Motion Pattern Recognition Based on Surface Electromyography Data and Machine Learning Classifiers: Preliminary Study

Objective: The aim of this preliminary study was to recognize motion patterns by classifying time series features extracted from electromyography (EMG) data of the upper limb muscles. Methods: In this study, we tested six machine learning (ML) classification models (decision trees, support vector machines, linear discriminant, quadratic discriminant, k-nearest neighbors, and efficient logistic regression) to classify time series features segmented from processed EMG data that were acquired from eight superficial muscles of two upper limbs over performing given physical activities in two main stages (supination and neutral forearm configuration) in initial and target (isometric) positions. Results: Findings indicate that in aiming to classify stages of the upper limb with the highest performance, the following ML models should be used: (1) K-NN cityblock (F1 equals 0.973/0.992) and K-NN minkowski (0.966/0.992) for the left limb in initial or target position; (2) K-NN seuclidean (0.959/0.985) and K-NN minkowski (0.957/0.986) for the right limb in initial position; (3) K-NN cityblock (0.966/0.986), K-NN seuclidean (0.959/0.985), and K-NN minkowski (0.957/0.986) for the right limb in target position. Conclusions: Upper limb positions tested in this study can be recognized based on classification of surface EMG data by using the k-nearest neighbors models (K-NN cityblock, K-NN seuclidean or K-NN minkowski) that have to be trained separately for the right and left upper limbs.