Emotion recognition based on time-frequency distribution of EEG signals using multivariate synchrosqueezing transform

This paper investigates the feasibility of using time-frequency (TF) representation of EEG signals for emotional state recognition. A recent and advanced TF analyzing method, multivariate synchrosqueezing transform (MSST) is adopted as a feature extraction method due to multi-channel signal processing and compact component localization capabilities. First, the 32 participants' EEG recordings from DEAP emotional EEG database are analyzed using MSST to reveal oscillations. Second, independent component analysis (ICA), and feature selection are applied to reduce the high dimensional 2D TF distribution without losing distinctive component information in the 2D image. Thus, only one method for feature extraction using MSST is performed to analyze time, and frequency-domain properties of the EEG signals instead of using some signal analyzing combinations (e.g., power spectral density, energy in bands, Hjorth parameters, statistical values, and time differences etc.). As well, the TF-domain reduction performance of ICA is compared to non-negative matrix factorization (NMF) to discuss the accuracy levels of high/low arousal, and high/low valence emotional state recognition. The proposed MSST-ICA feature extraction approach yields up to correct rates of 82.11%, and 82.03% for arousal, and valence state recognition using artificial neural network. The performances of the MSST and ICA are compared with Wigner-Ville distribution (WVD) and NMF to investigate the effects of TF distributions as feature set with reduction techniques on emotion recognition.