Machine Learning-Based Prediction of Unplanned Readmission Due to Major Adverse Cardiovascular Events (MACE) Among Hospitalized Patients With Blood Cancers

OBJECTIVES: Survivors of blood cancer (BC) may face an elevated risk of developing subsequent cardiovascular diseases due to being exposed tocardiotoxic cancer therapies. We aim to develop a machine learning (ML) model predicting 90-day unplanned readmission due to Major Adverse Cardiovascular Events (MACE) among hospitalized patients with BCs.

METHODS: We included patients aged 18 years or older primarily diagnosed with BCs (leukemia, lymphoma, myeloma) from the Nationwide Readmissions Database (NRD) between 2017 and 2019. The outcome was unplanned readmission for MACE within 90 days after discharge. MACE included acute myocardial infarction, acute coronary syndrome/ischemic heart disease, stroke and transient ischemic attack, heart failure, revascularization procedures, and any cardiovascular death. Tree-based ML cost-sensitive model (XGBoost) were trained on the 2017-2018 NRD and tested on the 2019 NRD data. A total of 217 clinically relevant variables were included for developing the model. Model hyperparameters were tuned with the stratified five-fold cross validation using the Bayesian optimization. The Youden’s J statistic was applied to find the optimal classification threshold. The balanced accuracy, area under the receiver operating curves (AUROC), area under precision-recall curves (AUPRC), and F₂ score were evaluated for the model performance. Shapley additive explanation (SHAP) values were calculated to quantify the importance of features.

RESULTS: Among 26,415 patients with BCs, 389 (1.473%) experienced 90-day unplanned MACE readmission. The model, with the classification threshold of 0.021, demonstrated a reliable performance (Balanced accuracy = 0.657, AUPRC = 0.036, AUROC = 0.71, F₂ = 0.133). SHAP analysis identified the most influential predictors to be older age, having certain comorbidities (heart failure, coronary atherosclerosis and other heart disease, type 2 diabetes), and experiencing elective index admission.

CONCLUSIONS: The tuned XGBoost model reliably identifies hospitalized patients with BCs at risk for MACE readmission, offering implications for improving discharge management to prevent unplanned readmission for MACE among older patients with BCs.