Machine learning model of an EMG motion intention detection system for robotics rehabilitation control

• Main Author: Sy, Armyn Chang
• Format: text
• Language: English
• Published: Animo Repository 2017
• Subjects: Robotics in medicine; Medical instruments and apparatus; Electromyography; Medical rehabilitation; Engineering
• Online Access: https://animorepository.dlsu.edu.ph/etd_doctoral/535
• Institution: De La Salle University

A key factor in physical rehabilitation is the active participation of the patients in exerting effort to complete the repetitive physical motion practices in the affected parts of their limbs. These conditions pose a difficult challenge for robotic devices in determining when to provide assistance. Unlike their human counterparts who can sense and feel the reactions and effort levels of their patients, robotic devices often do not have such capabilities. This research aims to mimic the sensing of motion intention, through the use of electromyography (EMG), which are signals emanating from contracting muscles. In EMG signal analysis, factors such as muscle size, movement velocity, initial angle at the start of the movement (initial muscle flexion) affects the EMG signal amplitude. It is therefore the objective of this research to determine how various levels of these factors would affect EMG amplitudes and predict their effects on EMG behavior through a machine learning model. Eight healthy subjects performed bicep curl movements at three different initial angles performing each at three increasing movement velocities with at least thirty movement repetitions. The results were summarized, processed, and statistically analyzed in order to determine the significance of the above stated factors in affecting the EMG signal amplitude, followed by the training of a neural network model using the gathered data sets. The trained neural network model was able to predict the behavior of the actual EMG signal amplitudes given varying levels of the above-mentioned factors. Simulating the network, using test data not included in the training set, yielded similar results. The networks predicted value can now be utilized in determining the threshold value needed in motion intention detection systems for robotic rehabilitation devices.