Machine learning-based prediction of activities of daily living in patients with stroke and other acquired brain injuries.

In rehabilitation research, machine learning (ML) algorithms are being increasingly applied to predict the ability of patients with stroke and other acquired brain injuries to perform activities of daily living (ADL). However, many previous studies have relied on complex data, such as imaging or blood tests, which require advanced equipment and specialized expertise and are difficult to apply in general clinical settings. We aimed to predict ADL independence in patients with stroke and other acquired brain injuries at the start of rehabilitation using routine assessment data. This retrospective study involved 379 patients with stroke or other acquired brain injuries who received rehabilitation at our hospital between September 2008 and August 2022. Three ML models - support vector machine (SVM), random forest (RF), and Light Gradient-Boosting Machine (LightGBM) - were used to classify Functional Independence Measure scores. Predictors included age, sex, Japan Coma Scale score, and Brunnstrom recovery stage. All models performed above chance, with SVM achieving the highest F1 score (0.862), followed by RF (0.857) and LightGBM (0.838). Analysis of variance revealed a significant effect of the classifier on performance (P < .001), with SVM and RF significantly outperforming LightGBM. SHapley Additive exPlanations analysis identified lower- and upper-limb Brunnstrom recovery stage, impaired consciousness, and age as key predictive features. ML algorithms can accurately predict ADL independence using initial rehabilitation assessment data without reliance on advanced medical technologies. The use of these algorithms may enhance clinical decision-making, and they may be particularly valuable in resource-limited settings.