Convolutional Feature Engineering for Cross-day Personal Identification using Wrist Electromyography.

Surface electromyography (EMG) has emerged as a promising biometric modality for person identification. However, its performance can deteriorate over multi-day scenarios, necessitating robust feature extraction methods. With recent advancements in artificial intelligence, deep feature extraction and classification techniques have gained momentum in the biosignal domain. In this study, we propose MyoBM-Net, a convolutional feature extraction method for identification applications using EMG signals. Instead of using the conventional frequency domain-based feature extraction, 1D convolution and 2D convolution layers are utilized to extract spatial and channel-specific information, respectively. The performance evaluation utilized wrist EMG data from 43 participants on three different days across one month while performing hand/wrist gestures. A cross-day analysis, with training and testing data collected on separate days, was conducted to assess the robustness of EMG-based biometrics in practical settings. In cross-day identification, MyoBM-Net achieved a median rank-1 accuracy of 98.5% and outperformed the conventional feature extraction method. The proposed method resulted in a lower DBI value of 1.73, highlighting its strong potential for use in personal identification applications. The source code and results are available at https://github.com/pradhanashirbad/MyoBM-Net.