Project Overview
Electromyography (EMG) signals are generated by the electrical activity of skeletal muscle fibers, and are used by health professionals for several purposes, including disease diagnosis and rehabilitation. These kind of signals are random, highly non-stationary, and display subtle variations associated with anatomical and motor coordination differences between subjects. Recently, with the increase number of automatic learning techniques, it is now possible to develop and train systems to recognize movements, action potentials and diagnose neuromuscular diseases, overcoming the limitations presented by the nature of these signals.
In this project eight computational models were trained to recognize six hand gestures and generalize between different subjects using raw data recordings of EMG signals from 36 subjects. In addition to the application in the health sector, digital inclusion should be aimed at individuals with physical disabilities, with which, this models could contribute to the development of identification and interaction devices that can emulate the movement of hands.
Results Highlight
The accuracy obtained in the cross-validation procedure for Random Forest (96.25%), ANN (96.09%) and CNN (95.84%) is comparable to the literature reports and outperforms the results presented as state-of-the-art for this classification task. Furthermore, the dataset used in this research, which contains data from 36 patients is bigger than those previously used in most multi-subject EMG signal classification tasks.
Products
[1] A. Mora-Rubio et al., “Multi-subject Identification of Hand Movements Using Machine Learning,” Sustainable Smart Cities and Territories, Springer International Publishing, p. 117-128, 2022, doi: 10.1007/978-3-030-78901-5_11.