Development and Validation of Machine Learning-Based Prediction for Dependence in the Activities of Daily Living after Stroke Inpatient Rehabilitation: A Decision-Tree Analysis

Background and Purpose: Accurate prediction using simple and changeable variables is clinically meaningful because some known-predictors, such as stroke severity and patients age cannot be modified with rehabilitative treatment. There are limited clinical prediction rules (CPRs) that have been established using only changeable variables to predict the activities of daily living (ADL) dependence of stroke patients. This study aimed to develop and assess the CPRs using machine learning-based methods to identify ADL dependence in stroke patients. Methods: In total, 1125 stroke patients were investigated. We used a maintained database of all stroke patients who were admitted to the convalescence rehabilitation ward of our facility. The classification and regression tree (CART) methodology with only the FIM subscores was used to predict the ADL dependence. Results: The CART method identified FIM transfer (bed, chair, and wheelchair) (score <= 4.0 or > 4.0) as the best single discriminator for ADL dependence. Among those with FIM transfer (bed, chair, and wheelchair) score > 4.0, the next best predictor was FIM bathing (score <= 2.0 or > 2.0). Among those with FIM transfer (bed, chair, and wheelchair) score <= 4.0, the next predictor was FIM transfer toilet (score <= 3 or > 3). The accuracy of the CART model was 0.830 (95% confidence interval, 0.804-0.856). Conclusion: Machine learning-based CPRs with moderate predictive ability for the identification of ADL dependence in the stroke patients were developed.