eeg-emotion-recognition

EEG Emotion Recognition project, experiment on SEED (SEED-IV), DEAP dataset

Database

• DEAP:

  • 32 participants
  • 40 one-minute videos
  • 40 channels (32 first channels are EEG)
  • 4 unique emotion state: arousal, valence, dominance, liking
  • After watching each video, each person was rating emotion point, ranging from 1.0 to 9.0 for each state. We applied our proposal method for binary classification for each state. Ex: low/high arousal, low/high valence, .v.v.
  • The continuous rating is thresholded at 4.5. Therefore, if the rating was higher or equal to threshold, the emotion of this video was classified as high emotion, and vice versa.
  • Metadata:
    • Sampling rate: 128Hz

• SEED-IV:

  • 15 participants
  • 3 sessions, 1 session ~ 24 trials (24 videos) = 24 * 3 == 72 unique videos
  • 62 EEG channels
  • 4 emotions: happy, sad, neutral, fear
  • For each session, 24 videos showed for each person and EEG signal are collected with the 62-channels ESI NeuroScan System.
  • Different trials (videos) have different samples because the length’s film clips are not the same.
  • Metadata:
    • Sampling rate: 200Hz

Some common pitfalls:

• Train-test-split scenarios.
• Curse of dimensionality: various number of custom features → decrease model’s acc

Train-test-split scenarios:

• Random-train-test
• Subject-dependent: video-based. Ex: 40 videos, 30 for training, 10 for testing
• Subject-independent (a.k.a Cross-subject & leave-one-subject-out): 15 people, train on 14, test on the left one.

Preprocessing and feature extraction:

• Bandpass filter

  • DEAP: 4.0 - 45.0 Hz
  • SEED: 1.0 - 50.0 Hz

• Artifact removal

• Segment signal:

  • DEAP (1s) - 128 samples / segment
  • SEED (4s) - 800 samples / segment

• Feature smoothing:

  • LDA: linear dynamical system
  • Moving average

• Feature extraction:

  • PSD (FFT) - STFT
  • DE - Different Entropy
  • Wavelet Transform (DWT, CWT)
  • Ensemble Wavelet Transform (entropy; statistic: median, mean, std, var, ..; crossing; …)
  • Ensemble features extraction (PSD, ensemble wavelet, hjorth, entropy, ...):

• Feature selection:

  • ANOVA
  • RFE (most reference), ...

Model

- CNN-1D, CNN-2D
- LSTM (apply on raw data and feature extracted)
- SVM
- RF
- Gradient Boosting

Result

• WIP