The goal of this project was to train a convolutional neural network to classify EEG signals as belonging to one of two individuals performing the same task (counting squares of a chosen color).
Models were trained for 7 randomly selected pairs of subjects, using 2 different data folds per pairing and the model architecture described below.
UC Berkeley-Biosense Synchronized Brainwave Dataset by Chuang, Merrill, Maillart & Students of the UC Berkeley Spring 2015 MIDS Immersion Class (2015).
The dataset contains 1-second windows of single-channel EEG readings sampled at 512 Hz. The signals were recorded using a consumer-grade headset with a single-electrode while the subjects responded to different types of stimuli.
To increase the number of available training samples, a cropped training strategy inspired by Schirrmeister et al. (2017) was used. For each subject, all samples recorded during the "colorRound" stimulus were concatenated in chronological order. New 1-second samples were then generated as sliding time windows, such that the start and end of each sample i+1 was shifted by 375 ms in relation to sample i.
Figure 1
Data Augmentation by Cropping
Note. The first 3 original samples recorded for subject 1 during the color task are shown in the first subplot; followed by subplots displaying the first 3 training samples generated by the cropping process (offset=375 ms). Dotted pink lines in the first subplot mark the original sample borders. EEG values have been scaled to fall in the [-0.5, 0.5] range.
Models with larger kernel sizes in the convolutional layers were found to be more powerful but have extremely high variance. It was found that more stable results could be achieved by using a "tribrid" architecure which allowed for a low-level representation of the input to flow through three different processing paths. These paths had the same overall architecture - four 2-layer convolutional blocks followed by two blocks of dense layers - but different kernel sizes in the convolutional layers.
The outputs of the processing paths were concatenated, and additive (self) attention was applied, before they were merged by summation.
The idea was for the paths to "collaborate" by focusing on different features in the input, but taking into account their counterparts' representations.
Figure 2
Architecture of the Model
Table 1
Results for Subject Pairings
| metric | subj. 3 & 7 | subj. 1 & 27 | subj. 8 & 4 | subj. 7 & 28 | subj. 7 & 1 | subj. 1 & 28 | subj. 6 & 18 |
|---|---|---|---|---|---|---|---|
| binary_accuracy | 0.722200 | 0.828900 | 0.881600 | 0.802600 | 0.684200 | 0.631600 | 0.947400 |
| precision | 0.750800 | 0.842400 | 0.836600 | 0.756000 | 0.751300 | 0.625000 | 0.928600 |
| recall | 0.710500 | 0.815800 | 0.947400 | 0.894700 | 0.552600 | 0.657900 | 0.973700 |
Note. Results for each subject pairing were averaged over folds.
notebooks:
-
EEG_data_inspection.ipynb: exploratory data analysis
-
EEG_subject_classification: contains the code for data augmentation and the classifier
plots:
-
data_augmentation_example.png
-
model_architecture_plot.png
-
pandas 2.2.2
-
numpy 1.26.4
-
matplotlib 3.10.0
-
keras 3.8.0
-
tensorflow 2.18.0
Chuang, J., Merrill, N., Maillart, T., & Students of the UC Berkeley Spring 2015 MIDS Immersion Class. (2015, May 9). Synchronized Brainwave Recordings from a Group Presented with a Common Audio-Visual Stimulus. UC Berkeley Spring 2015 MIDS Immersion Class.
Schirrmeister, R. T., Springenberg, J. T, Fiederer, L. D. J., Glasstetter, M., Eggensperger, K., Tangermann, M., Hutter, F., Burgard, W., & Ball, T. (2017). Deep Learning With Convolutional Neural Networks for EEG Decoding and Visualization. Human Brain Mapping, 38(5), 5391–5420.