Deep Residual Shrinkage Networks for Adaptive Classification of EEG Error-Related Potentials

Error related potentials(ErrPs) can reflect the brain's response to errors. In recent years, it has been used to study the neural mechanisms of human cognition, such as error detection and conflict monitoring. Although deep learning methods are commonly used for EEG analysis, they are rarely used for adaptive classification of ErrPs. In this paper, the deep residual shrinkage networks(DRSN) based on ResNets is used for adaptive classification of ErrPs signals. DRSN aims to improve the feature learning ability of the model under high noise or complex background. Its main innovation is introducing residual shrinkage units into ResNets, which integrates a soft threshold into ResNets as a nonlinear conversion layer, and automatically determines the threshold so as to reduce the influence of noise in EEG signals. Therefore, DRSN can also save more preprocessing steps related to denoising and simplify the signal processing of ErrPs. In addition, DRSN also introduces focal loss to reduce the impact of sample imbalance on classification results. The public dataset BNCI2020 dataset22 is used to verify the adaptive classification performance of DRSN. The classification results of DRSN greatly outperforms other existing adaptive classification methods. In the case of completely blind adaptive classification, the average classification accuracy of DRSN reaches 73.08%. Under the condition that the original data is just preprocessed by common average reference(CAR), the average adaptive classification accuracy of DRSN reaches 79.65%, which is about 60% higher than other adaptive classification methods using CAR, and in this case, DRSN significantly increase the classification accuracy by about 7% of subjects with weak ErrPs quality. In addition, compared with ResNets, DRSN is less time-consuming after introducing the residual shrinkage module for improving the efficiency.