Know Labs narrows accuracy gap between its noninvasive glucose sensor and current CGMs

Know Labs’ work in 2023 has been focused on two core areas, both with a goal of ushering its noninvasive glucose sensor technology closer to FDA review. Not only did the company debut in June the first physical prototype of the device that’ll house the technology, but it has also released a slew of studies in recent months showing progressive improvements in the system’s glucose-tracking abilities.

Those reported improvements have been dramatic, with the sensor’s accuracy nearly doubling between a proof-of-concept study published in February and the company’s latest study results, which it shared Wednesday.

Know Labs is racing to produce the first-ever FDA-cleared glucose monitor that collects readings completely noninvasively, merely by being held against a user’s skin. Its device relies on what the company has dubbed its Body-Radio Frequency Identification (Bio-RFID) sensors, which send radio waves through the skin to measure molecular signatures in the blood, and on a light gradient-boosting machine (LightGBM) artificial intelligence model that analyzes the sensors’ readings to calculate glucose levels.

The accuracy of Know Labs’ sensor—like other glucose monitors—is reported as a measure of mean absolute relative difference, or MARD, which assesses the size of the gap between a device’s blood sugar readings and those gathered by a reference device—so, the lower the MARD, the better.

With recent technological advances, the MARD ceiling for an accurate CGM has fallen to about 10%; the newest offerings from major diabetes devicemakers Dexcom and Abbott are closer to the 8% mark.

Know Labs is rapidly whittling down its own MARD levels. In the study results (PDF) published this week, the company said its technology had been able to predict blood glucose levels with a MARD of just under 11.3% when compared to readings from a Dexcom G6 CGM. That’s a far cry from the nearly 20% difference it reported in February, and it’s a further improvement from the 12.9% MARD it calculated just a few weeks ago, in a late May study.

The new research was published as a preprint and hasn’t yet undergone formal peer review, but was reviewed by Know Labs’ own scientific advisory board, which formed in March with a trio of experts.

Know Labs attributed the improved accuracy to new data preprocessing techniques used in the system’s analysis and to the fact that it fed the system more than double the amount of user data than before.

In the study, the researchers trained the LightGBM machine learning model using more than 3,300 reference device values collected over the course of 330 hours from 13 participants. The May study, in contrast, relied only on about 1,500 collected observations taken from five people.

The company said in its announcement about the study results that it plans to spend the rest of the year continuing to collect even more real-world glucose readings and using them to further refine the LightGBM AI. Some of those data will come from Know Labs’ Generation 1 prototype device itself, which is being tested every day, per the company, and is “expected to generate tens of billions of data observations to process.”