Prediction of Observational Gait Analysis Score in Stroke Using Sagittal Plane Gait Video and Clinical Assessment Data

Observational gait analysis is commonly used in clinical settings to assess gait dysfunction and to make treatment plans. Gait Assessment and Intervention Tool (G.A.I.T.) is one of the most useful scales for observational gait analysis. However, observational gait analysis such as G.A.I.T. requires experienced clinical skills and adequate time to score. This study proposes machine learning-based prediction method of the G.A.I.T. score on individuals with stroke by pose estimation from a single RGB camera and clinical assessment data obtained in conventional rehabilitation. Twenty-five individuals with subacute stroke participated in this study. The participants were captured by single RGB camera for self-selected speed gait on the sagittal plane. In total, 40 features from conventional clinical assessment and gait parameters by single RGB camera-based pose estimation were used as input data. Five different machine learning regression models predicted the overall score of G.A.I.T. related to lower limb and trunk motions in the sagittal plane. The predicted score was compared to the actual score evaluated by an experienced physical therapist. Model performance was assessed by root mean squared error (RMSE) and coefficient of determination. The results showed that CatBoost with Boruta achieved the lowest RMSE and the highest R2 among the five models for predicting the overall score of the G.A.I.T. related to movement on the sagittal plane. This study reveals that the proposed prediction method using clinically available RGB camera gait video and clinical assessment data has the potential to predict the G.A.I.T. score on individuals with stroke.