Distinguishing physical actions using an artificial neural network

Abstract

Analysis of electromyography (EMG) signals of normal physical actions have found to be important in order to detect certain abnormalities of the musculoskeletal system and diagnose abnormalities in patient behavior. This paper presents the results of the development of an Artificial Neural Network (ANN) for classification of EMG signals, according to the type of human behavior. The developed ANN is able to distinguish between 10 normal behaviors: bowing, clapping, handshaking, hugging, jumping, running, sitting, standing, walking, and waving. Feedforward neural network architecture was developed using dataset from UCI Machine Learning Repository database. QPC of each episode in EMG signal were obtained using bispectrum signal analysis. Training of ANN was performed using k-fold cross validation and impact of different number of neurons in hidden layer on system output was evaluated. Finally, the single-layer, feedforward neural network architecture with 17 neurons in hidden layer achieved the best performance and had sensitivity of 86.67% and of specificity 85.00%. The overall accuracy of developed structure is 86.25%.