Classification of Activities of Daily Living (ADLs) Based Upper Limb Movements Using Machine Learning & Neural Network Classifiers

Abstract

Real time natural control of assistive, rehabilitation and prosthetic devices has gained significant importance over the last few decades for the people suffering from motor disabilities due to stroke, any spinal cord injury or amputation. Although surface electromyography (s-EMG) has been used as a viable controlling interface for several robotic devices specifically designed for post stroke therapeutic services. But these conventional controlling strategies are not feasible to design the rehabilitation or HMI systems based on simultaneous movements of multiple degrees of freedom (DOF). This paper presents a novel control strategy for HMIs which is based on the coupled use of EMG and inertial sensors. EMG and kinematic data of healthy and stroke subjects for commonly used daily life activities has been recorded. Multiple machine learning models including LDA, QDA, LSVM, QSVM, Fine KNN, Ensembled discriminant, and ensembled KNN have been applied. Besides this a tri-layered neural network classifier has also been implemented. A comparative analysis has been performed for the classification outcomes of all the applied models for EMG, IMU & EMG+IMU data. Overall, the KNN model performed well for all types of datasets with an average accuracy of 98.5% but results clearly demonstrated that average classification accuracies for all the applied models have significantly improved for EMG+IMU data which indicates that sensor fusion based control strategy for prosthesis can achieve higher performance than conventional control systems for each task. This study is an effort to provide a new EMG+IMU based technique for fast, efficient, and reliable control of robotic, rehabilitation and assistive devices for multiple movements with varying DOF.