Development and Interpretation of Machine Learning Models for Short-Term Functional Recovery and 90-Day Readmission...

Area of Science:

• Cardiovascular Disease Research
• Machine Learning in Healthcare
• Predictive Analytics

Background:

• Early prediction of functional recovery and readmission is crucial for cardiovascular disease (CVD) patient management.
• Accurate prediction supports rehabilitation planning and individualized patient care.
• Machine learning (ML) offers potential for developing predictive models in CVD.

Purpose of the Study:

• To develop and interpret ML models for predicting 1-week functional recovery in CVD patients.
• To develop and interpret ML models for predicting 90-day readmission in CVD patients.
• To create a web application tool for risk estimation and model interpretation.

Main Methods:

• 553 CVD patients were divided into training (n=388) and testing (n=165) cohorts.
• Least absolute shrinkage and selection operator regression identified key clinical, functional, and laboratory variables.
• Ten ML algorithms were trained and evaluated using AUC, precision-recall curves, and calibration plots; Shapley additive explanations (SHAP) were used for interpretation.

Main Results:

• XGBoost model achieved an AUC of 0.856 for 1-week functional recovery prediction (accuracy 77.0%).
• Key predictors for recovery included Barthel Index, NYHA class, creatinine, serum chloride, and sodium.
• Random Forest model achieved an AUC of 0.662 for 90-day readmission prediction (accuracy 90.9%), with creatinine and lipid profiles as top predictors.

Conclusions:

• ML models demonstrate potential for predicting short-term functional recovery in CVD patients.
• Predicting 90-day readmission using ML remains a challenge.
• A Shiny web application prototype can aid individualized risk estimation and post-discharge management, pending external validation.