Neural Noise- Recurrent and convolutional neural networks have been used to classify EEG signals for guided imagery and mental workload detection (Nature, March 2025).1 1
Weekly enrichment (2026-07-20)
- The article (Scientific Reports, 2025) compares four deep-learning models — a 1D-CNN, an LSTM, a 1D-CNN-LSTM hybrid, and EEGNet (a 2D-CNN) — for binary classification of two mental states: Guided Imagery relaxation versus a mental-workload task.2
- The cohort was 26 right-handed males aged 19–24; EEG was recorded on a 256-channel EGI dense-array amplifier, the 10–12 minute window of each recording was band-pass filtered 1–45 Hz and split into 1-second segments, yielding 3,120 samples (1,560 per class).2
- Two electrode configurations were compared: the full 256 channels (FULL-256) and a 26-electrode central-parietal/occipital “cognitive” subset (COGN-26).2
- Six-fold cross-validation accuracies (FULL-256 / COGN-26) were: EEGNet 0.7615 / 0.7646, LSTM 0.7250 / 0.6833, 1D-CNN 0.7682 / 0.8094, and 1D-CNN-LSTM hybrid 0.7726 / 0.7556.2
- The 1D-CNN was best on COGN-26 (80.9% accuracy, F1 ≈ 0.78, precision ≈ 0.90), the hybrid was best on FULL-256 (77.3%), and the single-layer LSTM performed worst throughout.2
- In Leave-One-Subject-Out validation the 1D-CNN reached 78.87% on COGN-26; because the 26 cognitive electrodes matched or beat the full 256-channel set, the authors conclude the full array adds little value.2
- Classification was end-to-end on raw signals (no manual feature extraction or artifact removal), preprocessed with the MNE-Python and pyprep/PREP pipeline, with the CNN trained at a learning rate of 1e-5.2
- For context, an open preprint of the same work is on arXiv (2405.16901); the authors’ earlier band-feature GLM reached 90.77% on full recordings and a prior fNIRS-EEG CNN workload study reached ~89%, situating these end-to-end results among comparable EEG workload classifiers.3