Leveraging Machine Learning for Gait Phase Classification with Varied Training Methods

Abstract

Accurate gait phase recognition is crucial for real-time analysis and intervention in rehabilitation, biomechanics, and prosthetics. However, achieving this is challenging due to the diverse machine learning (ML) training methods. This study employs ML algorithms to classify gait phases, focusing on stance and swing phases, utilizing open-source data from $\mathbf{1 0 0}$ participants ($41.91 \pm 5.3$ yrs). The classification algorithms considered are k-Nearest Neighbor (k-NN), Logistic Regression (LR), Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), and Naive Bayesian (NB) algorithms. The study evaluates these algorithm performances using two training methods with five and ten lower body movements: one randomly selects $80 \%$ of stance and swing phase data for training, while the other divides data by participants, allocating $80 \%$ for training and $20 \%$ for testing. When assessing accuracy with five movements, RF achieved $99.8 \%$ for both training methods. With ten movements, RF achieved a high accuracy of $99.9 \%$ using the second method. Notably, the performance of all ML algorithms exhibited improvement when considering data from ten movements versus five. Additionally, it was observed that the second training method proved more effective with five-movement data compared to data involving ten movements. This comprehensive evaluation highlights the potential of machine learning algorithms for accurate gait phase recognition in diverse applications with varied training methods.