Multiclass classifiers for hand-gesture recognition of electromyographic signals from WyoFlex Band

This manuscript analyses the performance of different machine learning models classifying hand gestures from electromyography (EMG) signals. The EMG information is obtained from the WyoFlex armband, a wearable bracelet with four sensors that capture the EMG of the forearm. This study considers four different models, the K-Nearest Neighbor (KNN), Support Vector Machines (SVM), Artificial Neural Networks (ANN), and Long Short-Term Memories (LSTM) for classifying four and six hand gestures. The methodology comprises different scenarios, including the application of the Synthetic Minority Over-Sampling (SMOTE) algorithm to increase the cardinality of the dataset and the Minimum-Redundancy-Maximum-Relevance (MRMR) technique to reduce computational costs. These models are compared considering the classification of four and six different hand gestures in three different scenarios: a) including the SMOTE technique, b) including the MRMR procedure without SMOTE, and c) including both MRMR and SMOTE. The results show an overall accuracy between 79.32% to 94.90% on the four gestures classification and from 75.14% to 92.80% for six gestures classification, with the SVM classifier producing the best performance in training time and accuracy. The application of the MRMR algorithm yields a reduction of the training time in all models applying the K-fold and Leave-One-Subject-Out (LOSO) cross validation methods, the accuracy is not strongly affected reducing the number of features in almost all the models.