Detection of Error-Related Potentials in Stroke Patients from EEG Using an Artificial Neural Network

Abstract

Error-related potentials (ErrPs) have been proposed as a means for improving brain-computer interface (BCI) performance by either correcting an incorrect action performed by the BCI or label data for continuous adaptation of the BCI to improve the performance. The latter approach could be relevant within stroke rehabilitation where BCI calibration time could be minimized by using a generalized classifier that is continuously being individualized throughout the rehabilitation session. This may be achieved if data are correctly labelled. Therefore, the aims of this study were: (1) classify single-trial ErrPs produced by individuals with stroke, (2) investigate test-retest reliability, and (3) compare different classifier calibration schemes with different classification methods (artificial neural network, ANN, and linear discriminant analysis, LDA) with waveform features as input for meaningful physiological interpretability. Twenty-five individuals with stroke operated a sham BCI on two separate days where they attempted to perform a movement after which they received feedback (error/correct) while continuous EEG was recorded. The EEG was divided into epochs: ErrPs and NonErrPs. The epochs were classified with a multi-layer perceptron ANN based on temporal features or the entire epoch. Additionally, the features were classified with shrinkage LDA. The features were waveforms of the ErrPs and NonErrPs from the sensorimotor cortex to improve the explainability and interpretation of the output of the classifiers. Three calibration schemes were tested: within-day, between-day, and across-participant. Using within-day calibration, 90% of the data were correctly classified with the entire epoch as input to the ANN; it decreased to 86% and 69% when using temporal features as input to ANN and LDA, respectively. There was poor test-retest reliability between the two days, and the other calibration schemes led to accuracies in the range of 63-72% with LDA performing the best. There was no association between the individuals' impairment level and classification accuracies. The results show that ErrPs can be classified in individuals with stroke, but that user- and session-specific calibration is needed for optimal ErrP decoding with this approach. The use of ErrP/NonErrP waveform features makes it possible to have a physiological meaningful interpretation of the output of the classifiers. The results may have implications for labelling data continuously in BCIs for stroke rehabilitation and thus potentially improve the BCI performance.

Keywords: brain–computer interface; calibration; classifier interpretation; error-related potentials; neurorehabilitation; stroke.

Figure 1

Timeline of a single trial. The participant is cued to perform a specific movement after which the movement is attempted immediately after the presentation of the cue. The sham feedback was provided to the participant three seconds after the attempted movement.

Figure 2

Top: Grand average across 25 participants for the ErrP and NonErrP epochs for recording day 1 and 2. The shaded area indicates the standard error across participants and the solid line is the mean. Bottom: Single trials for participant 1 for recording day 1 and 2. The shaded area indicates the standard deviation, and the solid line is the mean across the trials. Time ‘0 s’ is the onset of the presentation of the feedback. The signals in all trials are from the electrode position FCz. The vertical dotted lines indicate the part of the signal that was used for the classification analyses.

Figure 3

Classification accuracies associated with different calibration schemes using features or the entire epoch as input for the artificial neural network and classification of features using linear discriminant analysis. The bars represent the mean ± standard error across the participants. For the between-day calibration, the three bars under “Day 1” represent training the classifier on data from recording day 1 and testing on recording day 2 and vice versa for the three bars under “Day 2”.