This is a guiding template for short-term exploratory projects. Many sections are optional either in their presence or scope and can be excluded based on the project at hand.
What we're exploring: Simple approaches for detecting faulty EEG recording segments.
Currently expected approach: * Compute multi-taper spectogram over 30s windows. * Extract individual PSD curves from the 30s windows. * Check whether this curve is an outlier compared to a collection of curves that were well-recorded.
Specifically: * Use the stages dataset to collect the “well-recorded” collection of PSD curves. * Learn some embedding based on this data. * Embed our recordings, and then manually try to find clusters of our outliers.
If we can then assign a probability of a window being an outlier, this can be used downstream for the majority voting of the U-Sleep. Thus, segments that e.g. lose contact over night may be ignored, and only well recorded ones are considered.
Important Note: * In the current multi-band EEG recording device, the derived is based on electrodes that are also used within the EEG channels. * Instead of trying to find “faulty EOG”, we can thus indirectly identify electrodes within the EEG.
Should we proceed?: Yes/No/Maybe
If yes, what needs to happen?