Important nuance here is the difference b/w the parent poster's "accurate enough to make insulin dosing decisions" criteria (extremely difficult+risky)

vs

"useful enough signal for general population / those at risk of type 2 diabetes / those with diabetes who don't require insulin treatment / etc" (much larger market for Apple w/ greatly reduced requirements)

^^ This - the ECG & Blood Oxygen levels are "good enough" for the public to consume and track, but perhaps not quite at the level of monitors used in medical settings to make care decisions.

Yeah I don't know about ECG but I checked my (non apple) watch with a real finger-clip SpO2 from the doctor and while it's pretty accurate under highly controlled situations (sitting still and positioned perfectly) it's really lousy under real-world conditions, moving around etc.

And I didn't just check the bog standard 98% measurement but also tried to hold my breath etc.

I can't speak for the Apple Watch, but the O2/BPM readings on some other watches are often dropped/delayed with the operating system taking priority. I was trying to use the heart rate data for a pet project and found its practically useless, on Samsung's watches its just enough data to show a few fancy numbers and thats about it.

On the contrary, the Apple Watch ECG is quite accurate, not quite enough to replace the system the doctor makes me wear, but enough that apple can get a general idea of any defects and tell you to go to the doctor. The "regular" hear rate sensor isn't anywhere near good enough to be taken seriously, but its leaps ahead my previous samsung.

I dream of the day where my Apple Watch can replace the myriad of annoying wires and sensors I have to wear to monitor myself. Even then, the Apple Watch has greatly helped me out when I've fallen and need assistance, and I've gifted several to my family just in case.

ECG is a relatively easy thing to measure, so that single axis they measure is quite accurate. But the conclusions one can make from that is limited (you basically take a video from a single point). Doctors use 12 axes instead to get a real-time, 3D vector of the heart.

Yeah I heard Apple Watches have better sensors, I wonder though. I can't use them myself at all as I can't use iOS (too limited for my requirements).

I'm using a $70 Amazfit right now but previously I had a Samsung, neither is very good.

To be honest, “real finger-clip” sensors are similarly fuzzy in their results - after all they work the exact same way. Maybe using percentiles is not ideal as it gives a sense of more accuracy than what makes sense.

They don't work quite the same way. The finger clips shine through the entire finger. The illuminator is on one side and the sensor on the other. The wrist-based sensors rely on reflections through the skin which is a lot more iffy.

Google tried its glucose monitoring contact lens and failed, the program was discontinued in 2018.

https://en.wikipedia.org/wiki/Google_Contact_Lens

For something that has to be monitored very closely, nothing beats an actual blood sample. And to get a blood sample, it can't be non-invasive.

to say google discontinued a project after a short period of time does not exactly speak to the level of difficulty or if that difficulty was even the reasoning for the discontinuing. someone at G could have just gotten bored or moved to a different department, or a new darling employee stole focus, or whatever other reasonings we've heard when G projects are shuttered.

> the program was discontinued in 2018

To be fair, the Verily folks did design the algorithm that predicts where your blood sugar will be in X minutes in Dexcom CGMs, which lives on. Dexcom now owns the patent.

> nothing beats an actual blood sample

Technically, interstitial fluid can get you pretty dang close, which is considered minimally invasive.

Nothing beats a blood sample, but interstitial fluid (which is invasive kind of) is good enough for insulin dosing. Most CGMs (Dexcom, Freestyle, etc) work this way. They don't sample blood, they sample the fluid between cells.

Replying to my own comment: I found a 2021 paper that goes over all the different technologies being considered, and the wide variety is frankly staggering. Table 10 lists ten different principles under consideration (fluorescence, spectroscopy, etc.) with their various pros/cons. Not ELI5, but as close as you might get:

https://iopscience.iop.org/article/10.1088/2516-1091/abe6f8#...

Seems like spectroscopy seems particularly promising, and according to the article, that's the approach Apple is taking. (Makes sense for a watch.)

And from the paper, the challenges with spectroscopy are signal interference -- scattering, water absorption, and distinguishing between glucose and other sugar components. So it ultimately sounds like a fiendishly complex signal processing problem. Which is exactly the kind of problem that is very very hard yet maybe not impossible.

That is the type of problem that takes a company of Apple’s size to tackle.

The IMHO gold standard book to get an understanding of the history of non-invasive glucose monitoring and the different technological approaches is "The Pursuit of Noninvasive Glucose: Hunting the Deceitful Turkey" by John Smith: https://www.researchgate.net/profile/John-Smith-241/publicat...

We - at DiaMonTech - working on a solution with mid-infrared spectroscopy and showed good accuracy in a published paper in the "Journal of Diabetes Science and Technology". We plan another clinical trial this year, but yes, this took some time. We started in 2015 and still have some work to do.