CENTER OF PRESSURE-BASED MACHINE LEARNING GAIT SCORE AND CLINICAL RISK FACTORS TO PREDICT FUNCTIONAL OUTCOME IN ACUTE ISCHEMIC STROKE

초록

Background and Aims: The Center of pressure (COP) offers biomechanical insights in gait analysis. Although machine learning (ML) methods have been applied to stroke researches, it remains unknown whether machine learning-enabled COP analysis improves outcome prediction beyond clinical risk factors. Methods: In this prospective study, we assessed the gait function of acute ischemic stroke patients who were capable of a 10m walk with or without a aid using a pressure-sensitive mat. Variables extracted from the COP data of each footfall were utilized to develop 3 ML models including logistic regression, LightGBM, and MLP for prediction of 3-month poor functional outcome, defined as modified Rankin scores ⩾ 2. Output score (GAIT score) derived from ML model with best performance was extracted. Multivariable models were constructed using clinical variables and GAIT score from ML model to evaluate the improvement of model performance when adding GAIT score. Results: Of the 185 included patients, we analyzed 10,804 pairs of consecutive footfalls, and 37.8% of them experienced the unfavorable outcomes. Of three ML models, MLP demonstrated the highest predictive performance as the area under the receiver operating characteristic curve (AUROC) of 0.799. In the multivariable regression, age, initial NIHSS, initial Fall Efficacy Scale-International were associated with unfavorable outcome and the model had AUROC of 0.730. Output score derived from the ML model improved the performance of the multivariable model with an AUROC of 0.812. Conclusions: Early gait information associated with COP measurement has the potential power and additive benefits when combined with clinical variables to predict unfavorable outcomes in acute ischemic stroke.