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

Background: Early prediction of functional recovery and post-discharge readmission may support multidisciplinary rehabilitation planning, follow-up stratification, and individualized management in patients with cardiovascular disease (CVD). This study aimed to develop and interpret machine learning (ML) models for predicting 1-week functional recovery and 90-day readmission. Methods: A total of 553 CVD patients were randomly assigned to a training cohort (n=388) and a test cohort (n=165). Candidate clinical, functional, and laboratory variables were selected using least absolute shrinkage and selection operator regression. Ten ML algorithms were trained and compared: logistic regression, support vector machine, neural network, k-nearest neighbors, decision tree, random forest (RF), gradient boosting machine, light gradient boosting machine, extreme gradient boosting (XGBoost), and naïve Bayes. Model performance was evaluated using area under the receiver operating characteristic curve (AUC), precision-recall curves, calibration curves, decision curve analysis, and confusion matrices. Shapley additive explanations values were used for interpretation, and final models were implemented in a Shiny web application. Results: Overall, 361 of 553 patients (65.3%) achieved 1-week functional recovery, and 60 patients (10.8%) experienced 90-day readmission. For 1-week functional recovery, XGBoost achieved test-set AUC 0.856, accuracy 77.0%, sensitivity 76.9%, and specificity 77.2%. Key predictors were Barthel Index, New York Heart Association class, creatinine, serum chloride, and serum sodium. For 90-day readmission, RF achieved test-set AUC 0.662, accuracy 90.9%, sensitivity 22.2%, and specificity 99.3%, with creatinine, low-density lipoprotein cholesterol, total cholesterol, C-reactive protein, and creatinine clearance as top predictors. Conclusion: ML models showed promise for predicting short-term functional recovery in patients with CVD, while prediction of 90-day readmission remained challenging. The Shiny-based dual-outcome tool may serve as a research prototype to support individualized risk estimation, model interpretation, and multidisciplinary post-discharge management, but external validation and prospective evaluation are required before clinical application.