Single-Channel EMG-Based Hand Gesture Classification Using Optimized Neural Networks

The increasing number of people with disabilities has driven the advancement of assistive technology to aid their activities. One of the current developments in assistive technology is bionic hands which involve machine learning to classify hand gestures. This paper presents the outcomes of a study on hand gesture classification using muscle activity signal data from electromyography sensors while performing four types of hand gestures. Data was gathered from eight subjects, with six subjects allocated as training data and two subjects as test data. The data was recorded using a single-channel surface electromyograph sensor placed on the surface of the extensor digitorum muscle. The selected muscle was focused on classifying extension actions at the Metacarpophalangeal (MCP) joint of the thumb, index and middle fingers. Signal data was then feature extracted through preprocessing and model creation stages. The classification model was created using the Artificial Neural Network (ANN) method and leveraging Genetic Algorithms to obtain the most optimal hidden layer configuration to enhance model accuracy. After four rounds of experimentation with Genetic Algorithm hyperparameter tuning, the best model evaluation result from the four experiments is 42.87%. Subsequently, considering the application of Principal Component Analysis (PCA) to reduce feature dimensions and the differences in signal patterns from subjects, the next experiment was conducted without PCA, eliminating some weakly correlated linear features and excluding data from one subject with distinct patterns. In this experiment, the accuracy result on the test data for all gestures increased to 59.52% and 84,85% on train data.