From sequential to simultaneous prosthetic control: Decoding simultaneous finger movements from individual ground truth EMG patterns.

Myoelectric bionic limbs hold the promise of restoring functionality and improving life quality for people with amputation. With recent advances in surgical reconstruction, which created additional signal sites for myoelectric control, intuitively controlling all fingers of a prosthetic hand became a possibility. To fully utilize a multiple degree of freedom (DoF) bionic hand, the fingers need to be controllable both individually and simultaneously. However, training algorithms to decode motor intent typically requires large sets of labeled data. This data requirement grows combinatorically with each additional DoF, complicating the training process for multi-DoF control. Here, we evaluated a method to create labeled simultaneous data from linearly combining individual movement data. We found that a classifier trained on such artificial data performed equivalently in decoding 3 DoF real-time finger movement to a classifier trained on ground truth data. However, its effectiveness diminishes with more complex tasks, i.e., 5 DoF finger control. In both cases, linearly combining individual movements decreased the time to acquire labeled data to train the classifier by up to 85%.