Decoding neural signals for direction in brain-computer interface using echo state networks and readouts

Abstract

Brain-computer interface (BCI) for movement control using the electroencephalogram (EEG) has been intensively investigated. However, there are limitations of EEG-BCI studies. Most of BCI paradigms based on sensory and movement process in a brain. In example of motor imagery paradigm, BCI systems based on mapping imagery movement of body parts (e.g., left or right hands) to movement directions (left or right directional movement of a machine or cursor) are less intuitive and much inconvenient due to complex training procedures, direct decoding methods for detecting user intention about movement directions are critically required. Another limitation is conventional BCI decoder required EEG system with high sampling rate and numbers of channels. These high specification of EEG recording system not suitable portable and daily usage of BCI. To solve these problem, we investigated the echo state networks and Gaussian readout as a novel BCI decoder. The echo state networks have universal approximation property with simple linear readouts. Therefore, well-designed readouts of the echo state networks can decode the any neural signals with simple learning rules. Readouts of the proposed decoder have Gaussian property in terms of neural code of direction selectivity in a brain. Because the echo state networks based on cortical column model (e.g. microcircuit model), Gaussian readouts of the echo state network can model the neural code of direction selectivity. To investigate performance of our proposed decoder, total sixteen healthy subjects participate the EEG-BCI task and recorded the EEG signals during they thought the intent to move with eight directions using EEG system having low sampling rate and low channels. The decoder accessed the classification performance of eight directions of stimulus in ten-fold cross validation and compared the performance outcome with that of a conventional machine learning method and deep learning methods. We showed that this decoding method successfully classified eight directions of intended movement (approximately 95% of an accuracy) and but best performance of conventional machine learning and deep learning method are approximately 50%. In theorem of the echo state networks, if readouts of the echo state well-designed for specific task, decoding directions, in this case, the decoder can apply to any type of neural signals. Therefore, we also tested the decoding performance of this method with four ECoG subjects with reaching-and-grasping task in execution task and imagination task. In ECoG study, locations and numbers of electrodes were limited in human study case because ECoG systems requires a surgical procedure. These limitation makes difficulty of decode the user intention and decrease performance of BCI. The decoder also successfully decodes directions of ECoG signals and show over the 0.7 correlation between arm movement and decoding trajectory. These results show that the echo state network and Gaussian readouts can be a useful decoding method to directly read user intention of movement directions even using an inexpensive, portable EEG system and many case of ECoG systems.