Emotion detection from EEG using transfer learning.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Electroencephalography (EEG)-based emotion detection faces challenges due to limited data.
• Transfer learning offers a potential solution for data scarcity in complex machine learning tasks.

Purpose of the Study:

• To develop an effective EEG-based emotion detection system using transfer learning and novel feature combinations.
• To evaluate the performance of the proposed method in both subject-dependent and subject-independent scenarios.

Main Methods:

• Utilized Resnet50 as the base model for transfer learning.
• Developed a novel feature representation using Mean Phase Coherence (MPC) and Magnitude Squared Coherence (MSC) in image matrices.
• Incorporated Differential Entropy (DE) features into the diagonal of the image matrix.
• Employed 10-fold cross-validation for subject-dependent accuracy and leave-one-subject-out (LOSO) for subject-independent accuracy on the SEED EEG dataset (3 classes).

Main Results:

• Achieved 93.1% accuracy in subject-dependent emotion classification.
• Obtained 71.6% accuracy in subject-independent emotion classification.
• Both accuracies significantly exceed chance level for 3-class classification.

Conclusions:

• The combination of MPC and MSC features shows significant promise for EEG-based emotion classification.
• The transfer learning approach effectively addresses data limitations in EEG emotion detection.
• Future work may explore data augmentation, enhanced classifiers, and improved feature engineering.