Offline Versus Real-Time Grasp Prediction Employing a Wearable High-Density Lightmyography Armband: On the Control of Prosthetic Hands

Several studies over the last decades have investigated the use of myoelectric upper-arm prostheses to restore lost functionality and perform complex activities of daily life. Muscle-machine interfacing methods employing electromyography, sonomyography, forcemyography, and mechanomyography, have been developed offering users intuitive control of prostheses. However, these methods suffer their own drawbacks and the practical, robust, real-time control of prostheses in the execution of complex everyday life tasks remains difficult to accomplish. In this study, a wearable high-density lightmyography armband is proposed, and the offline and real-time grasp prediction schemes are compared in an attempt to deepen our understanding in real-time decoding employing lightmyography signals. Thus bringing lightmyography closer to advanced real-time prosthetic control scenarios. Offline experiments were conducted where models decoding 10 classes were trained and tested, achieving accuracies exceeding 91% with a mean of 94.11% when a Random Forest model was employed and a mean of 95.87% when a Convolutional Neural Networks model was employed. However, as models were deployed in real-time, a decline in performance was notable, reducing the accuracies to 60-70%. Other limitations were observed in this transition. Although average accuracies were high, the precision and recall for a small fraction of grasps were often low, hindering overall real-time performance. Moreover, prolonged usage led to decreased decoding accuracy, with accuracies dropping to below 35%, indicating that fatigue, sweat, or sensor shift may have affected the signals over time. Thus for the real-time application of lightmyography, it is recommended to discriminate between a smaller number of classes for which accuracies are higher (around 80% for 6 classes without prolonged use).