Noninvasive brain-computer interfaces (BCIs) assist paralyzed patients by providing access to the world without requiring surgical intervention. While the performance of noninvasive BCI is hindered by long training times and variable user proficiency, it may be improved by deep learning methods, such as convolutional neural networks (CovNets). Prior work has suggested that the application of deep learning to EEG signals collected over the motor cortex during motor imagery based BCI increases classification accuracy in standard sensorimotor rhythm (SMR) BCI datasets. It remains to be seen whether these improvements can generalize to practical scenarios such as continuous control tasks (as opposed to prior work reporting one classification per trial), or whether valuable information remains latent outside of the motor cortex (as no prior work has compared full scalp coverage to motor only electrode montages). Here we report that deep learning methods significantly increase offline classification accuracy on an independent, large, and longitudinal online motor imagery BCI dataset with up to 4-classes and continuous 2-dimensional feedback. Improvements in classification accuracy were found to negatively correlate with initial online BCI performance, suggesting deep learning methods preferentially benefit BCI participants who need it most. The CovNets also significantly increased the information transfer rate (ITR) of the BCI system: They produced a two-fold increase in ITR without loss in classification accuracy when comparing CovNet models trained with full scalp EEG coverage to the traditional motor cortex specific decoding. Our results suggest that a variety of neural biomarkers useful for BCI, including those outside the motor cortex, can be detected through deep learning methods.
bioRxiv Subject Collection: Neuroscience