Back to projects

EMO-CARE — Subject-Independent EEG Emotion Recognition

View resource →

  • PyTorch
  • EEG
  • emotion recognition
  • deep learning

What it is

A deep-learning framework for EEG emotion recognition that generalizes to unseen subjects — the hard, deployment-relevant setting.

It uses 16 parallel Temporal Convolutional Network streams (kernel sizes 3–33, dilations 1/2/4) to capture emotional dynamics across timescales, then an 8-head self-attention over the 16 scale tokens to weight which temporal scales matter. Evaluated rigorously with Leave-One-Subject-Out cross-validation on SEED, SEED-V, and DREAMER (94.3% / 91.2% / 84.3%), with ~15 ms inference suitable for real-time use.

Stack

PyTorch 2.0 · multi-scale TCNs · self-attention · LOSO with statistical validation

Links