Feature extraction from peak detection algorithms for enhanced EMG-based hand gesture recognition models

Achieving precise control of hand prosthetics is essential for enhancing the quality of life for individuals with arm amputations. Electromyography (EMG) signals are widely used as interfaces for these prosthetics. A key challenge in developing such interfaces is hand gesture recognition, where EMG signals are analyzed to predict the intended hand shape of the user. This study advances machine learning models for EMG-based hand gesture recognition by exploring the use of peak detection algorithms for feature extraction. We evaluated three distinct algorithms and a baseline without peak features with features derived from the detected peaks as well as other common methods. Our initial hypothesis was that peak features from a savitzky-golay filter would provide the highest accuracy because it uniquely smoothens the signal. Contrary to our hypothesis, however, when trained with a random forest model, the features from the wavelet-based peak detection algorithm achieved a higher classification accuracy than the maxima-based and savitzky-golay-based algorithms as well as the data without peak detection features when classifying two gestures (hand open/close). Further, the extracted peak features were among the most important features for all three algorithms. These results demonstrate that peak features, particularly those extracted using the best-performing, wavelet-based approach, can enhance the performance of hand gesture recognition models. This improvement could significantly benefit patients relying on prosthetic devices by enabling more accurate translation of their intended motions into device actions, ultimately improving their quality of life.