Wearables & sensors

In one line

Wearables moved physiological measurement out of the clinic and onto millions of wrists. The signals are often genuinely good. The unsolved problem isn't the sensor — it's what happens to the alert.

The line between consumer and clinical has gone

It used to be simple: consumer devices counted steps, medical devices measured. That distinction is dead. Consumer wearables now do single-lead ECG, irregular-rhythm notification, SpO₂, continuous glucose, sleep staging — and several have regulatory clearance for specific claims.

The important nuance: clearance is per-claim, not per-device. A watch cleared to notify you of irregular rhythm is not cleared to diagnose atrial fibrillation, and its step counter isn't cleared for anything. The same object is a medical device for one function and a toy for the next — which is exactly the intended-use logic that governs SaMD generally.

The Apple Heart Study, and what it really showed

The largest reference point: ~419,000 participants, published in NEJM (2019).

The headline was that a watch could identify irregular pulse. The finding that matters is in the detail: of those notified, around a third had AF confirmed on subsequent ECG patch monitoring. Among those over 65 it was higher.

Read that honestly and it says two things at once. It is a real signal — a consumer device detecting a serious, treatable, often-silent condition at population scale is a genuine achievement. And it is a large false-positive burden — most people notified did not have confirmed AF at that moment, and each one generated anxiety, a consultation, and a workup.

Neither half is the story on its own. That's the shape of every wearable finding: real signal, real noise, and the ratio depends entirely on who's wearing it.

The base-rate problem, which nobody wants to discuss

Here's the maths that decides whether a wearable helps or harms.

Screening a low-prevalence condition in a low-risk population produces mostly false positives — however good the sensor is. This isn't a criticism of the device; it's arithmetic. A healthy 25-year-old getting an AF notification is far more likely to be a sensor artefact than an arrhythmia, because 25-year-olds mostly don't have AF.

So the same watch is a useful screening tool on a 70-year-old and an anxiety generator on a 25-year-old. The device cannot tell the difference. The population does.

Which means the interesting question was never "is the sensor accurate?" It's "who should be wearing it?" — and consumer distribution answers that question with "whoever buys one."

The deluge, and the 3am problem

This is where digital health repeatedly under-thinks.

A patient's CGM streams every five minutes. Their watch flags an irregularity at 03:14. The data arrives at your platform. Who is looking?

If the answer is "nobody until Tuesday", then you have created a record that an alert existed and was not acted upon — which is a clinical and medico-legal position materially worse than not having collected it. You cannot un-know a signal you received.

So before ingesting any wearable stream, the questions in order are:

1. Who reviews it, in what timeframe, and are they staffed and paid for that?
2. What is the escalation path when the signal is real at 03:14?
3. What is the documented, agreed answer for signals nobody reviews?

Most remote monitoring programmes fail here, not at the sensor. The technology arrives years before the staffing model, and the staffing model is the product.

Data quality, briefly

Wearable data is messy in specific ways: worn intermittently, worn wrong, worn by someone else, artefacts from movement, and gaps whenever it's charging. Absence in a wearable stream means nothing at all — the same non-event problem — and a model that treats missing as zero has learned something false.

And as with any sensor, the number on the screen is the end of a long chain of inference, not a measurement of the patient.