보행 형태의 분류를 위한 인공지능(AI) 모델 비교 및 특성 중요도 분석 = Comparison of Artificial Intelligence(AI) models and analysis of feature importance for gait pattern classification

다국어 초록 (Multilingual Abstract)

Purpose: This study aimed to classify normal and asymmetrical load gait patterns using machine learning algorithms and to identify the priority of variables that determine these gait types.
Methods: Eighty males in their 20s (age: 23.84±2.31 years, height: 172.52±7.67 cm, body weight: 63.07±10.42 kg) participated in the study. Participants performed gait at their self-selected comfortable speed, they were instructed to step on the force plate with their right and left feet sequentially (20 trials). Asymmetrical load gait was induced by carrying a shoulder bag equivalent to 10 % of their body weight on the left shoulder. A total of 46 gait parameters including kinematic and kinetic variables were initially analyzed, which were reduced to 23 variables following collinearity analysis. Four models (k-Nearest Neighbor, Support Vector Machine, Decision Tree, and Random Forest) were employed to evaluate classification accuracy and feature importance.
Results: Among the four models, the Random Forest model demonstrated the highest classification accuracy at 0.587.
Feature importance analysis revealed that lateral ground reaction force (LGRF: lateral), external rotation moment of the knee (R-knee moment: external rotation), and anterior knee force (Lknee force: anterior) were the most influential variables.
Conclusion: This study confirmed that the Random Forest model provides the most robust performance in classifying gait patterns. The identification of lateral ground reaction force and knee rotational/anterior stability as key indicators underscores the critical role of mediolateral dynamic stability and knee control mechanisms in distinguishing asymmetrical load gait.