Emotion recognition from EEG signal enhancing feature map using partial mutual information

The brain signal is the most potent information source to recognize emotion, a fundamental trait of human beings, for providing various emerging personalized courtesies or services instantly to individuals. For emotion recognition (ER), Electroencephalography (EEG) is a preferred brain signal, where the crucial and challenging task is accurately extracting features from complex EEG signals using appropriate computational intelligence or machine learning techniques. Recent ER methods mostly use EEG channel connectivity features to identify the emotion. Specifically, to construct a connectivity feature map (CFM), Pearson correlation coefficient (PCC), mutual information (MI), normalized MI (NMI), and a few other techniques are used. Notably, in the existing ER methods, CFMs are predominantly in the two-dimensional (2D) form, i.e., using the signals from two EEG channels. This study proposes an enhanced CFM that uses partial MI (PMI) by introducing an extra third channel to expose more information and strengthen the feature extraction ability of ER. The proposed technique calculates the PMI-based connectivity features for each pair of EEG channels and presents CFM in 2D and 3D forms. Convolutional Neural Network (CNN) is used to classify emotion using 2D and 3D CFMs. In creating CFMs from EEG signals, rigorous tests have been performed on the DEAP benchmark EEG dataset. As PMI exposed additional information, the enhanced CFM has been found to deliver better ER performances than the one that uses typical MI or NMI, revealing the proposed one outperforming the existing related contemporary methods.