Application of Gated Recurrent Unit in Electroencephalogram (EEG)-Based Mental State Classification

The classification of mental states based on electroencephalogram (EEG) recordings has recently gained significant interest in cognitive monitoring and human-computer interaction fields. Due to high signal variability and sensitivity to noise, correct classification is still tricky, even with advances in the analysis of EEG signals. Among deep learning models, Gated Recurrent Unit (GRU) models have established great potential for sequential EEG data analysis. The applications of the GRUs are less reviewed in tasks concerning classification cases of mental states compared to hybrid and convolutional models. Based on this paper, we will propose a method for developing a model based on the GRU network trained with raw EEG data in the classification tasks of mental states of concentration and relaxed conditions. We analyzed 400 EEG recordings taken from 10 subjects within a controlled environment and collected using the Muse EEG Headband. The mean, standard deviation, skewness, kurtosis, power spectral density, zero-crossing rate, and root mean square were extracted as statistical features from the raw EEG data. After parameter tuning, the GRU-based model achieved an excellent average accuracy value of 95.94% and also yielded precision, recall, and F1-scores within the range of 0.95 to 0.97 over 5-fold cross-validation. This shows that GRU works well in classifying mental states based on the EEG data.