Affective computing for emotion identification using dual-stage filtered multi-channel EEG signals

The dual-stage correlation and instantaneous frequency (CIF) thresholding approach for retrieval of noise-free desired frequency band of EEG signal is proposed for affective emotion identification task. Initially, the raw electroencephalogram (EEG) signals are breakdown applying the empirical mode decomposition technique to produce intrinsic mode functions (IMFs). The noisy IMFs are eliminated by applying correlation thresholding. Secondly, these noise-free EEG signals are divided into several modes using a non-linear chirp variational mode decomposition approach to retrieve desired frequency bands (4-30Hz) by applying the IF-based filtering method on the modes. The power spectral densities extracted from filtered modes are fed to ML-based classifiers to classify emotions into arousal, valence, and dominance groups. This study also shows the efficacy of ensemble ML (EML): random forest (RF) and bagging over conventional ML (CML): support vector machine and logistic regression classifiers. The RF reported the highest average F1-scores using 10-fold cross-validation for arousal, valence, and dominance are 83.99%,75.94%, and 88.86% respectively. Similarly, the respective average accuracies of two-EML are~1.47%, ~1.27%, and~0.3% higher compared to two-CML classifiers. To summarize, the proposed CIF-based filtering approach is useful for affective emotion identification under the framework of EML classifiers.