Research On Channel Selection And Explainable Emotion Classification Model Based On EEG Signals

In recent years,with the increasing pressure of life,people suffer more and more from mental stress as well as the potential risk of mental illness.The proportion of mental illness in the Chinese population has surpassed heart disease and cancer,and has become a huge burden on China’s medical system.In this case,people have conducted extensive research on mental illness in recent years.Emotion is an important wind vane of human mental state,which is closely related to people’s mental health.Therefore,emotion recognition is an important research hotspot.Emotion recognition method based on electroencephalogram has become one of the important research directions at present.There are many research achievements in the direction of EEG emotion recognition.However,there are still problems to be solved.(1)Using a large number of channels in research without channel selection brings information redundancy and computational burden,which is harmful to turning research results into portable products.(2)In order to pursue the accuracy of classification,the interpretability of the model is neglected,making the model like a “black box”,which is no help to perform targeted adjustment on the model.To solve the two problems,the following works have been done in this paper.(1)To solve the channel selection problem,an emotion-related channel selection algorithm based on brain functional connection has been proposed in this paper.The algorithm measures the synergistic effect intensity of the brain region corresponding to the channels in the period by calculating the correlation between the channels,then use the statistical test method to measure the value of each channel in discriminating emotions based on a large number of functional connectivity samples,which helps selecting channels strongly related to emotions.The proposed algorithm is compatible with the brain function view in brain science,and the results are statistically significant.(2)To solve the problem of the interpretability of model,a new emotion classification model SENet is proposed which is the combination of DNN and SE Block.During the training process,SENet is able to focus on valid information of the samples in the way of adaptive weighting,which makes it more precise in emotion classification and interpretable by showing the feature channel weight distribution obtained by the model.Experiments have been designed to test and verify the proposed algorithms based on DEAP dataset.Samples for emotion classification have been constructed on both the selected channel collection and the full channel collection to perform a comparative experiment on emotion classification to compare the classification rate of them.The results shows that these two kind of classification rates are at the same level,which proved the effectiveness of the proposed algorithm.To verify the proposed SENet algorithm,A further comparative experiment on emotion classification based on the former one has been conducted to test the classification rate of SENet and other methods.The results shows that SENet gains 0.6385 in average after model ensemble,which outperforms the algorithms proposed by other papers.At the same time,the feature weight distribution as well as the channel weight distribution has been exported,which have been discussed and analyzed to verify the interpretability of the model.