Machine learning for prediction of sleep stage

Hello,

I have two sets of physiologic recordings. One is a scored sleep study and the other is recordings from the brain. I have the sleep study findings loaded in to Matlab so that I can index using it, and the recordings from the brain analyzed using pwelch and plotted with spectrogram. When I overlay the two plots, there is an increase in the frequency band 15-30Hz seen in the brain recordings whenever the patient is in an "awake" period.

I have another 9 sets of brain recordings from other patients, but their respective sleep studies have yet to be analyzed. While I wait for that to happen, I want to predict what that output will look like solely from the brain recordings.

I believe I should be able to take the first half of the brain recordings from the patient with a scored sleep study, train a machine learning algorithm on this, and then test it on the second half. The data sets are very large. Fs = 1024, brain recording = [33000000, 1].

Any advice on how to proceed is much appreciated, happy to provide more information regarding state of the data as well. Thank you

[jgg]

The problem is that, while your data is very large, it's time series in nature, so you might not have enough information.

A first cut would be to try something like general linear regression: try to predict whether someone is awake or asleep based on the data. You could do this in a bunch of ways, but I think that would be my first try.

It would help if you could show us a plot of what your data looks like.

[Star Strider]

I did something similar to this in the early 1990s, but with specific neuropsychology laboratory tasks reflected by EEG activity pattern, rather than sleep stage. (We were doing pilot workload studies.) We used principal component analysis and linear discriminant analysis with the BMDP statistical routines, and FORTRAN.

I would first do a principal component analysis on your short-time-Fourier-transform EEG frequency data. That will make it easier for you to select out the frequencies-of-interest, since such studies are data-intensive. Then use that information (that will yield the specific frequencies most prominently associated with each sleep stage) as input to your classifier.

I would be surprised if there isn’t already a substantial literature on this that could help you in your research. If there is not, search PubMed, specifically for authors Pfürtscheller and Wolpert (individually, although they publish together). They have continued research in the area of brain-actuated control and task determination (I have not), and their approaches to data acquisition and analysis, even if not specifically related to sleep staging, will be valuable.

I’m not listing this as an Answer because it isn’t one.

[Greg Heath]

PLSREGRESS tends to be more helpful for ranking classification inputs.

Hope this helps.

Greg

[atek]

Thank you everyone for your input - it is much appreciated. @Star Strider, nice to see you again! @jgg, please let me know if this is useful. Left is the signal unprocessed, right is signal processed to remove values outside of 7*SD

[Star Strider]

You, too!

I wouldn’t eliminate the 'outliers'. Include them, and let your spectrogram (STFT) sort them, unless you know they are sampling artefacts. (These can occur in some instances, usually in recording pulse stimuli or from environmental contamination, but rarely in the physiological signals themselves.) You have to take a closer look at the signal to determine that (although the spike on the right in the left plot could qualify if you know that what generated it was not the EEG). Otherwise, determining what are EEG signals and what are artefacts isn’t possible from the resolution of the image you attached. It would also help to have units (sec, µV) on the axes.

[atek]

Ok, makes sense. The very large signal on the right is certainly artifact, I know the amplifier was overloaded at that point. I will adjust the axis label, thank you.

My plan is to watch a couple of the Matlab webinars and then try your suggestions. I will update with specifics if something arises