Compared to existing methods for muscle activity registration, electromyography signals in modern bionic devices are the most commonly used and extensively studied. However, their application in prosthetic control may be limited by both physical and physiological factors. To overcome these challenges, there is a proposed need for research and development of more precise and adaptive algorithms for processing EMG signals, capable of compensating for individual user characteristics and external interferences. This study examines various approaches to enhance forearm prosthesis control efficiency using electromyography, employing different classifier models based on machine learning, particularly neural network classification. A comparison of the accuracy of different models using various sets of signal features in the time and frequency domains was conducted, identifying the most crucial features for classification. Among the classifiers, the Random Forest method stands out for its high classification accuracy in both frequency (0.89) and time domains (0.88), compared to other methods. For most classifiers, except for the convolutional neural network, classification accuracy in the time domain is higher than in the frequency domain, with the neural network achieving 0.87 and 0.88 accuracy, respectively. At the same time, the most effective signal features identified include Wilson amplitude, zero crossings and maximum amplitude.