Generalizable gesture classification of HDsEMG using volume representations of muscles averaged across multiple individuals

Human hands can perform far more gestures than the number of muscles controlling them, as most gestures result from coordinated combinations of muscle activations and relaxations. This complexity poses a key challenge for human-machine interfaces performing gesture classification based on electromyography (EMG). Rather than identifying all conceivable gestures, it may be simpler to instead identify the activity of the individual muscles which generate a variety of complicated gestures. Here we suggest a three-dimensional model with volume representations of individual digit extensor muscles, averaged across multiple individuals, and evaluate its application and performance in hand gesture classification. Time-domain peaks in high-density surface EMG data from different hand gestures were extracted and localized within the model, from which a gesture classification scheme was generated for both single and multi-label cases. The model was created and tested on a publicly available dataset with 19 participants, leveraging a leave-one-out approach to assess inter-subject generalizability, and multi-label data to assess generalizability to gestures not included in the creation of the model. For single-label classification performance, true positive rates were between 61.9 and 95.1%, with false positive rates between 0 and 24.1%, for different single-digit extensions. The multi-label test demonstrated some degree of generalizability in identifying completely new gesture compositions, while simultaneously maintaining the leave-one-out approach for inter-subject generalizability. A model generated with this approach could be used for gesture classification by anyone, without individual modelling data, with the potential to generalize to any number of gesture compositions.