Clinical Sciences/Health Conditions
Lang Chen, MS
Graduate Student
School of Medicine, Southeast University, NO. 87 Ding Jia Qiao, Nanjing, PR China.
Nanjing, Jiangsu, China (People's Republic)
Chuwei Tian, MD
Doctor
School of Medicine, Southeast University, NO. 87 Ding Jia Qiao, Nanjing, PR China.
Nanjing, Jiangsu, China (People's Republic)
Jinyu Wang, PhD
Professor
Department of Rehabilitation, Zhongda Hospital, School of Medicine, Southeast University, Nanjing, Jiangsu, PR China.
Nanjing, Jiangsu, China (People's Republic)
Existing machine learning (ML) studies on stroke prognosis often overlook rehabilitation intervention—a critical covariate—and only use patients’ baseline characteristics as model input, limiting clinical relevance. This retrospective cohort study aimed to develop and validate a set of ML models for predicting activities of daily living (ADL) outcomes in stroke patients receiving multiple rehabilitation interventions, while exploring the contribution of key predictors.
Design:
Data were collected from the Department of Rehabilitation, Zhongda Hospital affiliated to Southeast University, Nanjing, China (data spanning January 1, 2022 to May 1, 2025). A total of 942 patients were included, with their demographic data, stroke characteristics, baseline assessment scores, rehabilitation interventions, and laboratory results extracted as features. Nine ML models were constructed to predict ADL, defined as a BI score ≥ 60 at discharge indicating favorable outcomes. The area under the receiver operating characteristic curve (AUROC) served as the primary performance metric, and Shapley Additive Explanations were used to identify key predictors. An external validation set, consisting of 191 patients from Xuyi People’s Hospital (data spanning January 2023 to May 2025), was used to assess model generalization.
Results:
In internal validation, the support vector machine (SVM) model exhibited the best performance (AUROC = 0.854). In external validation, random forest (AUROC = 0.922) and SVM (AUROC = 0.921) ranked among the top-performing models. The BI score on admission was the most critical predictor of ADL outcomes, followed by Brunnstrom stage and age as key baseline features. Among rehabilitation interventions, speech therapy, pelvic floor muscle rehabilitation, and virtual reality-based rehabilitation showed high importance.
Conclusion:
This study is the first to integrate rehabilitation interventions into ML models for stroke ADL predictions, with robust external validation results. The findings can support clinicians in making informed decisions and facilitate precise rehabilitation for stroke patients.
