Multi-Class Gait Phase Recognition using Machine Learning Models with Two Training Methods

Abstract

Walking is a fundamental human activity, yet understanding its intricacies is crucial for diagnosing and treating various gait abnormalities and musculoskeletal disorders. This study investigates the machine learning-based classification of gait phases into five subphases with two training methods. Using data from 100 individuals obtained from an open-source platform. The gait cycle was divided into the following subphases: Initial Contact, Loading Response, Mid Stance, Terminal Stance, and Pre-Swing. In the first approach, a stratified random sampling method was employed, allocating 80% of data from each subphase for training and 20% for testing. The second method involved training with data from 80% participants and testing with data from 20% participants. Out of different ML algorithms, the Random Forest (RF) achieves the highest performance (99.2%) with similar results across both methods. The k-Nearest Neighbors (k-NN), the Logistic Regression (LR), the Decision Tree (DT), and the Support Vector Machine (SVM) show comparable performance ($\sim 91 \%$) with marginal improvement with the second method. Naive Bayes (NB) has a significantly lower performance ($\sim 77.7 \%$). These findings suggest that RF is a robust and effective algorithm for this task. Furthermore, the study indicates that applying machine learning techniques to datasets divided by subjects presents promising opportunities for creating precise gait analysis systems. This can assist clinicians and researchers in enhancing patient care and gaining deeper insights into human locomotion.