Good morning,
Colleagues and I are working on a deep learning algorithm for sleep staging from PPG. We used the MESA and CFS datasets, as well as other datasets from other sources, and all of those featured a consistent and/or coherent performance using the same pipeline.
However, when using the STAGES dataset, we are meeting great difficulty to reach the expected range of performance, and this performance varies across collection site (Bogan - BOGN folder - especially features a terrible performance compared to the others).
So my questions are:
- What differences can we expect between STAGES and datasets such as MESA or CFS in terms of PPG quality, sleep scoring, maybe synchronization or any particularities that might have escaped our understanding, that might explain what we observe ?
- Same question across collection site within the STAGES dataset.
Thank you for your time !
Thanks for checking out the site. We (the NSRR team) haven't done much work with the STAGES dataset. MESA and CFS were centrally scored at an academic research center. STAGES, on the other hand, relied upon the clinical sleep scoring groups at each of the sites. Perhaps that explains some of the difference?
References:
- Notes/images we received about STAGES PSG collection: https://sleepdata.org/datasets/stages/files/original/Site%20Specific%20PSG%20Notes%20and%20Information
- Old thread where I briefly discussed the STAGES PSG scoring: https://sleepdata.org/forum/stages-ground-truth-hypnogram/
Thank you for your response,
I doubt all the difference we see is due to that, although probably some is. Maybe the extent to which subjects suffer from sleep disorders also contributes to it. It works too well for a severe synchronization problem, and that should not happen as we used the edf files datetime and the csv timestamps to align PPG to sleep stages.
We will continue to investigate that problem on our side.
Cheers !