Wearable HRV and readiness scores lack proof, cardiologist Eric Topol writes

The critique lands as Apple expands exactly those features. On 9 September the company said it was revamping the Apple Watch health sensing system. It added a Readiness score from 0 to 10 and raised how often it reports heart rate variability by a factor of 24, according to Topol's Ground Truths post. He notes the same announcement introduced a longevity tab and a "Health Age".

His argument is narrow and technical. The numbers are real. The meaning attached to them is not yet.

Heart rate variability is the millisecond-scale variation in the time between heartbeats. It is driven by the interplay of sympathetic and parasympathetic nerve inputs. Topol writes that more than three decades of research have associated low HRV with higher all-cause and cardiovascular mortality. The link is looser with early cognitive impairment, dementia, mental illness, Type 2 diabetes and substance abuse. But he is blunt about the limit: HRV is a surrogate marker, and no cause-and-effect relationship has been established. Raising your HRV, he writes, does not mean your health outcomes will improve. "In fact, there is no hard evidence for that."

That matters because the marketing has moved faster than the evidence.

Most of the outcome research Topol cites used HRV derived from an electrocardiogram. Consumer wearables measure something else: pulse rate variability from optical plethysmography, the green lights on the back of the watch. The two are not the same, he writes. They are only concordant when the delay between the electrical signal and the pulse stays constant, which in practice means at rest. Mattress sensors that infer HRV from movement are further removed again.

Even within HRV there is no single measurement. Different devices use different metrics. Apple has relied on SDNN, the standard deviation of normal-to-normal intervals, while nearly all others use RMSSD, which Topol describes as generally considered the more accurate metric. Recording windows differ too, from a few minutes to a full night's sleep.

Studies are small and short

The post also questions the research behind the composite scores. Topol writes that the studies underpinning readiness metrics tend to be small and short. A composite number bundling several inputs hides which one is doing the work. He points readers to the underlying papers rather than to the product pages.

The wearable market has been heading this way for a while. The pitches have grown broader. Apple is not alone in attaching a single number to recovery, strain or biological age.

Independent work on what wearables actually capture is thin. Some of it cuts against the framing. A scoping review of wearable device data evaluated in clinical studies is among recent preprints on the subject. A separate study circulating this autumn found that step counts outperformed smartwatch intensity algorithms in health phenotyping. Neither settles the HRV question. Both suggest the derived metrics are doing less work than the raw signal.

Topol's post is the most direct public challenge to the category from a prominent cardiologist in some time. It is also a familiar shape of argument in digital health. A measurement that is genuinely physiological gets promoted into a promise about lifespan.

Where the evidence is stronger

Other parts of the field show what better-supported digital health claims look like. They are far from a wrist score.

Northwestern University researchers reported in January a soft, flexible robotic probe, about three times the width of a human hair. It can be inserted through the port already used in fetal surgery. It continuously tracks a fetus's heart rate, heart rate variability, blood oxygen and temperature. The study was published on 26 January in Nature Biomedical Engineering. "Our ability to monitor the fetus hasn't changed in 40 years," fetal surgeon Dr Aimen Shaaban said in the university's announcement. The device has been tested in a large animal model, not in people.

On the public health side, the evidence is epidemiological and unglamorous. A study of more than 2.5 million children was published in The Pediatric Infectious Disease Journal and reported by PsyPost on 4 October. It found no association between receiving the MMR vaccine before age two and autism, and no link between the timing of the shot and diagnosis. Lead researcher Todd Burstain, a physician at UW Medicine, told PsyPost that earlier European studies without age anchoring "could potentially introduce an immortal bias into the data".

And in Canada, a federal consultation reported data on young men. Self-reported mood disorders among men aged 15 to 24 rose from 11 per cent to 16 per cent between 2012 and 2022, CBC News reported on 5 October. Generalized anxiety climbed from four per cent to 10 per cent. The share describing their mental health as very good or excellent fell from 70 per cent to 52 per cent. "There's a common problem everywhere and it's how men and boys are feeling," federal Health Minister Marjorie Michel told CBC.

None of that arrives as a score out of 10.

Elsewhere in the stack

The rest of the week's material is a reminder of how much of digital health rests on infrastructure nobody markets.

DigitalOcean has quietly ended its Open Source Credits programme. These are the cloud credits that projects used to run build servers, documentation sites and release mirrors. According to It's FOSS, maintainers of Node.js received an email. It said the company is no longer accepting new applications or approving renewals, extensions or additional credit requests. It also said the [email protected] inbox will no longer be monitored. There was no blog post or social media announcement. The publication reported that the lead developer of Pidgin confirmed the email on Mastodon.

On the regulatory side, the privacy group noyb published a leaked Irish Presidency document. Noyb says it would make use of personal data "in the context of AI" automatically lawful across the EU. Companies would be assumed to have an overriding legitimate interest rather than needing consent. "Under these proposals, the profits of AI companies would trump European's fundamental right to privacy," noyb founder Max Schrems said in the group's statement. The document is dated 21 September and is with the Council.

Researchers at UC Riverside, meanwhile, reported in September on large language models. Confidence and correctness arise from different internal features, they found, after examining Meta's Llama-3.1-8B and Google's Gemma-2-9B. "A model can answer with certainty, but also be wrong," lead author Het Patel said in the university's release. Suppressing features tied purely to uncertainty sharply reduced accuracy.

Google, for its part, added Oklahoma state IDs to Google Wallet on Android on 5 October, per 9to5Google. It is the thirteenth option in the list. The digital ID works at some TSA checkpoints, but the state still requires users to carry the physical card.

Two smaller entries point the same direction as Topol. A Kenyan health assistant called Tabibu answers questions from published Ministry of Health, Social Health Authority and Pharmacy and Poisons Board documents. It shows its sources and states plainly that it does not diagnose. A personal security writeup, Triple Sec, proposes a 3-4 word passphrase plus a 6-digit PIN for roughly 60 bits of entropy. Both are attempts to make a system's limits legible before someone relies on it.

That is the gap Topol is pointing at. A watch can measure the interval between beats with real precision. What it cannot yet do, on his reading of the literature, is tell you what that interval means for how long you will live.