An EEG-based framework of EMD and CNN for arousal and valence recognition

Emotion plays an important role in shaping our actions, choices, and cognitive functions. Recognizing their significance, electroencephalography (EEG) emerges as a suitable approach for studying and discerning emotions, given its advantages of cost-effectiveness, non-invasiveness, and exceptional temporal resolution.

In the current research, we attempted to introduce a new method according to empirical mode decomposition (EMD) and convolutional neural network (CNN) for emotion recognition. The multichannel intrinsic mode functions (IMFs) extracted through EMD were segmented, and signal power in each EEG segment was calculated as feature of high/low valence and arousal states. In the proposed approach, 17 × 17 images were created from EEG signals in each emotional state according to signal power in one-second segments of extracted IMFs, which were utilized as inputs for the proposed deep learning model based on CNN. The suggested technique is utilized on the DEAP database, consisting of emotional EEG data, and its outcomes are benchmarked against relevant previous studies.

Our approach yielded an accuracy of 91.21% for the random splitting evaluation technique and obtained a mean accuracy of 93.24% through the Leave-One-Subject-Out cross-validation process. The findings showed that our approach outperformed the previous studies which utilized the same EEG database.

Hence, the utilization of EMD in conjunction with CNN offers enhanced analysis capabilities to effectively detect emotions from nonstationary EEG signals due to their nonlinear and nonstationary signal processing attributes.