Xdawn denoising of binary labelled data. Avoiding overfitting.

💬 Support & Discussions
,
1

Hi all,

python 3.8.13
mne 1.6.1

I’m looking to do some temporal decoding of a simple reaction task, classifying two groups. Before this I’d like to apply Xdawn as a denoising step.

My understanding is that if I were to feed my labelled epoch data through Xdawn with n_components=2, Xdawn would fit the spatial filters to differentiate the two groups, not to remove noise.

To avoid this I used the following approach:

# Create a copy of the target epochs
nolabel_epochs = epochs.copy()

# Create uniform labels for all epochs
nolabel_epochs.event_id = {'uniform':0}
nolabel_epochs.events[:,-1] = 0

# Estimates signal covariance from raw
signal_cov = mne.compute_raw_covariance(raw)

# Create Xdawn instance with n=1
xd = Xdawn(n_components=1, signal_cov=signal_cov)

# Fit Xdawn on uniform labelled epochs
xd.fit(nolabel_epochs)

# Apply Xdawn to labelled target epochs
epochs_denoised = xd.apply(epochs)

The denoised epochs see the expected result, which is an increase in ERP. I wanted to double check that this wouldn’t result in any overfitting behaviour, i.e. applying spatial filtering based on group labels.

Thank you in advance,
Tom

1 Like
2

Interesting idea !

If you then contrast two conditions within epochs I think it’s ok. To me it’s a bit like when you do a beamformer and you contrast conditions in source space. The beamformer is typically estimated from the full epochs like you do here.

Alex

2 Likes