The pretraining cohort there came from five million people who had consented to their Fitbit or Pixel Watch data being used for health and wellness research, collected over a year in over 100 countries (Google Research, 2026)

Consented. Opted in.

That is not a random sample of the population a health system serves. It is a sample of people who already own a smartwatch and already agreed to hand over their data for research. The YouGov figure still applies here: 47% of UK adults aged 25 to 34 own a smartwatch, against 19% of those 55 and over (YouGov, 2025). SensorFM’s training population is drawn from the same group: people who already own a smartwatch. Five million data points does not change who they were sampled from.

The consented/opt-in is important as it selects twice over. First for who owns the device, second for who agrees to hand the data over for research.

So who does SensorFM represent?

The pretraining cohort is split into four age bands: 18 to 39 (38.7%), 40 to 59 (35.3%), 60 to 79 (21.6%), and 80 plus (2.1%). The evaluation cohort (testing the model) skews younger. 40.0% are 18 to 39, 45.1% are 40 to 59, 13.7% are 60 to 79, and 80+ accounts for ~0.1%. The performance and results of the model are not broken down by demographics. Even the model’s win rate (beating benchmark baselines in 34 out of 35 health tasks) is an average across the full cohort. Whether it holds for the 60-to-79 group is not reported, and with only 19 people aged 80 plus, no result for that band could be reported with any confidence.

This is clearly discussed as a limitation in the paper. The study population reflects the characteristics of the self-selecting device users. I’d take a punt that this is not truly representative of the US population, never mind the rest of the world.

The SensorLM model from the same family shared this DNA has a pre-training cohort median age of 41.7 (Zhang et al. 2025). he skew toward working-age adults of higher socioeconomic status isn't a one-off glitch in a single dataset. It’s a systemic feature of who buys these devices and who opts into research.

We are leaving patients behind

This is not a new observation, nor a guess. Torous, Rodriguez and Powell described a three-tier digital divide back in 2017: who owns the device, who actively downloads and engages with health apps or studies, and who goes on to generate the sustained, high-quality longitudinal data the field is actually built on (Torous et al., 2017). Each tier filters out more people, and the filtering is not random.

A recent study of cancer survivors showed that wearable activity tracker ownership was significantly associated with holding a college degree and higher household income, independent of age (Wang et al., 2026). The people most likely to own the device are already a specific socioeconomic group.

Unfortunately, we as clinicians are also biased. A 2026 study across French and Belgian oncology centres found that healthcare providers' perception of a patient's digital capacity was shaped by age, socioeconomic status and perceived frailty, independent of the patient's actual ability to use the technology (Prospero et al., 2026).

What an unfortunate paradox.

These are the exact people we should be targeting. The ageing population are more likely to be at risk of falls, frailty and cardiac arrhythmias, and therefore benefit from the information gained from these wearables (Savelieva et al., 2023). The people least likely to be offered a digital tool are often the ones most likely to benefit from it.

References

Google Research. SensorFM: Towards a general intelligence and interface for wearable health data. Google Research Blog, 2026.

Narayanswamy G, Xu MA, Heydari AA, et al. Towards a General Intelligence and Interface for Wearable Health Data. arXiv. 2026;2605.22759.

Patel MS, Emedom-Nnamdi P, Lapen K, Dee EC. Wearable-Derived Data for Patient Frailty: Extending Hospital Frailty Risk Score While Confronting Bias and Inequities. J Gen Intern Med. 2026. doi:10.1007/s11606-026-10576-3

Prospero E, Yazigi A, Rouzier R, Franzoi MA, L’Orphelin JM. Digital illiteracy and access to remote patient monitoring in oncology: a multicentre study of healthcare providers’ perceptions and patient-reported digital literacy. EClinicalMedicine. 2026;94:103866. doi:10.1016/j.eclinm.2026.103866

Savelieva I, Fumagalli S, Kenny RA, et al. EHRA expert consensus document on the management of arrhythmias in frailty syndrome. Europace. 2023;25(4):1249-1276. doi:10.1093/europace/euac123

Torous J, Rodriguez J, Powell A. The New Digital Divide For Digital BioMarkers. Digit Biomark. 2017;1(1):87-91. doi:10.1159/000477382

Wang SD, Song J, Pincever J, et al. Mobile Health Technology Ownership and Use Among Cancer Survivors in a Health System. Cancer Rep (Hoboken). 2026;9(4):e70536. doi:10.1002/cnr2.70536

YouGov. Nearly half of Britons concerned about wearable tech privacy. YouGov Surveys: Serviced, fieldwork 6-7 March 2025, nationally representative sample of 2,090 UK adults aged 18+. Published 17 March 2025.

Zhang Y, Ayush K, Qiao S, et al. SensorLM: Learning the Language of Wearable Sensors. arXiv. 2025;2506.09108.

It found that 2 devices (FitBit Charge 6 and Google Pixel Watch 2) had the lowest error (around 4.5 bpm). Other devices ran 2-3 times higher - 9 to 14 BPM (Gielen et al., 2026).

The study cohort consisted of 45 adults aged 21 to 68. Mean age was 34 (SD 12, i.e. most (68%) people were within 12 years of this mean).

Who buys a device?

47% of UK adults aged 25 to 34 own a smartwatch compared to 19% of those aged 55 and over (YouGov, 2025). 66% of over 55’s own no smart wearable at all. Two-thirds of the population most likely to be enrolled in a remote monitoring pathway do not own one of these devices and, on the evidence of the same survey, are the group least willing to pay for one.

Who are these devices tested on?

In addition to the study mentioned above, a further study tested devices across three physical intensities in 40 young adults (Jeony et al 2026). The devices in this study were significantly different in error rates with the Polar Verity Sense and Apple Watch SE outperforming the Polar Pacer Pro and Galaxy Watch 6. The authors’ conclusion?

Another finding stood out to me. The error rate was highest at low intensity and decreased as intensity increased.

That pattern is not a one-off. A separate 2026 validation of the Galaxy Watch 6 across a maximal ramp test found median absolute percentage error dropping from 2.90% in the 50 to 60% zone of maximum heart rate to between 0.60 and 0.75% above 70% (Inoue et al., 2026).

Considering PPG’s limitations with movement, I find this really interesting with considerations for the use of these devices in health over performance. Two independent studies agree that these devices are at their worst when the wearer is barely moving. This is the state a seventy-eight year old on a virtual ward is in for almost the entire monitoring period. The sport use case sits in the intensity band where the hardware performs best. The clinical use case sits in the band where it performs worst, and nobody has validated it there.

What does a 10bpm error actually mean?

On its own, nothing.

It needs context.

Consumer wearable HR validation research has a threshold. The convention is a mean absolute percentage error (MAPE) at or below 10% (Chow and Yang, 2020; Choe and Kang, 2025). It is used explicitly as the validity cut-off in device comparison work. 4/10 of the devices in the Gielen study had a MAPE of 11 - 16%.

Clinical context, not numbers.

Bradycardia is conventionally below 60 beats per minute and tachycardia above 100. A device carrying average absolute error of 9 to 14 bpm cannot reliably tell you which side of either line a patient sits on. For a resting heart rate somewhere in the seventies, that error band spans the range in which you would either do nothing or perhaps be concerned.

A meta-analysis of 56 Apple Watch studies, pooling 148 heart rate effect sizes, found mean bias close to zero but limits of agreement running from roughly ± 11 bpm (Choe and Kang, 2025). The average is excellent. The spread around it is 22 bpm wide. Averages are what manufacturers report. Individual patients live in the spread.

And on the rare occasion somebody splits a device study by age, the gap shows up. Testing the Garmin Vivosmart HR+ and Xiaomi Mi Band 2 in 20 young and 20 older adults, concordance for the Garmin fell from 0.92 in the young group to 0.80 in the older group, and the authors flagged sporadic extreme outlier readings in both (Chow and Yang, 2020). Both devices passed the 10%threshold overall. The age gradient was still there, sitting underneath a headline that said the device was fine.

What’s actually going on?

NHS England’s procurement guidance for technology-enabled virtual wards sets out continuous monitoring via wearables as a service design option, alongside spot monitoring, and instructs commissioners to involve a CSO in assembling the clinical safety assessment.

It is unlikely the NHS will be using consumer wearables at scale, but what if a patient already owns their own?

A UK study interviewed UK stakeholders involved in delivering, designing or evaluating frailty virtual wards. Smartwatches had been piloted (unspecified devices). One frontline healthcare professional reported that they “don’t seem to work on very frail skin” and that some were physically too heavy for patients. The study’s broader finding was that procurement practices were misaligned with service needs, producing equipment that was not fit for purpose (Lindsay et al., 2025).

Most clinicians interviewed thought remote monitoring was inappropriate for most frail patients, on clinical, practical and social grounds. Connectivity in patients' homes was poor. Setup time exceeded the length of many admissions. Staff workload rose in ways that were not visible in the business case.

Where are the practicality and usability sections? Are HFE experts not being consulted?

The point is that this is commercially rational. A manufacturer validates against the population that buys the product. Nobody involved is behaving badly. The result is still a body of evidence that describes performance in thirty-somethings and is then cited in business cases about seventy-somethings.

What happens when someone does test the older cohort?

The BASEL Wearable study compared 5 direct-to-consumer devices tested for atrial fibrillation against a 12-lead ECG in patients presenting to a tertiary cardiology service. 31% of the participants had AF. The median age was 66.7 (Mannhart et al., 2023). Sensitivity ranged from 58% to 85%. Specificity ranged from 69% to 79%. However, between 17% and 26% of rhythms from the tracings were reported as inconclusive. In an older, likely comorbid, real-world cohort, roughly one reading in four was not a wrong answer but no answer.

In comparison, the Doherty umbrella review reported a near-100% sensitivity and 95% specificity for AF. Doherty and colleagues qualify that number themselves. They note that participants in the arrhythmia studies were generally drawn from clinical populations already carrying a diagnosis of atrial fibrillation. It is not a screening figure.

Nor is it the only figure in their review. A second included review of smartwatch and smart band studies reported sensitivity ranging from 67.7 to 100% and specificity from 67.6 to 98%, depending on algorithm, population and testing conditions. That range brackets the BASEL results rather than contradicting them.

BASEL is not a screening study. It is a referral cohort with 31% prevalence, which is close to the most favourable condition these algorithms will ever meet. They still landed where they landed. Part of the reason is methodological: BASEL counted the inconclusive tracings, and the authors attribute the reduced sensitivity and specificity directly to them. Exclude the readings the algorithm declined to call and the figures improve considerably. Patients do not get to exclude them.

Why the gap will not close on its own

Set against the broader picture from Doherty and colleagues, the age problem is a refinement applied to a base that is barely there. Of roughly 310 consumer devices on the market, around 11%have been validated for even one biometric outcome, and the metric-level validations actually completed represent about 3.5% of what a full picture would require.

There is no commercial incentive to fix this. The manufacturer’s customer is 34. The health system’s patient is 74. Validating in the second population costs money, risks producing worse numbers, and serves a buyer the manufacturer does not yet have. Left alone, the incentive runs the wrong way.

Which means the NHS is not adopting a validated technology into a new population. It is adopting a consumer product away from the only population it was ever tested on.

What I would want to see

Four things, none of them difficult.

• Age-banded reporting as standard. Not mean cohort age. Performance by decade, with the sample size behind each band visible.

• Recruitment matched to the deployment population, not the customer base. If the procurement case is for over 65s, the evidence should be too.

• Comorbidity reported alongside age. Arrhythmia, heart failure, diabetes and peripheral vascular disease all sit on the causal path between the sensor and the number. Age is partly a proxy for them.

• Inconclusive rate reported as a primary outcome. BASEL is the model. A device that declines to produce a reading a quarter of the time has a problem that sensitivity and specificity will never show you.

Age is the easiest demographic variable in the world to collect. Every one of these studies already has it. Nobody is required to report the split, so nobody does.

Until that changes, a wearable score in an older patient is a number carrying an error bar that nobody has measured. That is a reasonable thing to be curious about. It is not a reasonable thing to commission a service around.

References

Choe JP, Kang M. Apple watch accuracy in monitoring health metrics: a systematic review and meta-analysis. Physiol Meas. 2025;46(4). doi:10.1088/1361-6579/adca82

Chow HW, Yang CC. Accuracy of Optical Heart Rate Sensing Technology in Wearable Fitness Trackers for Young and Older Adults: Validation and Comparison Study. JMIR Mhealth Uhealth. 2020;8(4):e14707. doi:10.2196/14707

Doherty C, Baldwin M, Keogh A, Caulfield B, Argent R. Keeping Pace with Wearables: A Living Umbrella Review of Systematic Reviews Evaluating the Accuracy of Consumer Wearable Technologies in Health Measurement. Sports Med. 2024. PMID: 39080098.

Gielen J, Van Oost CN, Debard G, et al. Accuracy of Optical Heart Rate Measurements for 10 Commercial Wearables in Different Climate Conditions and Activities: Instrument Validation Study. JMIR Form Res. 2026;10:e85186. doi:10.2196/85186

Inoue A, Soares JPF, Antunes-Santos F, et al. Heart Rate Estimation Using the Galaxy Watch During Maximal Cardiopulmonary Exercise Testing: Cross-Sectional Validation Study. JMIR Cardio. 2026;10:e81917. doi:10.2196/81917

Jeong Y, Yoon S, Lee J, Jeong WM, Ha MS. Comparison of heart rate measurement accuracy among commercially available photoplethysmography-based wearable devices across exercise intensities. Phys Act Nutr. 2026;30(2):142-150. doi:10.20463/pan.2026.0032

Kitagaki K, Hongo Y, Futai R, et al. Validity of Heart Rate Measurement Using Wearable Devices During Cardiopulmonary Exercise Testing in Patients With Cardiovascular Disease: Prospective Pilot Validation Study. JMIR Cardio. 2025;9:e77911. doi:10.2196/77911

Lindsay RK, Cunnington P, Dixon-Woods M. Why is implementing remote monitoring in virtual wards (Hospital at Home) for people living with frailty so hard? Qualitative interview study. Age Ageing. 2025;54(1):afaf003. doi:10.1093/ageing/afaf003

Mannhart D, Lischer M, Knecht S, et al. Clinical Validation of 5 Direct-to-Consumer Wearable Smart Devices to Detect Atrial Fibrillation: BASEL Wearable Study. JACC Clin Electrophysiol. 2023;9(2):232-242. doi:10.1016/j.jacep.2022.09.011

Merchant RA, Loke B, Chan YH. Ability of Heart Rate Recovery and Gait Kinetics in a Single Wearable to Predict Frailty: Quasiexperimental Pilot Study. JMIR Form Res. 2024;8:e58110. doi:10.2196/58110

Mou C, Yan X, Miao X, Zhu L. Accuracy of wearable devices in predicting falls in older adults: a systematic review and meta-analysis. Front Public Health. 2026;14:1778750. doi:10.3389/fpubh.2026.1778750

NHS England. Virtual wards enabled by technology: guidance on selecting and procuring a technology platform. england.nhs.uk.

YouGov. Nearly half of Britons concerned about wearable tech privacy. YouGov Surveys: Serviced, fieldwork 6-7 March 2025, nationally representative sample of 2,090 UK adults aged 18+, weighted by age, gender, education and region. Published 17 March 2025.

Melanin absorbs light. The more melanin in the skin, the more of that light is absorbed before it reaches the blood, and the weaker the signal the sensor has to work with. This isn’t a manufacturing flaw or a bad batch. It’s optics. Skin tone bias in PPG isn’t a possibility to rule out. It’s a default to design against, and then prove you’ve beaten it.

With sex differences, the physiology is real, but the measurement bias is subtle. With skin tone, the measurement bias is documented, large, and clinically dangerous, and SensorFM is built squarely on the signal where it can show up worst.

Two lights, two problems

Consumer devices don’t use one wavelength for everything. Green light (around 530nm) handles heart rate and HRV. It’s used because it gives a strong AC signal during movement, which is what continuous tracking needs. But epidermal melanin absorbs green light heavily, which degrades signal-to-noise on darker skin, especially under exercise.

Red (around 660nm) and near-infrared (around 940nm) handle SpO₂. Oxygenated and deoxygenated haemoglobin absorb these wavelengths at different rates, which is what lets a device calculate oxygen saturation via a ratio-of-ratios calculation. Melanin alters baseline attenuation across both wavelengths unevenly, which skews the calibration curve the SpO₂ estimate depends on.

That’s why SpO₂, not heart rate, is the single most pigment-sensitive metric in optical sensing.

What does the evidence say?

Pulse oximeters, the fingertip kind, in every hospital, estimate blood oxygen using the same optical principle as your watch. And they are less accurate on darker skin, in a way that causes harm.

The clinical wake-up call came in 2020. A landmark NEJM study found hospital pulse oximeters missed hypoxemia roughly three times as often in Black patients as in White patients at the same reading (Sjoding et al. 2020). That paper is the reason pulse oximetry skin tone bias moved from academic finding to FDA-level concern over the following five years, and it's the reason wearables and foundation models built on the same optical signal need to compensate for this bias.

The larger cohort studies since confirm the pattern. Occult hypoxemia, a reassuring oximeter reading while actual blood oxygen is dangerously low, occurred at the same reading in 2.1% of Black and 1.8% of Hispanic patients, versus 1.1% of White patients, in a study of over 46,000 patients (Burnett et al. 2022). A separate cohort of over 36,000 ICU patients found Black patients over-represented in oximeter discrepancies and occult hypoxemia, and, critically, that those discrepancies tracked worse outcomes: longer ICU stays and higher mortality (Saidy et al. 2025). Same reading, different risk.

Now the constructive half, because it matters. A device engineered and calibrated on a balanced range of skin tones can close the gap: Masimo’s own laboratory data, on healthy Black and White volunteers taken down through hypoxia, show no clinically significant racial difference in accuracy (Barker & Wilson 2022). Read that as proof it’s solvable, not proof it’s solved everywhere, since it’s industry-authored. But the direction is the point. Inclusive calibration works. The bias is a design choice, not a destiny.

The consumer gap

Hospital oximeters are regulated medical devices with clearance requirements. Consumer wrist wearables are not held to the same bar, and it shows. Independent testing finds wrist heart rate accuracy that is generally fine at rest but degrades during exercise, and degrades more on darker skin, with large differences between brands. Some devices stay within a few beats per minute across skin tones; others drift by 10 to 15bpm at rest and worse under exertion (systematic review; Fitzpatrick-stratified device studies).

To be fair to the field, some validations do the right thing and pass. A machine learning wrist-PPG algorithm tested with an explicit skin tone subgroup, roughly a quarter of the sample at Fitzpatrick V to VI, met its accuracy threshold across skin tones (Chen et al. 2024). That’s exactly the breakdown I want to see more of. The problem isn’t that no one can do this. It’s that it isn’t standard, so “validated” on a box tells you nothing about which validation you’re getting.

Manufacturers know about this and have responded, at least at the hardware level. Garmin and Fitbit have reportedly configured devices to boost green LED intensity when the sensor detects a weak signal, pushing more light through higher-melanin skin to recover a usable reading (govtech.com, 2021). Google’s own account of the Pixel Watch 2 and Fitbit Charge 6 redesign describes a multi-LED, multi-photodiode array built to cover more skin surface, with skin tone named explicitly as one of the variables the team tested against during development (Google, 2023). That’s a real engineering response, not a marketing claim, and it’s worth crediting.

It doesn’t fully close the gap. A Monte Carlo modelling study of the Apple Watch Series 5, Fitbit Versa 2, and Polar M600 found that increasing skin tone and BMI together can still produce a theoretical signal loss of up to 61% in the Fitbit Versa 2 and around 32% in the other two devices, even accounting for the sensor geometry manufacturers use (Ajmal et al. 2021). A separate study on a commercial wearable found skin tone was the only significant predictor of heart rate accuracy at rest and while walking, tattoos included (Koerber et al. 2022 systematic review; tattoo-specific study, 2024). Compensation helps. It doesn’t make skin tone stop mattering, and none of these manufacturer mechanisms have been published with the kind of subgroup validation data that would let an outside reviewer check the claim.

There’s a plausible mechanical reason compensation degrades under motion. “Cadence lock”, where the sensor’s algorithm mistakes the rhythmic motion of a stride for the pulse signal, is a well-documented PPG failure mode, acknowledged even in manufacturer forums. Whether melanin-driven LED intensity boosting specifically worsens cadence lock isn’t something I’ve found stated in any published source. It’s a reasonable hypothesis given the underlying physics, not an established finding, and I’m flagging it as such rather than asserting it.

The regulatory gap this all sits inside is now moving, at least on the clinical side. The FDA published draft guidance in January 2025 requiring pulse oximeter manufacturers to test accuracy across skin pigmentation groups, using the Monk Skin Tone scale and a spectrophotometric measure (Individual Typology Angle) rather than self-reported race, with minimum sample sizes and explicit error thresholds. That guidance applies to medical pulse oximeters. Consumer wearables, classified as wellness devices, sit outside it entirely. The regulatory bar clinicians are about to hold hospital oximeters to has no equivalent for the device on your wrist.

Back to SensorFM

SensorFM’s inputs include SpO2, derived from PPG, the single measurement with the strongest documented skin tone bias in sensing. Its pretraining cohort, the trillion minutes, five million people, never recorded ethnicity: unrecorded, not merely un-analysed. Its downstream validation cohort, the phenotyped 13,985 used to prove the tasks, is about 79% White. The authors name the limitation themselves: the data “is skewed towards... White/Caucasians,” and performance “may not generalize directly to the overall population.”

Worth being precise about the pipeline: the entire pretraining set comes from Fitbit and Pixel Watch devices, over 20 models, exclusively, and the model doesn’t ingest raw PPG. It takes 34 pre-processed one-minute aggregate features, computed on-device. So the SpO2 reaching SensorFM has already passed through Fitbit and Pixel Watch’s own compensation, the dynamic LED boosting and multi-wavelength array covered above. It isn’t raw, uncorrected signal. But that compensation is documented as incomplete, so the aggregate feature still carries some residual, unquantified bias from the hardware layer beneath it. What’s undisclosed is whether any signal quality flag travels with those 34 features, so the model has no visible way to distinguish a degraded reading from a clean one. That’s the specific, answerable question left, and it belongs to the authors.

Worth pre-empting the standard defence: that self-supervised pretraining on massive data lets a model learn invariant representations regardless of demographic labels. That doesn’t hold for a signal-to-noise problem. Self-supervision cannot extract signal from noise that was never captured cleanly, so the pretrained representation for darker-skinned cohorts is built on a systematically noisier distribution, a bias that propagates rather than disappears at scale, and compounds at fine-tuning on a 79% White cohort.

None of this is inference. It’s the paper’s own disclosure, and it deserves the same weight as the headline result. A model built on the most pigment-sensitive signal available, trained on data that didn’t record pigment and validated on a mostly-White cohort, is making a claim about “human physiology” its own evidence base can’t yet check against most of the humans.

What I’d want to see next

• Report SpO₂ and heart rate task performance by Fitzpatrick skin type, the way the best consumer validations already do. Worth noting the Fitzpatrick scale was designed in 1975 for UV sun-reactivity, not optical absorbance at green or red/NIR wavelengths. It’s the current industry standard for subgroup reporting, but the Monk Skin Tone scale and direct spectrophotometric measures like Individual Typology Angle are becoming the better fit for optical validation specifically.

• Record skin tone in future data collection. You cannot audit what you never measured. The pretraining “n.c.” is the thing to change first.

• Validate beyond a roughly 79% White cohort before the phrase “general-purpose representation of human physiology” does load-bearing work.

• Disclose whether the model ingests raw or pre-processed PPG, and whether any signal quality gating exists.

• Apply something like the FDA’s Monk Skin Tone / ITA testing standard, currently proposed for medical pulse oximeters, to the consumer-grade SpO2 feeding into this model. Nothing requires SensorFM to meet a clinical regulatory bar. But borrowing its testing standard voluntarily would be the single fastest way to answer the question this whole post is asking.

None of these needs new science. The Masimo data show the underlying problem yields to inclusive calibration. The Chen et al. validation shows the reporting is routine when someone decides to do it. This is the post in the series where the ask is clearest and the precedent already exists.

The green light is a genuinely clever piece of engineering. But it can only report what the light can see, and on darker skin, it sees less. A model that never checks whether that matters isn’t wrong, exactly. It just hasn’t earned the word “general” yet.

References

1. Sjoding MW, Dickson RP, Iwashyna TJ, Gay SE, Valley TS. Racial Bias in Pulse Oximetry Measurement. N Engl J Med. 2020;383(25):2477-2478. https://doi.org/10.1056/NEJMc2029240

2. Burnett GW, Stannard B, Swift MD, et al. Self-reported Race/Ethnicity and Intraoperative Occult Hypoxemia. Anesthesiology. 2022;136(5):688-696. https://doi.org/10.1097/ALN.0000000000004153

3. Saidy S, Adams C, Patel R, et al. Pulse Oximetry Discrepancies and Occult Hypoxemia in ICU Patients. J Intensive Care Med. 2025;40(12):1269-1278. https://doi.org/10.1177/08850666241280000

4. Barker SJ, Wilson WC. Racial effects on Masimo pulse oximetry: a laboratory study. J Clin Monit Comput. 2023;37(2):567-574. https://doi.org/10.1007/s10877-022-00927-w

5. Chen W, Liu Y, Zhang H, et al. Multicenter Evaluation of Machine-Learning Continuous Pulse Rate Algorithm on Wrist-Worn Device. Digit Biomark. 2024;8(1):218-228. https://doi.org/10.1159/000538910

6. Ajmal S, Boonya-Ananta T, Rodriguez AJ, Du Le VN, Ramella-Roman JC. Monte Carlo analysis of optical heart rate sensors in commercial wearables: the effect of skin tone and obesity on the PPG signal. Biomed Opt Express. 2021;12(12):7445-7457. https://doi.org/10.1364/BOE.439893

7. Koerber D, Khan S, Shamsheri T, Kirubarajan A, Mehta S. Accuracy of Heart Rate Measurement with Wrist-Worn Wearable Devices in Various Skin Tones: a Systematic Review. J Racial Ethn Health Disparities. 2023;10(5):2433-2445. https://doi.org/10.1007/s40615-022-01446-9

8. Colvonen PJ, DeYoung PN, Bosompra NO, Malhotra A. Limiting racial disparities and bias for wearable devices in health science research. Sleep. 2020;43(11):zsaa159. https://doi.org/10.1093/sleep/zsaa159

9. Food and Drug Administration. Pulse Oximeters for Medical Purposes: Non-Clinical and Clinical Performance Testing, Labeling, and Premarket Submission Recommendations (Draft Guidance). U.S. Department of Health and Human Services; 2025.

10. Lipnick MS, Ehie O, Igaga EN, Bicker P. Pulse Oximetry and Skin Pigmentation: New Guidance From the FDA. JAMA. 2025;333(16):1393-1395. https://doi.org/10.1001/jama.2025.2140

11. Narayanswamy G, Xu MA, Liu X, McDuff D, et al. Towards a General Intelligence and Interface for Wearable Health Data. arXiv preprint. 2026;arXiv:2605.22759v1. https://doi.org/10.48550/arXiv.2605.22759

Here’s the situation. The model was not used to predict sex. None of the 35 results is broken down by sex, skin tone or device. As stated last week, the authors comment that the model is “skewed towards women,” that performance “may not generalize directly to the overall population,” and that future work “will require… subgroup-specific validation.”

What would the missing results breakdown actually have told us? Let’s break it down signal by signal.

The model learns from signals we already know differ by sex. Until someone publishes evidence that the representation holds up equally for women and men, "general-purpose representation of human physiology" isn't a finding. It's a claim awaiting a test. For a model whose entire premise is generality, that evidence isn't a nice-to-have. It is the claim. Until wearable ecosystems release systematic subgroup evaluations, showing that frozen representations maintain parity in fairness and utility across biological sex, the claim of universal physiological generality remains unproven in clinical terms.

Biological sex is a signal-level problem.

Raw physiology varies by sex. This can show up in inherent baseline variance or reactivity. It shows up in the five signals SensorFM uses (photoplethysmography, accelerometry, electrodermal activity, skin temperature and altimetry) and across other metrics.

When a baseline differs by sex, the same number carries different information depending on whose wrist it came from. Here’s where that shows up.

Heart-rate variability. Women show a lower mean RR interval, lower SDNN and lower total power, alongside greater high-frequency power — the band indexing vagal, parasympathetic activity — and lower LF/HF ratios, all despite higher resting heart rates (Koenig & Thayer). The difference isn’t only at rest: men and women shift their sympatho-vagal balance differently in response to standing (Hnatkova et al.). The pattern also reproduces outside Western cohorts — normative values in healthy young Nigerian adults show the same direction (Adamu et al.), which matters for a model claiming 100+ countries of coverage. Read against a male-skewed prior, a healthy woman’s profile can look like elevated sympathetic tone that isn’t there.

Sleep architecture. Women retain significantly more slow-wave (N3) sleep and less wake-after-sleep-onset into older adulthood than age-matched men (Taporoski et al.). The difference reaches into microstructure that conventional scoring misses. The sexes differ in cyclic alternating patterns during non-REM despite similar sleep stages (Hartmann et al.). A model that scores “deep sleep” from wrist signals is implicitly setting a threshold. Whose architecture is that threshold tuned to?

Electrodermal activity. Here the literature is less settled. EDA is the sweat-gland signal SensorFM lists among its five inputs, used as a proxy for sympathetic arousal and, increasingly, “stress.” Sex differences in EDA reactivity appear in some studies and not others: one study of emotional response coupling found women showed stronger concordance across autonomic channels including EDA (Rattel et al.), while other work found no sex difference elsewhere. The honest summary is that EDA baselines and reactivity are shaped by multiple factors, including sex, that a pooled model may not be handling explicitly, and nobody has shown that it is.

And then there’s the cycle. All of the above sits alongside infradian rhythms in skin temperature, heart rate and HRV across the menstrual cycle. Skin temperature and PPG-derived heart rate are both core SensorFM inputs. For roughly half the population, for several decades of life, some of this model’s key inputs oscillate monthly in a pattern that is essential to record.

Sensors - anatomy changes the measurement

The physiology is only half of it. The other half is that bodies differ in ways that change what the sensor can physically detect.

PPG. Fat distribution, wrist and forearm geometry, and local vascular perfusion all affect how well the green light reaches blood and returns. On top of that, differing forearm muscle activation during exertion generates distinct motion artefacts. There’s a neat illustration of what that does in practice: a machine-learning wrist pulse-rate algorithm validated against ECG met its accuracy threshold in every subgroup except men under motion, where error rose past the clinical bar, driven by stronger forearm contractions (Chen et al., Empatica). The authors argue, reasonably, that these isolated outliers wouldn’t undermine the device’s intended use. Set the clinical adjudication aside and keep the structural fact: a device can pass its headline accuracy target and still fall out of spec for one sex under one condition. You only know because they published the model breakdown.

That example cuts against the lazy version of this argument, and I want to keep it in view. Subgroup reporting isn’t absent from the field. It just isn’t standard, and it’s inconsistent when it happens. That’s the honest characterisation of the devices SensorFM’s data comes from.

A note on scope before I go further. SensorFM uses five modalities, and ECG and EEG aren’t among them. I’m including them anyway, because this series uses SensorFM as a lens on wearable sensing generally — and the same principle runs through every modality in the field, including the ones the next foundation model will almost certainly absorb:

• ECG (not a SensorFM input) — breast tissue and upper-torso geometry alter electrode placement and the recorded signal.

• EDA (one of SensorFM’s five) — differences in epidermal thickness, sweat-gland density and baseline skin hydration alter contact impedance.

• EEG (not a SensorFM input) — hair density and scalp thickness alter contact impedance, while skull thickness variation modifies volume conduction and cortical signal attenuation.

These are the ordinary anatomy of over half the population — and the pattern is consistent enough that you should expect some version of it in any sensor that touches a body.

Well… let’s just pre-train on everything!

That does not solve the problem. SensorFM learns by self-supervised masked reconstruction, predicting missing chunks of time-series across a vast population. This produces rich, general-purpose embeddings without anyone hand-labelling anything, and that’s genuinely powerful. But learning from uncurated data at scale doesn’t automatically mean learning every subgroup equally well. It can just as easily mean learning the majority pattern very well.

Three things happen when you pool raw multi-sensor data across 20-plus device models without any sex-stratified reporting:

1. Sensor noise and physiology become indistinguishable. The network sees altered optical signal-to-noise from tissue depth or movement, and genuine physiological variance like higher vagal tone, as one undifferentiated feature space. It has no way to tell you which is which — and neither do you.

2. Measurement artefacts get laundered into physiology. If an optical or electrodermal sensor loses fidelity under particular anatomical or movement conditions, a pooled model can read degraded signal quality as a physiological trait. The bias stops looking like a bug and starts looking like a finding.

3. Representation gets confused with performance. SensorFM’s latent space does organise itself by gender — the authors show it. That’s a real and encouraging result. But recovering sex-linked structure in an embedding is not the same as demonstrating equal predictive accuracy across sexes on downstream clinical tasks. The first is visible in a scatter plot. The second requires a stratified table nobody has published.

What “not shown” means - and what it doesn’t

I want to keep this distinction sharp, because it’s the difference between a fair critique and a cheap one.

“Not shown” does not mean “broken.” There are many positive attributes in this model: the embedding encodes gender on its own, and the authors report the model’s reliance on demographic features falls as it scales, consistent with the physiology-only representation doing real work. The commercial devices may perform evenly too. I can’t assert otherwise, and I won’t.

“Not shown” does mean the claim is currently untestable in the direction that matters. A general-purpose physiological model is, by definition, a claim about every subgroup it pools. The evidence that would test that claim — performance stratified by sex, by skin tone, by device generation — is exactly the evidence absent from the paper, on a model that predicts age, BMI, height and weight yet never reports how any result splits by sex. When the missing number is the one that would test the headline, its absence is the story.

The mitigations that already exist

It’s worth being fair here. Consumer wearables haven’t ignored this entirely.

Most devices take sex as an explicit input at registration and feed it into downstream formulas. VO2 max and calorie-burn estimates use sex-stratified constants precisely to avoid the pooling problem described above.

Platforms like Whoop and Oura have also moved away from population norms. Recovery and readiness scores are built from a 14 to 30 day personal baseline, so a reading gets compared against your own history rather than a generic template.

Cycle tracking goes further still. Oura, Apple Watch and Whoop explicitly model infradian rhythm, isolating the monthly temperature and HRV shifts described earlier so they read as expected variation, not anomaly.

We’ve done this before

Here’s what makes me impatient rather than merely interested.

None of this needs new science. It needs no new cohort, no new trial, no new sensor. The data to produce a sex-stratified breakdown already exists inside the analysis that produced the headline. Someone has to decide to publish it.

And we know what the alternative costs, because we’ve run this experiment already. Women were routinely excluded from clinical trials for decades — the NIH only mandated their inclusion in 1993 — and medicine is still unpicking the consequences of drugs tested largely on men and dosed for men. That wasn’t malice. It was a default that nobody interrogated until the harm was obvious in hindsight.

The difference this time is that we can see it happening while the technology is still being built. Foundation models for physiology are at the stage where the conventions are being set — what gets reported, what counts as validated, what a reviewer thinks to ask for. Those conventions are far easier to shape now than to retrofit onto a hundred million wrists.

Which brings it back to SensorFM specifically. I’ve ranged across modalities it doesn’t even use — ECG, EEG — because the point was never really about one model. It’s about what we’re willing to accept as evidence. But SensorFM is the sharpest example available right now, precisely because it’s the most ambitious: five signals, a trillion minutes, five million people, and an explicit claim to represent human physiology in general. The bigger the claim, the more the subgroup evidence matters. And on sex — the most basic biological split there is — that evidence isn’t in the paper.

So: not a takedown, and not a demand for perfection. Just the observation that “we averaged across five million people” answers a different question from “it works for you.” One is a description of the data. The other is a claim about a person.

Next in this series: the same question asked of skin tone — where the measurement bias isn’t subtle at all, and the evidence is a great deal harder to look away from.

Subscribe now

REFERENCES

Physiology and validation citations retrieved via PubMed; DOIs linked.

1. Koenig J, Thayer JF. Sex differences in healthy human heart rate variability: a meta-analysis. Neurosci Biobehav Rev. 2016;64:288–310. https://doi.org/10.1016/j.neubiorev.2016.03.007

2. Hnatkova K, Šišáková M, Smetana P, et al. Sex differences in heart rate responses to postural provocations. Int J Cardiol. 2019;297:126–134. https://doi.org/10.1016/j.ijcard.2019.09.044

3. Adamu DB, et al. Heart rate variability in healthy young adult Nigerians. West Afr J Med. 2024;41(6):675–681. (PMID: 39340793)

4. Taporoski TP, Beijamini F, Alexandria S, et al. Gender differences in the relationship between sleep and age in a Brazilian cohort: the Baependi Heart Study. J Sleep Res. 2024;34(4):e14154. https://doi.org/10.1111/jsr.14154

5. Hartmann S, Bruni O, Ferri R, Redline S, Baumert M. Characterization of cyclic alternating pattern during sleep in older men and women using large population studies. Sleep. 2020;43(7):zsaa016. https://doi.org/10.1093/sleep/zsaa016

6. Rattel JA, Mauss IB, Liedlgruber M, Wilhelm FH. Sex differences in emotional concordance. Biol Psychol. 2020;151:107845. https://doi.org/10.1016/j.biopsycho.2020.107845

7. Chen W, Cordero R, Lever Taylor J, et al. Multicenter evaluation of machine-learning continuous pulse rate algorithm on wrist-worn device. Digit Biomark. 2024;8(1):218–228. https://doi.org/10.1159/000542615

8. Narayanswamy G, Xu MA, et al. Towards a general intelligence and interface for wearable health data. arXiv 2605.22759v1 (2026). https://arxiv.org/abs/2605.22759

9. Age and sex differences in heart rate variability and vagal specific patterns — Baependi Heart Study. Glob Heart. 2020;15(1):71. https://doi.org/10.5334/gh.873

10. Varner KJ, Keeler Bruce L, Soltani S, et al. Sex differences in the variability of physical activity measurements across multiple timescales recorded by a wearable device: observational retrospective cohort study. J Med Internet Res. 2025;27:e66231. https://doi.org/10.2196/66231

11. Bruce LK, Kasl P, Soltani S, et al. Variability of temperature measurements recorded by a wearable device by biological sex. Biol Sex Differ. 2023;14:76. https://doi.org/10.1186/s13293-023-00558-z

SensorFM has a lot of us very excited. Developed by Google Research, it is trained on roughly a trillion minutes (or 19025.875 centuries if anyone is counting, give or take a leap year or two), from around five million people across 100+ countries. That is a huge amount of data. They used 20+ different Fitbit and Pixel Watch models. Their front page heading is that SensorFm is a “general-purpose representation of human physiology.”

Algorithms are only as good as the data they are trained on. If the data has bias, the algorithm will too. Evaluating SensorFM gives me a chance to talk about the parts of sensor and algorithm design that decide what these devices are actually measuring, as opposed to what the marketing says they measure.

Critical appraisal has a bias of its own: it dwells on the flaws. Read any paper, and most of the discussion is about what the authors could have done better. New technology gets the same treatment — and it’s easy to let useful critique slip into plain negativity. So let me start with what is genuinely impressive, because there’s a lot.

First, the impressive part.

Normally, every health question a wearable answers needs its own custom-built model, with someone deciding in advance which bits of the sensor data matter. It’s slow, specialist work. SensorFM skips it. It learns a single general picture of your physiology from five raw signals, photoplethysmography (the green light on the back of your watch), movement, electrodermal activity (a sweat-based signal), skin temperature, and altitude, and that one picture is good enough to answer 35 different health questions. And when they tested it by bolting only the simplest possible model on top, it still beat the traditional hand-built approach on 34 of the 35. One generalist, almost no task-specific tinkering, outperforming the specialists. That’s a real shift in how wearable data might be used, and I don’t want to undersell it.

Who is the average human?

These 35 questions are grouped into six health categories:

• cardiovascular

• metabolic

• mental health

• sleep

• lifestyle

• demographics

The model is asked to predict the answers in each of these categories, even demographic attributes. The headline “34 of 35” is a single number, averaged across five million bodies. It tells you the model does well on average. It tells you very little about who it does well for. Additonally, who is the average?

The demographic tasks are to predict age, BMI, height and weight. It sounds odd — why predict someone’s height from their watch? — but in a benchmark like this it’s a test of how much real physiological information the representation has captured. Some major demographics are missing. None of the results is broken by sex, skin tone or age group, device model, or country. The authors acknowledge these limitations themselves. In their own words, the pretraining data is “skewed towards women, who are more frequent adopters of Fitbit devices, and White/Caucasians.” They note that reported model performance “may not generalize directly to the overall population,” and that future evaluation “will require... subgroup-specific validation.”

To be precise about it, Google couldn’t record ethnicity for the five million, it’s de-identified consumer data. What they do say is that the Fitbit user base itself skews White/Caucasian. So the skew is acknowledged but never measured. We have no idea what the true ethnic breakdown of those five million people actually was — only that it almost certainly wasn’t representative.

Am I just being picky?

Or does this stuff actually matter?

It matters at the level of the raw sensor. The green light on your smart watch? Melanin absorbs it differently, which is why wrist heart-rate accuracy is well documented to slip on darker skin, especially during exercise. Heart-rate variability? Men and women sit at measurably different baselines, so the same number can mean different things depending on whose wrist it came from. These are the physiology the model is built on.

Coming up

Sex differences. HRV, electrodermal baselines and sleep architecture all differ by sex, with different clinical meaning attached to the same raw numbers. Does one “general” representation hold, or quietly average a real biological difference away?

Skin colour and pigmentation. PPG is the best-documented case of a sensor that performs unevenly across skin tones. What does that mean for a model built on PPG as a core input?

Device heterogeneity. The paper itself calls consumer wearables “heterogeneous in both hardware and signal processing” — then pools 20+ models and never breaks performance down by device. Do a handful of newer watches carry the result?

Geography. The training data spans 100+ countries. The validation was run on US participants only. “Global data” and “validated globally” are not the same claim.

The synthesis. Aggregate wins tell you where a model succeeds on average; they’re silent about where it fails for someone specific. I’ll turn all of this into a short checklist you can point at the next foundation model — because there will be a next one.

None of this is a takedown. It’s a calibration exercise, which is rather the point of this publication. The most exciting technology in a field is exactly the technology worth reading carefully, and “trained on everyone” is not the same claim as “validated for anyone in particular.”

More soon

Reference: Narayanswamy G, Xu MA, et al. Towards a General Intelligence and Interface for Wearable Health Data. arXiv:2605.22759 (2026). Google Research. https://arxiv.org/abs/2605.22759

WHOOP recently published Cristiano Ronaldo’s heart rate at each goal in Portugal’s 5-0 win over Uzbekistan at the World Cup.

Greg Grosicki, PhD, Staff Research Scientist at WHOOP said:

“On goals scored by teammates, his heart rate sat 25 to 30 beats lower, which tells you how much of that spike is his own exertion plus the rush of putting the ball in the net”

Although this is theoretically true and is a good talking point, and I am all for physiological signals being used to discuss cognitive and emotional processes, a few questions have to be explored.

1. Is the difference between the goals significant?

2. What is actually happening to his exertion at these levels?

3. Would heart rate variability add anything?

4. Can five single HR readings tell us anything at all?

I ask because these are the questions that separate a physiological talking point from a physiological claim.

Five readings are not a result

I am working backwards here, but bear with me. Start with the last question, because it governs the rest. Five single readings, one per goal, cannot carry much. There is no distribution behind them, no spread. You cannot ask whether 176 differs meaningfully from 152, and you certainly cannot split either number into effort and emotion, because both questions need variance and there is none here. Five isolated points are an anecdote, not a result. That is fair enough. The data from WHOOP is not being presented as a controlled study. But we need to be careful with our words.

What the levels actually show

Read the levels on their own terms and they are unremarkable for elite match play. Ronaldo is 41, so his estimated maximum heart rate is around 179 (Tanaka et al., 2001), give or take about ten beats. His resting heart rate is 43 (WHOOP, Podcast 272). Against that range, 176 is 98 per cent of his heart rate reserve, near-maximal. The 164 is 89 per cent, hard and above threshold. The 146 to 152 cluster is 76 to 80 per cent, vigorous but clearly submaximal. When he scores he makes the run, the sprint and the strike. When a teammate scores he may be jogging or standing. That difference alone moves heart rate 25 to 30 beats at these intensities. Nothing in the data isolates an emotional increment on top of the effort.

The tell is in his own two goals

His own two goals are the tell. They came 33 minutes apart, at 6 and 39 minutes, and they differ by 12 beats. A constant rush of scoring would not produce that. The number is tracking the passage of play and the state of the match, not a feeling.

Heart rate also lags. It climbs over roughly 30 to 60 seconds toward the current demand and decays over minutes afterward. Each reading is a smear of the preceding half minute of play plus the decay from the last sprint. It is not locked to the instant the ball crossed the line, which is the one moment the framing needs it to be.

HRV would not help

An obvious suggestion is heart rate variability, but at these intensities it would not help. By near-maximal effort the vagal input that drives variability is already withdrawn, so HRV has fallen to its floor with little left to read (Tulppo et al., 1996; Michael et al., 2017). A ramping heart rate breaks the stationarity these measures assume, and clean beat-to-beat timing is exactly what wrist and arm optical sensors lose to motion. HRV carries the same non-specificity besides, since effort and arousal suppress it through the same pathway. It earns its place at rest, overnight and in recovery, as an index of autonomic load across a tournament. It does not earn it mid-sprint.

How you would read emotion properly

In my previous work I have used video annotations combined with physiological metrics to investigate this. I have done it in sports medicine and in the operating theatre.

If you wanted to read emotion off a heart rate properly, you would want the opposite of his situation. Low exertion, so the signal is not swamped by physical work. Low motion, so the sensor is clean. And the subject’s own report, so you have something to check against. A fan sitting still at a match is a better experiment than an athlete finishing a sprint. Even then it has a hard limit. A single cardiac measure cannot tell you which emotion is present. How far autonomic activity differentiates emotions at all is contested, and the evidence that it does comes from multi-channel recording under controlled conditions, not one heart rate value in a stadium (Kreibig, 2010).

Heart rate is not a specific signal. It is moved by exertion, core temperature, hydration, caffeine and posture as much as by any emotion, and on its own it cannot say which. No single channel separates the feeling from everything else raising the rate.

How well was it measured?

There is another problem with this data, and it is not unique to WHOOP. The optical sensors these devices use are least accurate in the conditions WHOOP was reporting from, high intensity and rapid changes in activity. In controlled testing, absolute error during activity ran around 30 per cent higher than at rest (Bent et al., 2020). Before anyone interprets the 176, the prior question is how well it was measured. ECG based sensors, a chest strap or clinical leads, read the heart’s electrical signal directly and are the reference standard optical devices are judged against. Optical estimates blood flow and infers the rate from it, which is why it drifts under motion.

Calibrate confidence to evidence

This is not one company’s failing. A 2025 review of composite consumer scores found that none of the fourteen evaluated across ten manufacturers had undergone rigorous independent validation (Doherty et al., 2025).

None of this means heart rate is useless, or that physiology has no place in a conversation about how someone felt. It means confidence has to be calibrated to the evidence. WHOOP has the data, the athlete and the reach to set the terms of this conversation. Whether those terms are earned is a separate question. Answering it, for any device making claims like these, is the work.

Thanks to Greg Grosicki for sharing this data and getting the conversation started.

References

WHOOP. Resting heart rate figure (43 bpm), WHOOP Podcast, Episode 272.

Kryger, K. (2026). WHOOP data on Cristiano Ronaldo’s heart rate during record-breaking goals. Men’s Journal. https://www.mensjournal.com/fitness/cristiano-ronaldo-whoop-data-heart-rate-response-during-record-breaking-goals

Bent, B., Goldstein, B.A., Kibbe, W.A., & Dunn, J.P. (2020). Investigating sources of inaccuracy in wearable optical heart rate sensors. npj Digital Medicine, 3, 18. https://doi.org/10.1038/s41746-020-0226-6

Doherty, C., Baldwin, M., Lambe, R., Burke, D., & Altini, M. (2025). Readiness, recovery, and strain: an evaluation of composite health scores in consumer wearables. Translational Exercise Biomedicine, 2(2), 128-144. https://doi.org/10.1515/teb-2025-0001

Kreibig, S.D. (2010). Autonomic nervous system activity in emotion: a review. Biological Psychology, 84(3), 394-421. https://doi.org/10.1016/j.biopsycho.2010.03.010

Michael, S., Graham, K.S., & Davis, G.M. (2017). Cardiac autonomic responses during exercise and post-exercise recovery using heart rate variability and systolic time intervals: a review. Frontiers in Physiology, 8, 301. https://doi.org/10.3389/fphys.2017.00301

Tanaka, H., Monahan, K.D., & Seals, D.R. (2001). Age-predicted maximal heart rate revisited. Journal of the American College of Cardiology, 37(1), 153-156. https://doi.org/10.1016/S0735-1097(00)01054-8

Tulppo, M.P., Mäkikallio, T.H., Takala, T.E.S., Seppänen, T., & Huikuri, H.V. (1996). Quantitative beat-to-beat analysis of heart rate dynamics during exercise. American Journal of Physiology-Heart and Circulatory Physiology, 271(1), H244-H252. https://doi.org/10.1152/ajpheart.1996.271.1.H244