I'm one of the co-founders of Cardiogram and involved in this study -- happy to answer questions if folks have them!
(EDIT: heading to bed, but will answer questions in the morning!)
back
6 comments
What input signals do you have access to through the Apple Watch? Is it only Heart Rate output or do you get the actual interbeat intervals? Any chance of getting to the raw ADC output of the photodiode?
I've written a blog about this a while ago [0]. You're limited to what HealthKit allows you to query and the RR intervals are not available. With the release of WatchOS 4 however the SDNN heart rate variability can be queried.
[0] https://www.vanmil.org/heart-rate-variability-and-brain-wave...
Yes. Apple Watch takes measurements every five minutes (no RR intervals yet). There are a few popular watches, like Garmin, that do support RR intervals.
Interesting, thanks!
I didn't see a reference in the TechCrunch article -- is there a technical paper out?
Edit: A presentation is linked, as "study":
Link to presentation: http://www.abstractsonline.com/pp8/#!/4412/presentation/4522...
A followup: Does the neural network provide some extra insight into what part of the waveform/time-series of pulses is able to provide predictive information?
Correct -- the abstract will be included in a special issue of Circulation:
http://circ.ahajournals.org/content/136/Suppl_1/A21042
The study was presented at the American Heart Association's Scientific Sessions earlier today. We included some details beyond what appears in the TechCrunch article in a blog post here: https://blog.cardiogr.am/screening-for-hypertension-and-slee...
I'd expect a more mature version of this work to appear in a medical journal, but that often takes 6-12 extra months.
Couldn't find the measurement process.
What is the common root cause of error? Inaccuracies in measurement or the model classifying data incorrectly?
Sorry for the habitual rant, but I feel the need to point out those are not open articles. Your case is typically something of interest to people who have no reason to have access to medical journal stores and… :(
This thread is not the right place to grab pitchforks, just a reminder open source science was still an on-going effort.
What do you do with the data collected via your app?
Whats the end result of your research?
Did the 70% people you collected data are aware?
The 70% of people were aware—they enrolled in an IRB-approved study with UCSF Health eHeart Study, and went through an explicit informed consent process (as medical research generally requires). We wrote a bit more about the end goal in the last section of the blog post here:
https://blog.cardiogr.am/screening-for-hypertension-and-slee...
What are the false positive and false negative rates for sleep apnea detection? Would it be able to detect sub-clinical symptoms?
Second the other comment on technical paper. Is it available?
The accuracies quoted here are c-statistics, or area under the ROC curve. An example operating point for sleep apnea is 97% specificity with a sensitivity of 52% (compared to a 20% diagnosis rate in today's healthcare system). If a specificity of 82% is acceptable, then we can detect about 75% of sleep apnea.
No technical paper yet (this was a cardiology conference), but we have some follow-up work in submission to an academic artificial intelligence conference, so hopefully we'll have a technical paper with more details out soon.
Is it this https://itunes.apple.com/de/app/cardiogram-for-apple-watch/i... ?
Could you please state, if the app can currently do that?
Do you have plans to commercialize this?
> Do you have plans to commercialize this?
Cardiogram is a venture-backed company.
That's correct. More from our investors:
https://a16z.com/2016/10/20/cardiogram/
https://homebrewvc.tumblr.com/post/152080260668/your-heart-beats-102000-times-a-day-cardiogram