Machine Learning Prediction of Independent Walking 8 Weeks After Rehabilitation in Patients With Stroke: A Retrospective Study

BackgroundEarly prediction of independent walking after stroke may support rehabilitation planning, but the prognostic relevance of somatosensory pathway integrity remains unclear.ObjectivesTo compare logistic regression, used as a conventional reference model, with five machine-learning classifiers (random forest, gradient boosting, support vector machine, k-nearest neighbors, and neural network) for predicting independent walking 8 weeks after inpatient rehabilitation admission.MethodsThis single-center retrospective cohort study included 169 patients with ischemic or hemorrhagic stroke and baseline Functional Ambulation Category scores ≤3 who underwent inpatient rehabilitation. Seven baseline predictors were NIHSS, FMA-LE, BBS, TIS, MMSE-K, SEP severity, and stroke type. Six classifiers were evaluated using stratified five-fold cross-validation and pooled out-of-fold predictions. ROC-AUC was the primary measure; average precision and calibration were additionally assessed.ResultsFifty patients (29.6%) were independent walkers at 8 weeks. Complete SEP impairment was present in 36.1% of non-independent versus 4.0% of independent walkers (p < .001). AUCs ranged from 0.839 to 0.884. The neural network had the highest AUC (0.884; 95% CI, 0.832-0.937) and average precision (0.771), a calibration slope of 0.904, and the lowest Brier score (0.122). No pairwise AUC difference remained significant after Holm adjustment. BBS and SEP severity were the leading neural-network predictors by permutation importance.ConclusionsBaseline clinical and somatosensory measures showed good internally validated discrimination. The neural network had the highest AUC point estimate and average precision, but no model showed statistically superior AUC after multiplicity adjustment. External validation is required before clinical implementation.