Comment by idealmedtech | Hacker News Reader

Comment by idealmedtech | Hacker News Reader

Parent Full thread

Hi, this is actually my field of expertise! IdealMedTech is a clinical glucose control startup targeting inpatient glucose control.

Having an MARD (mean absolute relative difference, or in math terms, E[|g-r|/r], g is measured glucose, r is reference glucose) of below 15% is actually considered very good for glucose monitoring.

We did a study that involved modelling different levels of sensor error, and you actually start to see diminishing returns on control once your MARD passes below 10%, with 15% being a good target.

Most sensors on the market that are considered best in class (eg Dexcom G6) have guaranteed MARD below 15%, usually sitting right at 10%. Getting lower than that is very difficult for interstitial glucose, especially using the classical glucose oxidase sensing mechanism.

There are other ways to sense glucose using, for example, Raman spectroscopy, and there have been companies successful in employing these techniques (OptiScan comes to mind), but I'm not yet convinced by the capacity of spectroscopy to deliver accurate results in the presence of interference from drugs and perfusion issues.

> Most sensors on the market that are considered best in class (eg Dexcom G6) have guaranteed MARD

Disclaimer: I work for Dexcom.

A few times every year, there's a ton of hype about some new noninvasive device, and the newbies get a bit worried, but the old timers just shrug it off.

We (humanity) will literally be able to regrow new pancreases with stem cells before we have accurate noninvasive blood glucose devices. Even Tim Cook with unlimited resources at his disposal admitted it's a much harder than they thought.

That said, I think there's a large market for people without diabetes who are interested in their blood glucose and who may not need super-accurate readings. For example, as a long distance runner, it'd be super helpful to know if my blood sugar is trending down before I bottom out, so I can know when to eat a gel pack, etc. It would also be useful for people trying to lose weight to learn how their eating habits and exercise affect their blood sugar levels.

Noninvasive devices may be better for this group of people who don't need the devices to be accurate enough to make decisions about insulin-dosing.

"A few times every year, there's a ton of hype about some new noninvasive device, and the newbies get a bit worried, but the old timers just shrug it off"

This.

But a buddy of mine may surprise you soon. And I worked in the field too.

And a lot of people are waiting to be (surprised). It super duper helpful for t2dm folks to know the "trend" - which food works better, which food doesn't, is the medication working, how much workout is required to drop the blood sugar levels, the dawn phenomenon etc.. with literally billions of people getting affected with t2d it would be a game changer. Still hoping for apple watch xx to have that feature.

Thank you for your work. My 9 year old son has worn a Dexcom for the past 3 years (Medtronic prior to that) and it's a godsend to be able to pair it with Sugarmate.io so we get phone calls in the middle of the night if his BG goes low.

In addition, I imagine confounds which are patient/location specific can be calibrated out to some extent, by comparing against finger-stick while fasting, and longer-term A1C.

Following a chain of references from the OP:

study of wearable: https://cb195ec0-5419-4138-91fc-7288f69116ec.filesusr.com/ug... whitepaper announcement: https://medium.com/know-labs/know-labs-releases-white-paper-... whitepaper: https://docs.google.com/document/d/e/2PACX-1vQXqlmKwPzhZIn9t...

The whitepaper gives actual prep methods for tissue phantoms. It shows a clean correlation of ca 1.72GHz transmission to glucose. Presumably altered parameters in tissue phantoms could help roughly characterize confounds.

This... brought back some memory...to average or not to average... C8 Medisensors achieved better weekly MARD in the early 201*. But the problem was, on a bad day, no one knows why it was bad..., so no way to alert the user (unless the reading was way off). Even though on "average", performance looks nice, but a bad day the data can be very problematic for a user, and the user has not way to trust the device based on MARD...

This is a great point, and precisely why we're using two sensors and a simple voting algorithm in our first in man trial (3+ sensors and associated algorithms were deemed too risky by FDA for initial studies).

Not sure what sensors you are using... why FDA thinks more sensors carry higher risk?

It's not the number of sensors, it's the novelty of the voting algorithm required to fuse those sensors that gives them pause. FDA likes to be _VERY_ conservative when it comes to trying new things. We're trialing our control software, which is already novel to them, so anything we can do that's not brand new helps reduce our risk.

More information should help make more informed devision, if the algorithm can not clearly explain it to FDA or FDA cannot understand that is the case... something seems wrong here... all the best wishes and also encourage more exploration.

Reply on news.ycombinator.com