Normal Inverse Gaussian Features for EEG-Based Automatic Emotion Recognition

Electroencephalography (EEG)-based emotion recognition is crucial in the domain of human–computer interaction (HCI), which gained significant attention in recent years. However, the nonstationarity and chaotic nature of the EEG signals pose challenges and restrict the state-of-the-art techniques from precisely identifying distinct emotional states from the EEG data and hence offer limited emotion recognition performance. To capture the underlying nonlinear characteristics of EEG, this study employed a novel local mean decomposition (LMD) algorithm, which decomposes EEG signals into product functions (PFs). Further PFs are modeled by normal inverse Gaussian (NIG) probability density function (pdf) parameters. Thus, these pdf features are fed to an optimized Adaboost classifier developed with the help of a cross-validation (CV) approach. The novelty of the work lies in the NIG modeling of LMD-domain PFs to identify specific emotions from the EEG signals. The significance of the NIG parameters is illustrated by qualitative, pictorial, and statistical analyses. To assess the efficiency of the proposed approach, intensive experiments are conducted on open-source datasets, Shanghai Jiao Tong University (SJTU) Emotion EEG Dataset (SEED), SEED-IV, and Database for Emotion Analysis of Physiological Signals (DEAP). The emotion recognition performance is evaluated in terms of heat maps, receiver operating characteristics (ROCs), and accuracy. The proposed emotion recognition system outperformed the state-of-art methods and achieved a maximum accuracy of 97.3%, 98%, and 98.6% with the CV approach and 93.23%, 94.87%, and 95.58% for the cross-subject validation (CSV) approach using DEAP, SEED, and SEED-IV datasets, respectively.