mrueschman

We did create ahi_a0h3a at a later date. Prior to that, a variable like rdi3p was commonly used as the primary AHI indicator in SHHS, though this variable required that apneas also have an associated >=3% desaturation (like the hypopneas). The meaning of "RDI" has also shifted a bit over time, so the variable name and its approach for including apneas had become a bit antiquated.

Just to give you a sense of how the scoring program works and how the different AHIs are computed, here's another description. The study is scored once. You cited the rules about hypopneas requiring an associated >=2% desaturation. The scoring program (Profusion) would then give us different variables ("components" for AHI calculations) representing counts of specific event types at specific desaturation/arousal criteria. For instance, we might get something like the following in an individual sleep study.

1. 50 total hypopneas (>=2% desaturation criteria as based on scoring manual)
2. 40 hypopneas with >=3% desaturation (subset of #1)
3. 25 hypopneas with >=4% desaturation (subset of #2)

Thus, if I want a "total AHI" (all hypopneas) I might use #1, if I want ahi_a0h3 I am going to use #2 (for the hypopnea component), and if I want ahi_a0h4 I am going to include #3.

Joachim,

I'm not aware of any projects/publications that looked to re-score parts of the SHHS dataset. The ahi_X variables, like ahi_a0h3a, were derived from the counts of apneas and hypopneas from the original scoring. These were created to mimic AHIs from other cohorts where we normally use "all apneas" and "hypopneas with >=X% oxygen desaturation" as the AHI numerator. The scoring software output counts of respiratory events that met different desaturation thresholds and with/without arousals factored in. You can see this by looking at the variables that comprise the ahi_a0h3a calculation.

Mike

Do you mean variables like the following? https://sleepdata.org/datasets/shhs/variables?search=shift

I opened the dataset and it looks like hstg2t1p is equal to stg2t1p/(time_bed/60), so it looks like time in bed is the denominator.

Could you shed more light on how many subjects you have found that are affected with the misalignment issue? Did you do a test across all 1,800+ subjects with overlapping PSG/actigraphy, or did you identify the problematic subjects above in a more piecemeal fashion? Originally you said you were "noticing many subjects with temporal misalignment".

Thanks!

Thanks again. This is still on our radar - I am trying to see if someone else on our team can take a closer look to gain a better understanding of the issue.

Thanks! How about these IDs? These subjects all used the same actigraphy device and the studies below occurred in chronological order, with some other subjects in between who did not have accompanying PSG data. Between Groups 1 and 2 I believe the actigraphy device may have gone to the manufacturer for maintenance/repairs. Between Groups 2 and 3 there was a daylight saving time shift. I am curious to see if there is a point in time where the watch shifts from being aligned to misaligned, so to speak.

Group 1

• 3931
• 1374
• 3529

Group 2

• 2828
• 1227
• 3819

Group 3

• 5793
• 2820
• 5303 (already done above)
• 6318 (already done above)

jongguri,

Could you check for this issue and post your plots for MESA ID 5303? This subject used the same actigraphy device as your 6318 example. 5303 was completed a week before 6318.

Thanks!

We will take a look at the examples you provided in more detail and report back.

Hey jongguri,

Thanks for diving into the MESA PSG and actigraphy data and posting this question. PSG and actigraphy were scored independent of one another, so some incongruity may be expected when overlapping the two data sources. The PSG and actigraphy devices were meant to be initialized and setup on the same computer such that the internal timekeeping would match.

Everything I see about subject 4892 looks "OK" from my end. I double-checked with the PSG scorer that the time for sleep onset (9:17 p.m.) is set correctly. The device was set to record between 9:00 p.m. and 7:00 a.m. The subject is very still (based on the activity counts) for some time before 9:17 - maybe they were lying in bed before actually trying to fall asleep. We see this a lot, which adds some guesswork to determining where sleep actually begins when reviewing the actigraphy data.

Please let me know if you find any other questionable disagreements when using the overlap file.

Mike

Hi biggull,

You need to have an approved data request to access most datasets.

Click "Request data access" next to the dataset(s) of interest to start the data request process. Thanks!