Interpretable Machine Learning for Predicting Discharge Outcomes in Rehabilitation Settings

Abstract

Rehabilitation discharge planning is a complicated clinical procedure that is determined by various patient and care-related variables. Interpretable machine learning has the potential to aid such decisions by making correct and clear predictions with the help of routinely gathered clinical data. This study aimed to develop and compare interpretable machine learning models for predicting discharge outcomes, with a focus on identifying patients at risk of non-home discharge. A retrospective observational analysis was conducted using an open clinical dataset of neurological and orthopedic rehabilitation patients. After removing duplicate records, 8,468 unique cases were included. The original multi-class discharge variable was converted into a binary outcome of home versus non-home discharge. Data preprocessing involved the removal of fully missing variables, imputation of remaining missing values, encoding of categorical variables, and exclusion of outcome-related features to prevent data leakage. Logistic Regression and Decision Tree models were trained using an 80:20 stratified train-test split with class weighting. The Decision Tree model demonstrated superior performance, achieving an accuracy of 0.906 and an ROC-AUC of 0.962, compared to 0.760 and 0.816 for Logistic Regression. Notably, recall for non-home discharge improved from 0.686 to 0.932. Feature importance analysis indicated that prediction was driven by a small number of variables, with minimal contribution from demographic factors. Interpretable machine learning shows strong potential for predicting discharge outcomes in rehabilitation. The Decision Tree model provided both high predictive performance and a transparent decision-making structure, supporting its potential role in clinical decision support.