Comprehensive Summary
This paper focuses on improving the quality of data used to train brain-computer interfaces for motor applications such as neuroprostheses. The neural response decoder plays a critical role in classifying ECoG signals as either usable or unusable for training the motor control decoder, forming part of a feedback-based system model. In this setup, ECoG features are simultaneously fed into both the neural response decoder and the motor control decoder, while the neural response decoder is further trained using the motor control decoder’s predictions of the patient’s intended movements. Results show that when the neural response decoder is provided with the target action intended for the motor control decoder, it can more accurately distinguish between correct and incorrect decoding of ECoG signals, though this process is limited to offline use. Nonetheless, higher accuracy can still be achieved when both decoders are pre-trained with data about the target action and the motor control decoder’s predictions are incorporated into the training of the neural response decoder.
Outcomes and Implications
Patient interaction with brain-computer interfaces can often be a tedious and frustrating experience. Patients using BCIs to replace lost functions often compare the devices with original function prior to injury, resulting in impatience, loss of confidence, or rejection of the BCI during the acclimatization or training phase. This training phase and further BCI-patient interaction can be significantly improved with models that use accurate and robust data. The proposed improvements to the Motor Control Decoder system in this paper increase the quality in signals classified by utilizing the Neural Response Decoder. This may have the potential to improve and accelerate the training process that patients undergo to use BCIs. Increased signal quality may also have the potential to improve the accuracy of the BCI’s output, which is this case, can be prosthetic motion.