Exploring EEG Features in Cross-Subject Emotion Recognition.

Area of Science:

• Neuroscience
• Machine Learning
• Affective Computing

Background:

• Recognizing emotions from electroencephalography (EEG) across different individuals is challenging due to feature generalizability issues.
• Previous research explored limited EEG features, yielding inconsistent findings on cross-subject emotion recognition.
• A systematic investigation of diverse EEG features is essential for advancing this field.

Purpose of the Study:

• To comprehensively investigate the effectiveness of various EEG features for cross-subject emotion recognition.
• To compare the performance of linear and non-linear EEG features across different datasets and methodologies.
• To identify robust EEG features and brain activity patterns indicative of emotions across subjects.

Main Methods:

• Evaluated 18 linear and non-linear EEG features on the DEAP and SEED datasets.
• Employed a support vector machine (SVM) classifier with a leave-one-subject-out cross-validation strategy.
• Utilized automatic and manual feature selection techniques to identify optimal feature subsets.

Main Results:

• Achieved highest accuracies of 59.06% (AUC=0.605) on DEAP and 83.33% (AUC=0.904) on SEED using automatic feature selection.
• Manual feature selection identified Hjorth mobility in the beta rhythm as a highly effective feature for cross-subject emotion recognition.
• Correlation analysis provided insights into feature implications for differentiating emotions across subjects.

Conclusions:

• The study demonstrates the feasibility of identifying robust EEG features for cross-subject emotion recognition.
• Findings underscore the importance of systematically exploring a wide range of EEG features for improved emotion detection.
• This research contributes to understanding brain-based emotion recognition and its potential applications.