[This review is intended solely for my personal learning]

Paper Info

DOI: 10.1007/978-3-031-05409-9_23
Title: A Swarm Intelligence Approach: Combination of Different EEG-Channel Optimization Techniques to Enhance Emotion Recognition
Authors: Sabahudin Balic, Lukas Kleybolte, and Christian Märtin

Prior Knowledge

  • EEG-based Emotion Recognition: EEG captures electrical brain activity via multiple electrodes. Its high temporal resolution makes it suitable for decoding emotional states in real-time. However, the full 32-channel setup is often computationally expensive and redundant.
  • Swarm Intelligence: Algorithms like Particle Swarm Optimization (PSO), Cuckoo Search (CS), and Grey Wolf Optimizer (GWO) mimic social behavior of animals to solve optimization problems.
  • Feature vs. Channel Selection: Traditional feature selection targets discriminative features across frequency bands, while channel selection focuses on spatially optimizing electrode positions.

Goal

To evaluate and compare different EEG channel selection techniques—both classical and swarm intelligence-based—to optimize emotion classification performance while significantly reducing computation time.

Method

Dataset & Emotion Model

  • DEAP Dataset: 32-channel EEG recordings from 40 subjects watching 32 emotion-evoking music videos.
  • Emotion Labels: Based on Russell’s Circumplex Model, which maps emotions in 2D space using Arousal and Valence dimensions (High/Low quadrants).

Preprocessing & Feature Extraction

  • EEG signals were downsampled to 128 Hz, filtered (4–45 Hz), and segmented into 2s windows (256 points) with 0.125s overlap.
  • FFT was used to extract power across five bands: δ, θ, α, β, γ.
  • Feature set: 160 features (32 channels × 5 bands).

Channel Selection Techniques

  • Classical Methods:
    • PCA: Unsupervised, variance-preserving projection.
    • mRMR: Supervised mutual information-based selection maximizing relevance and minimizing redundancy.
  • Swarm Intelligence Methods:
    • PSO: Simulates bird/fish social movement.
    • CS: Inspired by cuckoo parasitism with Lévy flights for global exploration.
    • GWO: Mimics hierarchical grey wolf hunting behavior. Each method selected the top 20 out of 32 EEG channels.

Classifier

  • Bidirectional LSTM:
    • Four LSTM layers (first bidirectional), interspersed with dropout.
    • Dense layers for binary classification of arousal and valence.
    • Trained for 200 epochs using Adam optimizer.

Results

  • mRMR emerged as the most time-efficient method with minimal accuracy trade-off.
  • Cuckoo Search achieved the highest accuracy, albeit with higher selection time.
  • Reducing to only 10 channels with CS still yielded 91.32% (arousal) and 91.84% (valence), cutting training to ~4h.

Limitations

  • Homogeneous Channel Count: Fixed reduction to 20 channels across all subjects limits the adaptive power of swarm intelligence.
  • Swarm Methods are Costly: CS and GWO, while accurate, introduce 4–7h additional selection time per subject.
  • Lack of Real-time Testing: Results are from offline DEAP dataset; online or real-world deployment was not evaluated.

Reference