EEG Enhancement by Auto DNNs with Regularization of Spatial Feature Loss

Abstract

Electroencephalography (EEG) can be applied in medical diagnosis forecasts via using Brain-Computer Interface (BCI) technology. EEG signals are low voltage signals that are susceptible to various types of noise such as 50 Hz power frequency, noise between the electrodes and the skin and so on. In this work, an enhancement method for EEG data based on a deep neural network (DNN) architecture search method in which the spatial feature loss acts as a regularizer while training the end-to-end network for best noise removal effect is proposed. The proposed system realizes noise reduction by using DNNs, which employs an alternative objective function combining spatial feature loss with time-domain feature loss. The spatial feature can be obtained by Common Spatial Pattern (CSP) algorithm. Experimental results show that auto DNNs with regularization of spatial feature loss can efficiently eliminate the simulated noise in EEG data and makes the mean square error between predicted values and real values as small as 0.06. In addition, the proposed objective function outperforms objective function with single time-domain feature loss. Meanwhile, the number of parameters in auto DNNs is obviously less than other models by 81.7% to 94.2% and also less when using proposed objective function than not use it by 28.6%. These results demonstrate that proposed DNNs based method can reduce parameters and computation. Therefore the proposed method is promising for the wearable application and embedded scenarios.