Comparison of fine-tuned single-source and multi-source approaches to surface electromyogram pattern recognition

Fine-tuning (FT) is an well-establish transfer learning technique that mitigates differences in data distributions of target and source domains by modifying parameters of a neural network pretrained with source data. This technique has been demonstrated to be effective for cross-subject classification using surface electromyograms (sEMGs). However, in the case of biomedical data such as sEMGs, there are two possible approaches to applying FT: (1) single-source approach, which uses multiple source subjects as a large single dataset, and (2) multi-source approach, which uses multiple source subjects as a collection of multiple small datasets. Furthermore, there is arbitrariness in the ensemble strategy for the final output in a multi-source approach. As it is unclear which approach is more valid, we compared the single-source approach and four multi-source approaches (using hard voting, soft voting, weighted voting, and supervised weighted voting) based on three public sEMG datasets: MyoDataset, NinaPro DB2, and NinaPro DB5. We found that the fine-tuned multi-source approaches outperformed the fine-tuned single-source approach on all datasets regardless of the ensemble strategies. In addition, we identified slight but significant disparities between ensemble strategies, but this varied in trend across datasets. This suggests that fine-tuned multi-source approaches are effective and better performance can be achieved with a suitable strategy. Our findings improve the usefulness of FT in sEMG-based human–computer interfaces.