They say "90% accuracy" for apnea - definitely not high enough for a diagnosis. The study's lead investigator is talking about this as a screening test, not as a firm diagnostic.

And given the involvement of UCSF in the study, I'd take this pretty seriously - just make sure to pay attention to the claims the study is actually making, and not the science-journalism hype.

It’s good enough to raise awareness and I believe sleep apnea could use that.

> Machines to diagnose apnea have lots of sensors including electrodes, "chest-band" pulsometers, and "saturationmeters".

It seems like a fallacy to claim that since diagnosis required complicated equipment in the past, that it will be necessary going forward. Machine learning is going to make tons of current diagnostic equipment look archaic in comparison.

> I have asked doctors about capabilities of smart devices for apnea control or detection, all I got was laughs.

Medical doctors are rarely up to date on the latest technology. I wouldn't be surprised if my doctor has never heard the term machine learning before in his life.

Also, their livelihood kind of depends on this technology not being available. Just food for thought.

It seems a fallacy to say that "machine learning" will replace the sensors needed to produce the data required to make a diagnosis. Sure the sensors can always be in a smaller, simpler form factor, but that's engineering the sensors specifically. You need data to make a diagnosis, and the magic of machine learning can't replace raw data coming from sensors.

I see this kind of confusion a lot, where "machine learning" is mistakenly used to mean obsoleting sensor data instead of obsoleting the hardcoded processing of said data.

> It seems a fallacy to say that "machine learning" will replace the sensors needed to produce the data required to make a diagnosis.

I agree. Good thing I didn't say that :) My comment implied exactly what you wrote, that the machinery will get smaller, and machine learning will be able to make more accurate predictions from a more limited data set. Although, if we consider that the Apple Watch will be strapped to your wrist for 16 hours a day, it might actually turn out to be an expanded data set, but a smaller number of sources.

> I see this kind of confusion a lot, where "machine learning" is mistakenly used to mean obsoleting sensor data instead of obsoleting the hardcoded processing of said data.

I believe that's a misinterpretation on your part. I haven't seen anyone claim that machine learning doesn't require data from sensors.

> It seems like a fallacy to claim that since diagnosis required complicated equipment in the past, that it will be necessary going forward. Machine learning is going to make tons of current diagnostic equipment look archaic in comparison.

As someone sais below "ML is not going to replace raw input from sensors."

> Medical doctors are rarely up to date on the latest technology. I wouldn't be surprised if my doctor has never heard the term machine learning before in his life.

Current dignose equipament is big, invasive and unconfortable; often they require you to stay one night at the hospital.

There are tests programs about using those devices to replace current tech (obvious reasons: less costs, less invasive tests, subject owns the hardware…)

This kind of procedures are not used because "doctors are old and non-tech people", it's because it's not working.

Sure, apps and smart devices could replace some day those devices but they need more and accurate sensors. Extrapolation of data from a good-enough pulse rate sensor it's not a replacement.

While I understand and mostly agree with your general sentiment, I think you're downplaying the extent of how valuable the machine learning is in this process. The reason why we need all these sensors and machines currently is because we need a reasonable picture of how all that data works together so that a human being can look at that picture as a whole and make some deductions.

This isn't the first time or the last time that computers have been able to find patterns in much simpler measurements simply because they're not human. A similar example, although not quite as advanced, is the ability for a computer to extract sound information from a black and white video. Computers have been able to recreate sounds from behind double-paned glass by analyzing vibrations captured through videos. Humans have had to rely on various arrays of lasers, sonar, and other directional instruments to get 1/10th of the accuracy that a computer algorithm has been able to achieve from a simple, low-quality video camera. The point is that more machines and sensors doesn't always equal a better diagnosis. Better analysis of existing sensors and tech, even if it seems to be lower fidelity, can actually yield more accurate results.

> As someone sais below "ML is not going to replace raw input from sensors."

You've misinterpreted my comment. I didn't claim sensors weren't required. I said that it's a fallacy to claim that the current diagnostic equipment is required. There are sensors on smart watches. Cardiogram is using data from those sensors.

> This kind of procedures are not used because "doctors are old and non-tech people", it's because it's not working.

It's not working yet (at least not enough to make a 100% accurate diagnosis). No one claimed it was. I hate to break it to you, but the equipment currently used in sleep studies isn't 100% accurate either. Neither is the human interpretation made by sleep doctors. However, to say that doctors are laughing it off makes those doctors look stupid, not Cardiogram. Based on the studies they've run so far, it's not hard to envision a future where smart watches and apps will replace a large percentage of doctors whose main source of income is sleep studies.

> Sure, apps and smart devices could replace some day those devices but they need more and accurate sensors. Extrapolation of data from a good-enough pulse rate sensor it's not a replacement.

Your claim seems to be that sleep doctors need input from a large number of sensors, so it can't be possible to make a diagnosis from a single heart rate sensor. It would appear that Cardiogram has presented evidence to the contrary:

https://blog.cardiogr.am/screening-for-hypertension-and-slee...

If you disagree, make a specific criticism of their evidence.

Nobody is saying to use this to replace going to the doctors.

It's for you to just download the app, try it for a night, and see the doctor if the result concern you. It doesn't have to be hundreds percent accurate.

Are you against teaching people to do self check-up of breast cancer, too? Because I'm sure medical process of checking for breast cancer requires more complicate tools than just feeling your breast with a hand.

Apnea has obvious signs (fatigue, day sleep, depression, …) that are the equivalent of the "self check-up" in your example of breast cancer.

If you suspect that you have apnea episodes you should go to a doctor.

80% of people with sleep apnea don't know they have it. Unfortunately, the signs are not obvious to most.

I'm against breast self exams, and so is the US Preventive Services Task Force. The evidence indicates there's no benefit but instead the practice calls harm (in anxiety, overdiagnosis, unneeded biopsies, etc.

https://www.uspreventiveservicestaskforce.org/Page/Document/...

> I have so many doubts about this. As you should. "New study suggests" is basically the same as saying "new study didn't find". It's clickbait or a PR piece.