Google Health Guardian on Pixel Watch: Blood Pressure and Insulin Resistance Trends Without FDA Clearance

Short answer: Google Health Guardian estimates blood pressure and insulin resistance trends from a Pixel Watch, without a cuff or a blood test, and Google ships them as wellness features rather than FDA-cleared measurements.

The first monthly reports arrived on 1 October 2026. Some owners with normal cuff readings or normal A1C tests were told their blood pressure or insulin resistance trend was High; one person with type 2 diabetes reported the opposite experience, a result that matched their condition [1]. Both reactions are what you would expect from a screening model with real but limited accuracy, delivered every month to everyone who wears the watch.

This post covers what each feature does, what Google’s published research shows about the models underneath, how that compares with Apple’s FDA-cleared hypertension notification, and what health plans, care providers and app developers should do when users start arriving with these flags.

What is in Health Guardian?

Google announced Health Guardian on 12 August 2026 and turned the features on in the Google Health app on 24 September [2][3].

Blood pressure and insulin resistance reports arrive on the first of each month. The features are free on Pixel Watch 3, 4 and 5, and require Google Health Premium on Fitbit Air [3].

The models are built on Google’s foundation models: WavesFM, trained on billions of minutes of heart and movement data, for blood pressure, and SensorFM, pretrained on more than a trillion minutes from 5 million opted-in users, for insulin resistance [2].

Why is it a wellness feature and not a medical one?

Because the FDA made room for it in January 2026.

The FDA’s revised General Wellness guidance describes non-invasive sensors that estimate physiological parameters, including blood pressure and blood glucose, as potentially low-risk wellness products when they are intended solely for wellness and their values are validated [4]. FDA Commissioner Marty Makary summarized the line plainly at the time: a device that simply provides information can do so without FDA regulation, but claims of being medical grade are a different matter [5].

Google’s labeling follows that line closely. The features are “trends”, not measurements; they are “not intended to diagnose, mitigate, treat, or cure any disease”; and users are told to consult a medical professional [2]. The word “trends” is doing regulatory work, as the analysis of Google’s insulin resistance claim argued when the feature was announced. The detailed boundary is covered in whether your health app is a medical device.

Apple took the other route with the same question. Its hypertension notification went through FDA 510(k) clearance in September 2025, with a published validation study [6]. That is the comparison that matters for anyone deciding how much to trust each signal.

How accurate are the models behind it?

Google has not published performance figures for the shipped features. What it has published is the research underneath them.

Blood pressure. In a prospective multicenter study posted in December 2025, Google Research and Beth Israel Deaconess Medical Center tested a model using photoplethysmography and accelerometer data from a consumer smartwatch, with no cuff calibration, in 196 people without known cardiovascular disease or blood pressure medication. Against 24-hour ambulatory blood pressure monitoring, the clinical reference standard, seven days of real-world wear gave 65.8% sensitivity, 90.0% specificity and a positive predictive value of 80.6% [7]. Google says blood pressure trends were also validated in its Fitbit Hypertension Labs study [2].

Insulin resistance. The WEAR-ME study of 1,165 people, published in Nature, predicted insulin resistance against HOMA-IR, an estimate from fasting glucose and insulin. A model using wearable data and demographics alone reached 60% sensitivity and 80% specificity; adding routine blood tests raised performance considerably [8][9]. HOMA-IR is itself an approximation of the gold-standard clamp test, which matters when “High” is the output a user sees [9].

Apple, for contrast. Apple’s cleared hypertension notification flagged 41.2% of people with hypertension at 92.3% specificity over 30 days against home cuff readings, and does not report a blood pressure number either [6]. The details are in cuffless blood pressure wearables.

Three cautions apply when reading the chart. The references differ: 24-hour ambulatory monitoring, home cuffs and HOMA-IR are not the same yardstick. The populations differ. And Google’s numbers are for research models, not necessarily the exact feature on the watch.

Why the first reports surprised people

Two design choices explain most of the 1 October reaction.

It reports every month, to everyone. Apple’s feature stays silent unless it sees a pattern over 30 days [6]. Health Guardian delivers a trend category to every eligible user on the first of the month [3]. A monthly report that sometimes says High will reach many more people with false positives than a feature that only speaks up when it is confident.

Specificity of 80% to 90% is still a lot of false flags. If a model is wrong for 10% of people without hypertension, a million wearers without hypertension produce about 100,000 High flags. For insulin resistance at 80% specificity, it is about 200,000 per million. Some of the users with normal cuff and A1C results were very likely among them [1][7][9].

The confusion is not proof that the models are bad. It is what screening at population scale looks like when the output goes straight to consumers rather than through a clinician.

What should health plans, providers and apps do?

Pixel Watch users will bring these reports to call centers, care managers and clinicians. Plans and products should decide in advance what happens next.

• Offer the confirmation test. A validated home cuff or a lab order turns a trend into a measurement. That is the step that matters for care, and it is also what the AMA survey says physicians need before they trust wearable data; see how doctors use wearable data.
• Record the confirmed value, not the flag. A trend category has no mmHg and no lab range. It should not become a diagnosis code, a coverage decision or a risk score input on its own.
• Expect volume, and design for it. Monthly reports to every wearer will generate far more High flags than an opt-in screening program. Triage with risk factors such as age, BMI and family history rather than treating every flag the same.
• Do not echo it as a diagnosis. An app that repeats “your blood pressure is high” is making a stronger claim than Google does. Keep the language at “your watch reported a higher trend; here is how to check”.
• Watch the subgroups. The WEAR-ME model performed much better in obese and sedentary participants than overall [9], which is also where confirmation is most worth paying for.

Can apps read Health Guardian data?

Not directly. Health Connect has blood pressure, heart rate, heart rate variability, sleep and oxygen saturation records, but no type for blood pressure trends, insulin resistance trends or sleep breathing quality [10]. The results live in the Google Health app.

What apps can read is the underlying physiology that Pixel Watch writes to Health Connect, and any real cuff readings a user’s blood pressure monitor writes there. That is the more useful input for most products anyway: a confirmed reading with a source beats a monthly category with none. How those readings should be stored alongside estimates is covered in the blood pressure guide.

The short version

Health Guardian brings blood pressure and insulin resistance trends to every Pixel Watch 3, 4 and 5, reported monthly as wellness features without FDA clearance [2][3]. Google’s research models behind them caught about two thirds of hypertension and 60% of insulin resistance, with specificity of 90% and 80% [7][8]. Apple’s cleared feature catches fewer cases but only speaks when it is confident [6]. The 1 October reports show what monthly screening for everyone produces: useful nudges, real cases found, and a lot of false alarms. Treat a High trend as a reason to measure properly, and build the confirmation step into any program that will see these flags.