چكيده به لاتين
Decoding continuous movement-related parameters like position and speed is valuable both for understanding the neuroscience of the movement and for designing effective Brain Computer Interfaces (BCIs). In this study, we have proposed a novel deep neural network structure for high accuracy decoding of continuous movement and continuous applied force from LFP signal. Our proposed method consists of a convolutional and a recurrent block. The convolutional section extracts the optimal features from the LFP data and the recurrent section estimates movement-related parameters form the features created in the previous block. To evaluate the network, we used two separate datasets. In the first one, LFP signal was used to decode continuous movement in 2D space and in the other one, LFP signal was used for decoding continuous applied force. In both datasets, the proposed network was able outperform multiple classical approaches including linear models, Kalman filter and another deep neural network designed for continuous decoding. Correlation coefficient (r) and coefficient of determination (R2) were used for validating the decoded signal. Average correlation coefficient for decoded position, speed and applied force were 0.89, 0.75 and 0.75 and the coefficient of determination were 0.75, 0.55, 0.54 respectively. In addition to accurate decoding of movement-related parameters, the network automatically performed all the feature extraction pipeline from raw LFP signal. This property makes it easier to generalize the proposed method for other subjects or datasets. Furthermore, the unique structure of the proposed network enabled us to investigate physiological properties which lead to accurate prediction of movement parameters from LFP signal. We are hoping that with the recent advancements in developing high accuracy, deep learning-based neural decoding models, the dream of designing effective BCI systems will become a reality.
