Abstract Objective. Wrist electromyography (EMG) is emerging as an enticing wearable input modality for human-machine interaction. Traditionally recorded from the forearm for use in transradial prostheses, wrist-based EMG sensors are now being integrated into devices such as watches and wristbands for hand gesture recognition (HGR). Consumer familiarity with wrist-worn devices makes wrist EMG a compelling option, but the need for individualized user calibration remains a challenge. Approach. This study therefore evaluated various cross-user models to reduce the calibration burden and compared wrist- and forearm-based models. Eight different machine learning architectures were evaluated across 33 users, using varying amounts of data from the end user. Main results. A temporal Convolutional Network-Bidirectional Long Short-Term Memory (TCN-BiLSTM) architecture, applied for the first time to EMG classification, was found to significantly (p<0.05) outperform other tested machine learning architectures. An Inter-Day Feature Set (IDFS) combined with Z-score normalization achieved the best performance when classifying five gestures (plus a rest class) using either wrist or forearm EMG. Consistent with other recent results, wrist EMG consistently outperformed forearm EMG in all analyses, including within- and across-user comparisons (p<0.05). In cross-user models, wrist EMG demonstrated a zero-shot performance of 78.2% compared to 71.6% for forearm EMG (p<0.05). Introducing one calibration repetition from the end user increased one-shot performance of wrist EMG to 91.6%, compared to 86.9% for forearm EMG (p<0.05). Adding further training repetitions boosted wrist EMG performance to 98.3%, compared to 97.4% for forearm EMG. Significance. These findings provide new evidence supporting the viability of wrist EMG for cross-user HGR models that generalize to new users with minimal calibration, suggesting promising potential for its broader adoption in wearable devices.
Botros et al. (Thu,) studied this question.