve analysis evaluates the net clinical benefit of each model across threshold probabilities. The results suggest limited but variable clinical utility for readmission prediction, consistent with the modest discrimination and class imbalance observed for this endpoint.

• 

  Figure 7. SHAP-based interpretation of the final XGBoost model for 1-week functional recovery. SHAP analysis was used to interpret the final XGBoost model for predicting 1-week functional recovery. (A) Global SHAP importance plot ranking predictors according to mean absolute SHAP value. Admission Barthel Index was the most influential predictor, followed by serum sodium, serum chloride, NYHA class, and creatinine-related variables. (B) SHAP beeswarm plot showing both the magnitude and direction of each predictor’s contribution to model output across individual patients. Each point represents one patient, with color indicating the original feature value. (C) SHAP dependence plot for admission Barthel Index, showing its nonlinear relationship with the predicted probability of functional recovery. Higher admission Barthel Index values generally contributed positively to the model output, indicating a higher likelihood of early functional recovery.

• 

  Figure 8. SHAP-based interpretation of the final Random Forest model for 90-day readmission. SHAP analysis was applied to the final Random Forest model for predicting 90-day readmission. (A) Global SHAP importance plot showing the predictors with the largest average contribution to model output. Serum creatinine ranked as the most important variable, followed by LDL cholesterol, total cholesterol, C-reactive protein, endogenous creatinine, serum sodium, troponin, serum chloride, blood glucose, LVEF, hemoglobin, age, NT-proBNP, and other laboratory or clinical indicators. (B) SHAP dependence plot for serum creatinine, illustrating how creatinine values influenced the model’s predicted readmission risk. (C) SHAP beeswarm plot displaying the distribution, direction, and magnitude of feature effects across patients. These findings suggest that renal function, lipid metabolism, inflammatory status, and cardiac-related indicators contributed to the Random Forest readmission prediction model.