EEGdenoiseNet

[scott-huberty]

This is the model I breifly mentioned in our last meeting: https://de.mathworks.com/help/signal/ug/ocular-artifact-removal-from-electroencephalogram-signals-using-deep-learning-regression.html

As compared to our WIP, There are some similarities (they use an LSTM architecture), and some differences (IIRC, they simulated EEG data with and without EOG artifact, and trained a regression model to take the the EEG signal (with EOG artifact) as input, and minimize the loss between the predicted clean EEG signal and the actual "clean" version of the EEG signal (i.e. the label).

Just leaving this here in case it is relevant to us down the road.

[christian-oreilly]

Just adding additional models we could test for improved performances:

• EEG-Conformer: https://github.com/eeyhsong/EEG-Conformer
• EEGPT: https://openreview.net/forum?id=lvS2b8CjG5
• EEGFormer: https://arxiv.org/html/2401.10278v1

I think the best would be to make an interface that allows selecting the model and wrapping up models into a class or something like that. Then, users could also use the library with new models.

[scott-huberty]

Sure, I think that is a nice idea

[scott-huberty]

Maybe it's possible to tinker with the ShallowFBCSPNet or Deep4Net CNN's from BrainDecode for the purposes of EOG regression:

https://braindecode.org/stable/auto_examples/model_building/plot_regression.html#sphx-glr-auto-examples-model-building-plot-regression-py