Development and internal validation of an artificial intelligence-driven precision rehabilitation model for ischemic stroke: a retrospective cohort study.

BACKGROUND: Functional recovery after ischemic stroke varies substantially across patients and stages of care. Conventional rehabilitation planning relies mainly on clinical scales and physician judgment, which may not fully capture the nonlinear interactions among neurological severity, functional status, comorbidities, treatment exposure, and rehabilitation stage. This study aimed to develop and internally validate an artificial intelligence-driven precision rehabilitation model for predicting poor rehabilitation outcome across the continuum of care in patients with ischemic stroke. METHODS: In this retrospective cohort study, 268 patients with ischemic stroke were divided into a training cohort (n = 188) and an internal validation cohort (n = 80) using stratified random sampling to preserve the outcome distribution. Candidate predictors included demographic characteristics, vascular risk factors, comorbidity burden, baseline National Institutes of Health Stroke Scale (NIHSS) score, baseline Modified Barthel Index (MBI) score, rehabilitation stage, and treatment-related variables. Poor rehabilitation outcome was defined as an MBI score below 85 or a reduction in NIHSS score of less than 2 points during the observation period. Logistic regression, random forest, extreme gradient boosting (XGBoost), and Light Gradient Boosting Machine models were developed and compared. Model performance was evaluated using discrimination, calibration, decision curve analysis, and SHapley Additive exPlanations. RESULTS: Poor rehabilitation outcome occurred in 104 patients (38.8%). In the internal validation cohort, XGBoost showed the best overall performance, with an area under the receiver operating characteristic curve of 0.862 (95% CI, 0.776-0.923), accuracy of 0.813, sensitivity of 0.806 (95% CI, 0.625-0.926), specificity of 0.816 (95% CI, 0.679-0.912), F1 score of 0.781, and Brier score of 0.139. Calibration analysis showed acceptable agreement between predicted and observed risks. Decision curve analysis indicated greater net clinical benefit than treat-all or treat-none strategies across threshold probabilities of 0.20-0.75. SHAP analysis identified baseline MBI score, baseline NIHSS score, age, structured rehabilitation therapy, rehabilitation stage, and comorbidity burden as the most influential predictors. Model-derived low-, intermediate-, and high-risk groups showed poor outcome rates of 12.8%, 34.4%, and 68.6%, respectively. CONCLUSION: The proposed artificial intelligence-driven model showed favorable discrimination, acceptable calibration, and clinically meaningful risk stratification for ischemic stroke rehabilitation. It may provide a preliminary framework for individualized rehabilitation risk stratification across the continuum of care; however, external multicenter validation and prospective clinical evaluation are required before routine clinical application.