Performance Evaluation of Pattern Recognition Algorithms for Upper Limb Prosthetic Applications

Poly-articulated, myoelectric hand prostheses reproduce complex multi-degree of freedom movements, which are fundamental to effectively assist upper limb amputees in the execution of daily life activities. In this scenario, the control system consists in a pattern recognition algorithm translating the recorded electromyographic (EMG) activity into joint movements. However, the low decoding performance typically reached by the control system results in poor stability of the prosthetic device. In order to solve this issue, here we tested several state-of-the-art classifiers for decoding multi-joint hand movements from electromyographic recordings of arm muscles, collected from healthy subjects. Specifically, we tested: NonLinear Logistic Regression (NLR), Regularized Least-Square, Artificial Neural Network, Support Vector Machine, and Linear Discriminant Analysis. We aimed at minimizing the number of EMG electrodes (6 maximum) by optimizing each classifier in terms of the F1Score, and then we compared the performance of the classifiers. We found that the NLR algorithm achieved the best results with only 3 EMG electrodes. The optimized algorithms were then tested on three right arm amputees controlling a virtual hand. We obtained that algorithm's performance was comparable with that obtained from healthy subjects. In particular, the NLR classifier achieved 99% correct classification for all the patients, indicating its potential effective use in prosthetic applications.