Machine Learning-Based Prediction of Unplanned Readmission Due to Major Adverse Cardiac Events Among Hospitalized Patients with Blood Cancers

Figures

• 

  Figure 1.. Study Flowchart. Caption: Patients Aged 18+ With a Primary Diagnosis of Blood Cancers (Leukemia, Lymphoma, Multiple Myeloma, Myelodysplastic Syndrome) Were Included. Exclusions Were Patients Who Died, Experienced Complex Care (e.g., Transfers or Combined Same-Day Stays), Were Discharged in Fourth Quarter, or had Missing Data. Data From 2017-2018 Were Used to Train Seven Machine Learning Models, With 2019 Data for Testing. SHAP Analysis was Performed for Model Interpretation, and High-Risk Patients (Predicted Readmission Risk >75th Percentile) Were Identified. SHAP Values for These Patients Were Used for Clustering Analysis, and Decision Rules Were Applied to Link Original Dataset Information With SHAP-Based Clusters for Interpretable Results. Note: NRD: National Readmission Database; SHAP: SHapley Additive exPlanations; HDBSCAN: Hierarchical Density-Based Spatial Clustering of Applications With Noise.

• 

  Figure 2.. Predictive Performance of Developed Models. Caption: These figures present the Predictive Performance Using PR-AUC (Left) and ROC-AUC (Right) of Developed Models. The CatBoost Algorithm Achieved the Highest AUC Values for Both ROC (0.737, 95% CI: 0.712-0.763) and PR (0.040, 95% CI: 0.033-0.050). However, the Differences in ROC-AUC and PR-AUC Among Most ML Algorithms Were Not Statistically Significant, Except for the CNB Algorithm. The Stacking Algorithm (Super Learner) Achieved a Slight Enhancement in Performance Compared to the Single Algorithms in These Two Metrics. Note: CB: CatBoost; LR: Logistic Regression With L2 Regularization; CNB: Complement Naïve Bayes; RF: Random Forest; SL: Super Learner; SVM: Support Vector Machine; XGB: XGBoost.

• 

  Figure 3.. SHapley Additive exPlanations (SHAP) Value Plots of Models Developed by Algorithm. Caption: These figures present the SHAP Summary Plot for Logistic Regression, Random Forest, XGBoost, and CatBoost. The figures illustrate the Impact of Each Feature (Predictor) on and Highlight Key Features of the Predicted Risk of 90-Day Unplanned MACE Hospital Readmission. The x-Axis Represents the SHAP Values, which Indicate the Direction and Magnitude of Feature Contribution to the Outcome Prediction. The Feature Value (Gradient Color) on the Left y-Axis Reflects the Impacts of Increasing (Red/Pink Color) or Decreasing of the Predicted Risk Among Features.

• 

  Figure 4.. Uniform Manifold Approximation and Projection (UMAP) Embedding of SHapley Additive exPlanations (SHAP) Values by Dimension. Caption: The Left-Hand Plot Shows the Two-Dimensional UMAP Embedding of SHAP Values for the Whole Population (n = 76 957), while the Hand-Right Plot Shows Three-Dimensional UMAP Embedding of SHAP Values for the Readmitted and High-Risk Populations (n = 16 692).