Eirikur Agustsson's wiki

Sleep Project Sleep Project Log

Below is the initial project plan.

SLEEP

Emmanuel Mignot and Simon Warby study the structure of sleep. They measure electrically the activity of the brain (EEG), they measure heartbeat and breathing as well. They analyze these signals to detect different `phases' of sleep. There are both `global' patterns (e.g. frequency of the EEG signal) and `local' events (e.g. `spindles' and `K-complexes') that allow the classification of sleep into a small number of phases (e.g. `deep sleep', `phase 2 sleep', `REM sleep' etc). Mignot and Warby suspect that the structure of sleep, as well as the details of the shape of the local events, are in part genetically determined and would like to explore this hypothesis.

In order to do so, they would like to automate the analysis of sleep. They suggest starting from detection and classification of spindles and K-complexes. They already have a collaboration with some people at Stanford to work on spindles. Something that they are missing, and appears to be very important at this point, is careful hand-annotation of spindle and K-complex events in a few hundred patients. This is because they need to create a training and test set for automated local event detectors.

Your goal will be to: (a) use techniques developed in our lab for annotation of large datasets via the Amazon Mechanical Turk to annotate `by hand' large numbers of sleep recordings using the Amazon Mechanical Turk (AMT) mechanism put together by Peter Welinder. Classify each event by degree of certainty: `clear case', `likely', `perhaps', `most likely not'.

(b) using the annotated data, understand whether the local events (spindles and K-complexes) come in different `types' or perhaps in a continuum, and whether there are patients whose spindles and K-complexes are different from the rest. Also, whether within the same patient there is much variability or not.

(c) compare annotations performed by experts (trained sleep technicians) with respect to annotations performed by regular people who sign up for AMT work. See if there is a difference in expertise and overall performance. (You will see statistical techniques for doing this in Peter Welinder's papers).

(d) If time is left, try your hand at building automated spindle and K-complex detectors. Spindle detectors are possibly less useful because someone at Stanford is already working on this problem, but you may get lucky and get something better / different from what they do.

MY WORK

Here are the main steps for your work, as I see them. You will work out the details with Peter. Many aspects of the project will change as we make progress and discuss the outcomes, but here you have a trace of what is likely to happen.

0. Study Peter's papers. You will need to understand both the statistical techniques, the AMT mechanisms, as well as the Python and other machinery that allows us to post tasks, and collect and analyze the data automatically. This was all built by Peter and he will be your guide.

1. Meet Simon Warby together with Peter. Simon will likely fly from Stanford in early july carrying with him a few DVDs containing the data of a few 100s nights of sleep (different patients). Simon will (a) educate you on what sleep is all about and what signals they collect, (b) what is the goal of this project, a selection of the relevant literature supporting the hypothesis that there is a genetic basis for sleep structure, (c) teach you how to read into Matlab / Python the data that he is giving you.

2. Learn to generate JPG or other image files from snippets of EEG data, so that sleep data may be presented to AMT workers as images.

3. Discuss with Simon what are the best instructions that may be given to AMT workers for classification of spindles and K-complexes.

4. Work out how to recruit and recognize sleep technicians as part of the AMT study.

5. Learn how to use Peter's system to collect data from AMT workers.

5. Run a first trial with 10 nights of sleep to see if the machinery works. Analyze those data. Discuss with Mignot, Warbly and myself. Decide how to adjust the `big' study.

6. Run major annotation with AMT. Analyze those data. Discuss what has been collected.

7. Consider similarities and differences of local events in different patients.

8. (Perhaps) develop your own local event detector

9. Write up your results.