Unsupervised Decoding of Multi-Finger Forces Using Neuronal Discharge Information with Muscle Co-Activations

Accurate finger force prediction is essential for intuitive human-machine interactions. Various studies have attempted to develop reliable decoders for multi-finger force prediction for the control of assistive robotic hands. However, most approaches were implemented in a supervised manner, i.e., data labels (e.g., finger forces) were needed to train or refine models, which might not be appropriate in certain situations, particularly in the cases of individuals with an arm amputation. In addition, previous studies have not addressed interference from the co-activations of unintended finger muscles. However, finger co-activations occur naturally in daily activities. Therefore, we developed an unsupervised neural-drive approach for simultaneous and continuous multi-finger force predictions, considering finger muscle co-activations instead of avoiding them. To this end, we collected high-density surface electromyogram (sEMG) signals from the forearm extensor muscle during single-and multi-finger isometric contraction tasks. Motor units (MUs) were extracted from sEMG signals of the single-finger tasks. Considering the different contribution of each MU, we assigned weights to MUs based on the firing statistics of the MUs for the target finger across trials. Due to the co-activation effect, where MUs from other fingers may influence the force of the target finger, we introduced an MU sharing procedure to incorporate these MUs. Compared with the supervised sEMG-amplitude methods, our approach demonstrated superior force prediction performance, as evidenced by a higher R²(0.72 ± 0.11 vs. 0.64 ± 0.073) and a lower root mean square error (5.95 ± 1.43 % MVC vs. 7.47 ± 1.81 % MVC). Our approach has the potential to enable intuitive neural-machine interfaces, allowing a wide range of human-machine system applications.