Predictive analytics for cardiovascular patient readmission and mortality: An explainable approach

Area of Science:

• Cardiovascular Medicine
• Health Informatics
• Machine Learning in Healthcare

Background:

• Cardiovascular patients face high rates of adverse outcomes after hospital discharge.
• Early identification and intervention are crucial for preventing readmissions and mortality.
• Predictive modeling can aid in identifying high-risk patients for targeted care.

Purpose of the Study:

• To evaluate machine learning algorithms for predicting unplanned readmission and death in cardiovascular patients.
• To assess the explainability of these predictive models.
• To identify key risk factors for adverse outcomes at 30 and 180 days post-discharge.

Main Methods:

• Gradient boosting machines (GBMs) were trained on electronic medical records, administrative, and mortality data.
• Data from 39,255 cardiovascular patients across four Australian hospitals (2017-2021) were utilized.
• Model performance was compared against LASSO regression, HOSPITAL, and LACE indices; Shapley values were used for explainability.

Main Results:

• GBMs demonstrated strong performance, with AUCs of 0.70 for readmission and 0.87-0.90 for mortality.
• Significant predictors for readmission included elevated red cell distribution width, advanced age, high troponin/urea, and lower albumin.
• Mortality predictors included elevated red cell distribution width, advanced age, high troponin/urea, and specific white blood cell counts.

Conclusions:

• An explainable predictive algorithm was developed to identify high-risk cardiovascular patients at discharge.
• The study successfully identified key clinical and sociodemographic risk factors for readmission and mortality.
• These findings support the use of machine learning for proactive patient management in cardiovascular care.