Shallow Convolutional Network Excel for Classifying Motor Imagery EEG in BCI Applications

Abstract

Many studies applying Brain-Computer Interfaces (BCIs) based on Motor Imagery (MI) tasks for rehabilitation have demonstrated the important role of detecting the Event-Related Desynchronization (ERD) to recognize the user's motor intention. Nowadays, the development of MI-based BCI approaches without or very few calibration stages session-by-session for different days or weeks is still an open and emergent scope. In this work, a new scheme is proposed by applying Convolutional Neural Networks (CNN) for MI classification, using an end-to-end Shallow architecture that contains two convolutional layers for temporal and spatial feature extraction. We hypothesize that a BCI designed for capturing event-related desynchronization/synchronization (ERD/ERS) at the CNN input, with an adequate network design, may enhance the MI classification with less calibration stage. The proposed system using the same architecture was tested on three public datasets through multiple experiments, including both subject-specific and nonsubject-specific training. Comparable and also superior results with respect to the state-of-the-art were obtained. On subjects whose EEG data were never used in the training process, our scheme also achieved promising results with respect to existing non-subject-specific BCIs, which is other progress to facilitate the clinical application.

Monte Carlo Dropout for Uncertainty Estimation and Motor Imagery Classification

“…Two popular public datasets from BCI Competition IV, specifically datasets 2a and 2b [29][30][31][32][33][34][35][36][37], were used.…”

“…A total of two sessions on different days was performed for each subject, completing a total of 288 trials per session. The first session was used here for training, and the other session was employed for testing, as done in previous works [29][30][31][32][33][34].…”

“…Each subject completed a total of five sessions summing 720 trials (first two sessions with 120 trials each, and the last three of 160 trials each). Then, the first three sessions were used for training, while the last two sessions were used for testing, as done in previous works [29,[35][36][37].…”

“…For this reason, as shown in Figure 1, the EEG signals are band-pass filtered here in a frequency range from 4 to 38 Hz through a Butterworth filter, aiming to preserve the ERD and ERS rhythms, rejecting undesirable physiological and non-physiological artifacts. Next, each filtered EEG trial x is standardized by applying the electrode-wise exponential moving standardization with a decay factor of 0.999 [29,31], according to the following equations:…”

“…The objective of this study is to propose methods for accuracy improvement and uncertainty analysis in a deep neural network scheme for MI-based BCIs in order to obtain robust predictions, and therefore, more reliable responses of robotic devices during motor rehabilitation. An existing Shallow Convolutional Neural Network (SCNN) [29] scheme was used in our study to test the main contributions of this paper, such as SCNN combined with Monte Carlo dropout (SCNN-MCD) to classify motor imagery EEG in BCI applications and estimate the model uncertainty. We investigate with SCNN-MCD how different uncertainty measures correlated with the predictive accuracy and also introduce a threshold method that rejects EEG inputs that produce predictions of very high uncertainty in cases when we should not rely on the prediction and minimizes the error rate of the classifier.…”