UCLA

Brain-computer interface (BCI) technology can return the ability to communicate to those suffering from neurological disease or paralysis. Common BCI systems require their user to constantly attend an external stimulus, which can be exhausting and may not be physiologically possible for some users. BCIs designed around sensorimotor rhythms, like those produced during the imagined movement of a body part, remove the need for a constant external stimulus. A BCI designed around motor imagery data would need to be able to determine the user’s intent in the shortest amount of time possible. This study looked at the effect of shortening the training sequence of data on a proven deep learning model’s (EEGNet) performance. We show that cutting the amount of data passed to the model by half did not negatively affect the model’s performance and suggest this is due to a decay in feature fidelity in the latter half of the data. Additionally, we designed a hybrid convolutional-recurrent neural network (CRNN) optimized to classify short sequences of data. This new model is able to achieve similar results as EEGNet across subjects at 44.98 percent and 49.37 percent accuracies, respectively. Our results suggest a motor imagery-based BCI could determine the user’s intent with as little as two seconds of data and pioneer a novel deep learning model that could perform even faster.