Interpretable Machine Learning for Stroke Recovery: Predicting Discharge and 3-Month Functional Outcomes

Abstract

IntroductionStroke is a leading cause of disability worldwide. This study uses Machine Learning models to investigate factors influencing modified Rankin Scale scores at discharge and three months post-discharge.MethodsData from 116 stroke patients were analyzed using four predictive models: Logistic Regression, Support Vector Machine, Random Forest, and Extreme Gradient Boosting (XGB). Shapley Additive Explanations (SHAP) were also employed to interpret factor significance.Results and discussionThe XGB model achieved an Area Under the Curve of 79% at discharge and 87% three months post-discharge. SHAP analysis revealed changing factor significance over time. The National Institutes of Health Stroke Scale was most critical at discharge, while post-discharge destination became more significant at three months. Age, time metrics, thrombolysis therapy, and management of long-term health issues also proved influential.ConclusionsFindings highlight the complex, evolving nature of stroke recovery. The shift in factor importance from clinical interventions to broader health management issues emphasizes the need for time-sensitive, multifaceted approaches to stroke care. This study contributes to understanding stroke recovery by identifying key influencing factors and demonstrating the value of SHAP for model interpretation. The insights gained have practical implications for rehabilitation practices. By identifying evolving predictors of recovery, the proposed framework may support early stratification of rehabilitation needs, assist clinicians in tailoring rehabilitation intensity and modality, and inform discharge destination decisions.

Ähnliche Arbeiten