Non-Invasive Glucose Monitoring: Zero Cleared Devices, So Apple and Google Infer Instead

On 21 February 2024 the FDA issued a safety communication telling people not to buy smartwatches or smart rings claiming to measure blood glucose without piercing the skin. None had been authorised. The agency said it had identified dozens of companies and multiple brand names selling them [1] [2].

Two and a half years later that has not changed. No non-invasive glucose device has been cleared or approved.

And yet in August 2026 Google shipped a wearable feature about metabolic health, and in September 2025 Apple shipped one about blood pressure, and both are real, and one of them is FDA cleared.

Nothing about the physics improved. What shipped answers a different question.

The bar the sensor route has to clear

To replace a blood draw or a finger stick you have to be about as good as one. The reference points are the cleared minimally invasive continuous glucose monitors, which report mean absolute relative difference in single digits: 8.2% for Dexcom G7 and 9.2% for FreeStyle Libre 3 on manufacturer figures [4].

Worth noting even here, because it recurs across this whole category: an independent head-to-head study found 8.9% for Libre 3 and 13.6% for Dexcom G7, materially worse than the manufacturer figure for one of them [4]. Vendor numbers and independent numbers diverge even among cleared devices.

Against that bar, the closest published non-invasive result is Know Labs, which reported 11.1% MARD against venous blood draws in March 2024 [3]. That is genuinely close. It was also produced in controlled measurement conditions rather than continuous real-world wear across body positions, glycemic ranges and populations, which is what the company has said its FDA trials are designed to test [3].

A decade, and what it produced

The direct-sensing route is not short of money or talent. It is short of clearances.

Rockley Photonics built photonic integrated circuits for spectroscopic sensing, held publicly known relationships with Apple and Medtronic, went public via SPAC, and ran a human study of 40 diabetic subjects over ten weeks. It filed for Chapter 11 in January 2023, was delisted from the NYSE, and emerged that June with roughly $35 million after a 46-day restructuring [5].

Movano has pursued radio frequency sensing, holds granted patents on RF glucose and blood pressure monitoring, and built what it described as the smallest RF integrated circuit designed for the purpose. It has stated an intention to run a clinical trial toward a 510(k) submission. It has not been cleared [3].

Know Labs has the best published accuracy number in the category and is heading into FDA trials rather than out of them [3].

Apple has worked on non-invasive glucose for years without shipping or, on the public record, submitting anything.

Two routes, not one pivot

What shipped instead of a glucose sensor answers a different question, and it is worth being careful about how that happened.

The direct route asks what is this person’s glucose right now, which requires detecting glucose against water, temperature, skin tone, perfusion and movement. The other route asks is there a pattern in this person’s data consistent with a condition, which requires only that correlated signals carry some information. Resting heart rate, heart rate variability, sleep and activity do carry some, and they come from sensors already on tens of millions of wrists.

Both shipping features below take the inference route. Neither measures the thing in its name. Both report a longitudinal pattern.

Where they differ is which regulatory door they walked through, and that turns out to determine what evidence each company owed the public.

Door one: infer a pattern and get it cleared

Apple’s hypertension notifications were cleared by the FDA on 11 September 2025 and rolled out to Series 9, 10 and 11 and Ultra 2 and 3 [6]. Apple’s own description of what the feature does not do is worth quoting, because it is the architectural move in one sentence: it “does not measure blood pressure directly, and does not surface a blood pressure reading to users.” Instead it “notifies users if optical heart sensor data shows signs of hypertension after 30 days” [7].

The pipeline is inference throughout. A deep learning feature extractor, trained self-supervised on large-scale unlabeled data from the Apple Heart and Movement Study, converts 60-second PPG segments into signal characteristics. Apple notes that this training data “did not include any blood pressure or hypertension data, focusing solely on identifying PPG signal features.” Linear models then score hypertension risk per segment, and results are filtered, averaged over 30 days and compared against a threshold [7].

Apple published the validation. The algorithm was developed using data from more than 100,000 study participants across seven studies, then validated in a pivotal decentralized study enrolling 2,229 participants, of whom 1,863 met usable data requirements. All were adults with no preexisting hypertension diagnosis. The reference standard was each participant’s own 30-day average of home blood pressure readings, categorised under the 2017 ACC/AHA guidelines [7].

The results, with the raw fractions Apple reports:

Read the first two rows together. The FDA cleared a feature that misses about 59% of hypertension overall, and roughly 70% of Stage 1.

That is not a scandal, and understanding why is the most useful thing in this post. Apple did not stumble into that number. It chose it:

A receiver operating characteristic (ROC) curve was generated based on the development data, and an operating point was selected to prioritize high specificity, minimizing false positives, while also notifying a significant portion of people with hypertension.

Door two: infer a trend and stay in wellness

Google’s Insulin Resistance Trends, announced 12 August 2026 as part of Health Guardian, makes the same architectural move. It uses resting heart rate, activity, sleep and related signals to report a direction of travel in metabolic health over multi-week windows [8].

It went through a different door. There is no 510(k) here. The feature sits under general wellness enforcement discretion, which the FDA’s revised guidance of 6 January 2026 widened specifically to cover non-invasive products that estimate or infer physiological parameters, provided they avoid disease references, diagnostic thresholds, clinical alarms and treatment guidance. We covered that shift in where the FDA line sits now.

What that door does not require is a published pivotal study. Google has not published device-level performance for the shipping feature. What exists is the underlying research, the WEAR-ME study of 1,165 participants, which for the configuration a watch actually has, wearables and demographics with no blood panel, reports auROC 0.70, sensitivity 0.60, specificity 0.80 against a HOMA-IR cut-off [9]. We went through those numbers and what they do to predictive value in the insulin resistance post.

The three doors, side by side

The row that matters commercially is the second-to-last one. The row that matters scientifically is the last one.

Doors two and three run on the same physics and the same information ceiling. What separates them is entirely a choice about what claim to make, and therefore what evidence to produce. Apple chose a narrow claim and paid for a pivotal study. Google chose a softer claim and did not have to.

Neither is illegitimate. But a reader cannot tell the difference from the marketing, and only one of them published the number.

What this means if you are building

Do not bet a roadmap on a sensor that has not cleared. The category’s history is announcements, partnerships, patents and pilot studies, followed by more announcements. Clearance is the only event that changes anything, and in non-invasive glucose there has not been one.

Inference ships, with a ceiling. You can build a real feature today on resting heart rate, HRV, sleep and activity. What you cannot do is exceed the information those signals carry, and no amount of model scale adds signal that is not in the input. A 0.70 auROC is what that correlation is currently worth, and the honest product design follows from that number rather than working around it.

Choose your door deliberately, and disclose either way. If you take the wellness route, nothing forces you to publish performance. That is exactly why publishing it is worth something. The teams that will still be trusted in three years are the ones whose numbers were available before anyone asked.

Pattern claims survive scrutiny that value claims do not. “Your blood pressure is 142/91” invites a direct comparison against a cuff, which you will lose. “A pattern in your data over 30 days is consistent with hypertension, see a clinician” is defensible, cleared, and more useful to the person reading it.

A note on sourcing, because this topic attracts nonsense

While researching this piece I was told, by multiple confident-sounding sources, that Apple had activated a non-invasive glucose sensor on the Apple Watch Ultra 3, and that microneedle interstitial fluid arrays are in shipping consumer wearables.

Both are false. Apple shipped hypertension notifications, which use the optical heart sensor and do not measure glucose or blood pressure. Microneedle sensing is real but is research and early-commercial, and the FDA’s January guidance treats minimally invasive sensing as invasive, which puts it outside the wellness route entirely.

This category has more low-quality content than almost any other in health tech, because the promise is enormous and the verification is tedious. The check that works is short: find the clearance, or find the published study, or treat it as a plan rather than a product.

The short version

Nobody has solved non-invasive glucose. A decade of photonics, radio frequency and spectroscopy has produced one bankruptcy, several patent portfolios, a best published figure of 11.1% MARD against a single-digit bar, and an FDA safety communication telling consumers not to buy what is being sold.

What shipped instead does not measure anything. It infers a longitudinal pattern from sensors that already exist, which sidesteps the physics problem and inherits a different ceiling, set by how much the correlated signals actually carry.

Apple took that through a 510(k), published a pivotal study of over 2,000 people, and disclosed 41.2% sensitivity. Google took the same idea through wellness enforcement discretion and has published nothing at the device level. Both are legitimate routes. Only one of them told you the number.