Machine Learning-Enabled Classification of Forearm sEMG Signals to Control Robotic Hands Prostheses

In this paper, we aim to use and compare different machine learning algorithms, including k-nearest neighbors (k-NN) and support vector machines (SVM), to classify surface electromyography (sEMG) signals that correspond to the flexing of the 4 hand fingers, and recorded through 8 sensor channels. k-NN algorithm was optimized to find the values of k and the best type of distance, while four different kernels were used for SVM to find the optimal one. Moreover, linear discriminant analysis (LDA) was used to reduce the number of dimensions and investigate the effect of reducing the features on classification accuracy. Finally, the best performing ML algorithm was used to classify again using all possible combinations of 2 channels to assess LDA results. Training the algorithms shows that SVM with a Radial basis function (RBF) kernel outperforms k-NN and other SVM kernels, with 100% classification accuracy. Moreover, dimensionality reduction with LDA shows that using only 3 features keeps the accuracy at 100%, suggesting that using less sEMG sensors may not affect the quality of classification, which was confirmed by the result that using only channels 6 and 7 yielded 100% accuracy. Our results can pave the way for implementing strategies to decrease the cost of manufacturing prostheses and accelerate the execution of the classification algorithm since it should be performed in real-time.