Machine-learning-based prognostic models for independence in toilet-related activities in patients with subacute stroke: a retrospective study.

Background: Independence in toilet‑related activities critically shapes discharge planning and caregiver burden after stroke. Reliable early‑stage prediction models could therefore aid individualized rehabilitation.

Objective: To compare the predictive performance of logistic regression (LR) and five machine learning algorithms - decision tree (DT), support vector machine (SVM), artificial neural network (ANN), k‑nearest neighbors (KNN), and ensemble learning (EL) - for toilet-related independence at discharge.

Methods: We retrospectively analyzed subacute stroke survivors admitted to Tokyo Bay Rehabilitation Hospital from March 2015 to September 2019. Independence was defined as a score ≥ 6 on four Functional Independence Measure (FIM) subitems (toileting, bladder management, bowel management, toilet transfers). Participants' characteristics and FIM subitems were entered as predictors. LR and five machine‑learning algorithms were trained with five‑fold cross‑validation. Model performances were evaluated by the area under the receiver‑operating‑characteristic curve (AUC).

Results: Of 824 participants (mean age 70.9 years), 453 (55%) were independent at discharge. In validation data, SVM (AUC = 0.9223) achieved, followed by LR (0.9202), ANN (0.9201), KNN (0.9072), EL (0.8961), and DT (0.8394). On test data, SVM and LR maintained AUCs of 0.9101 and 0.9078, whereas ANN declined to 0.8922. EL (0.9021) and KNN (0.9020) remained stable; DT (0.7864) performed the lowest. In LR, FIM-Bed to chair transfer was the strongest positive predictor, and age was the strongest negative predictor.

Conclusions: SVM provided the highest accuracy with minimal overlearning. LR offered similar performance and greater interpretability, supporting its clinical use. These models could provide valuable information in stroke rehabilitation.