[This article belongs to Volume - 38, Issue - 05]

Symbolic Regression Based Approach for Emotion Classification from EEG Signals

Symbolic Regression (SR) is an Evolutionary Algorithm (EA) variant, which aims to generate mathematical equations for various tasks such as classification and regression. SR uses Genetic Algorithm (GA) to generate mathematical equations that best fit input data which have the least error. SR is notably useful in cases where features in a data set are known to have distinct mathematical relationships among each other and its output. This research aims to assess the efficacy of SR on biomedical data, such as electroencephalogram (EEG) signals, to classify human emotions. As the data is purely numerical, SR can be used to generate mathematical equations which relate the features extracted from the EEG signals. Ongoing research in the field of emotion classification using biomedical signals shows that several deep learning techniques such as Convolutional Neural Nets and a combination of Neural Nets have been used. However, these models or methods do not provide further insights such as which frequency band provides more relations pertaining to emotions. In other words, the current state of the art lacks explainability. The proposed method provides further information on the dominant features and insights into frequency bands through the equations generated by Symbolic Regression. The flow of the proposed pipeline is as follows: the signals are sampled down, and noise is removed for computations of the discrete wavelet transform, which is used to extract two main features: differential entropy and energy spectrum. These features are conveyed to a symbolic classifier for the generation of equations and classification. Results show that SR achieved comparable results with an accuracy of 85 % on the SJTU (Shanghai Jiao Tong University, China) Emotion EEG Dataset (SEED) data set, along with an added advantage of being explainable due to its generated equations. This method assesses the capability of evolutionary algorithms for broader EEG-based classification tasks.