Wearable AI Devices: Real-Time Health Monitoring for Patients

 A wearable that detects something you cannot act on has not improved your health. It has given you a job.

That sentence describes the experience of a large number of people currently wearing health-tracking devices. They receive a notification about an irregular reading, feel alarmed, book an appointment, undergo an investigation, and are told everything is fine. The device performed as designed. The user is left with a bill, an afternoon lost, and a slightly worse relationship with their own body.

This is not an argument against wearable AI devices. It is the necessary framing for understanding which of them genuinely help, because the difference between the useful and the anxiety-generating versions is not the quality of the sensor.

What a Wearable Actually Measures

Almost nothing a consumer device reports is directly measured. Nearly all of it is inferred, and the inference quality varies enormously between metrics.

• Heart rate is estimated optically, by measuring light absorption changes as blood flows through the wrist. Reasonably accurate at rest, less so during vigorous or irregular movement.
• Heart rate variability is derived from those beat intervals and is sensitive to how the device processes them. Trends are more meaningful than absolute values.
• Blood oxygen is inferred from light absorption at two wavelengths. Accuracy in consumer devices is variable, and research has documented that pulse oximetry can perform differently across skin tones — an issue that received significant regulatory attention after the pandemic.
• Sleep stages are estimated from movement and heart rate. Consumer devices detect sleep versus wake reasonably well and stage classification considerably less well, compared against clinical polysomnography.
• Calorie expenditure is a model output based on movement and personal characteristics. It is the least reliable common metric by a wide margin.
• Stress is not a physiological measurement at all. It is a label applied to a pattern in heart rate variability, and it should be read as such.

Knowing which numbers are measurements and which are estimates changes how much weight to place on them.

Where the Evidence Is Strongest

Three applications have accumulated genuine clinical support, and they share a common structure: a specific condition, an actionable finding, and a defined pathway afterwards.

Atrial fibrillation detection. Large studies including the Apple Heart Study identified previously unknown irregular rhythms in participants, and regulators have cleared ECG and rhythm-notification features on several consumer devices. Atrial fibrillation is common, often asymptomatic, and materially raises stroke risk — and it is treatable. That combination is what makes detection worthwhile.

Continuous glucose monitoring. The strongest case in the category. For people with diabetes, continuous measurement has substantially changed management, replacing intermittent finger-prick readings with a trend that shows what food, exercise, and medication actually do.

Seizure detection. Devices cleared to detect certain seizure types and alert a carer address a specific danger — a seizure occurring when no one is present.

Note what these have in common. The condition matters, the finding leads to a different action, and there is a clinician at the other end.

The Screening Problem in Healthy People

The mathematics of testing changes completely when you move from patients to the general population, and this is where most consumer health monitoring runs into trouble.

When a condition is uncommon in the group being tested, most positive results will be false — even with a good detector. This is a property of prevalence, not a flaw in the device, and it cannot be engineered away.

The consequences accumulate:

• Investigations for findings that would never have caused harm.
• Anxiety that persists after reassurance.
• Health system capacity absorbed by worried, well people.
• Incidental discoveries that trigger further investigation of their own.
• Reduced trust when the alarm turns out to be nothing, which may suppress attention to a real signal later.

There is a live debate in medicine about continuous monitoring in healthy people, sharpened by glucose monitors becoming available without prescription to people without diabetes. The concern is not that the data is wrong. It is that normal physiological variation, displayed continuously and without clinical context, is easily read as pathology.

What Continuous Data Genuinely Adds

Against all of that, there is a real and underappreciated advantage.

Clinical measurement is a snapshot. A blood pressure reading taken in a clinic is famously unrepresentative — the setting itself changes the value. A heart rhythm recorded during a ten-minute appointment misses an arrhythmia that occurs twice a week. A patient's account of their sleep is a recollection, not a record.

Continuous measurement replaces a moment with a pattern, and for several conditions the pattern is what matters:

• Home blood pressure readings predict outcomes better than clinic readings.
• Intermittent arrhythmias are detected by monitoring across days rather than minutes.
• Heart failure decompensation is preceded by measurable physiological changes.
• Post-operative complications often show in vital signs before the patient reports feeling unwell.
• Chronic condition management improves when the clinician can see what happened between appointments.

Orthosomnia and the Anxiety Cost

Sleep medicine has documented a condition informally named orthosomnia: patients whose sleep worsened because of their preoccupation with optimising the sleep scores their tracker produced.

The mechanism is straightforward. Sleep is disrupted by anxiety. A device that assigns your night a grade creates a performance metric, and performance metrics generate anxiety. The person then sleeps worse, receives a worse score, and worries more.

The broader pattern applies beyond sleep. Continuous self-measurement can shift a person's relationship with their body from experience to surveillance, and for some people that is a genuine harm rather than a minor side effect. Anyone with a tendency toward health anxiety, or a history of disordered eating around food and exercise tracking, should approach these devices with real caution — and ideally with clinical input.

Medical Device or Consumer Product?

The distinction determines almost everything about how much weight a reading deserves, and manufacturers have commercial reasons to keep the line blurred.

A regulated medical device has been through a formal authorisation process for a specific claim in a specific population. It carries labelled indications, documented performance, and a manufacturer accountable for accuracy. A blood glucose monitor prescribed for diabetes management sits here.

A consumer wellness product makes no medical claim and is not held to that standard. It can report the same underlying signal — heart rate, oxygen saturation, sleep — with no obligation to demonstrate clinical accuracy.

The complication is that many devices are both. The same watch may contain a cleared rhythm-notification feature alongside a dozen wellness metrics that were never evaluated by anyone. Users experience them as a single product with uniform authority. They do not have it.

How to tell what you are looking at:

• Check whether the specific feature — not the device — is described as cleared or approved, and for what.
• Look for the labelled indication: which condition, which population, what age range.
• Treat marketing language about "wellness," "insights," or "trends" as an indication that no medical claim is being made.
• Be aware that clearance in one country does not transfer to another.

Who Reads the Data?

The unresolved operational problem is what happens to all this information.

A patient arrives at a ten-minute appointment with six months of continuous readings. The clinician has no time to review it, no tool to integrate it into the record, and no established standard for interpreting consumer-grade data. In many systems there is no reimbursement for the time it would take.

What functional programmes have in common:

• Data flows into the clinical record automatically rather than being shown on a phone screen.
• Automated filtering surfaces only what requires attention.
• A defined person is responsible for responding, with a defined timeframe.
• Thresholds are set for the individual patient, not from a generic default.
• Patients are told clearly what is monitored and what is not — a device is not an emergency service.

Without those, a monitoring programme generates data nobody owns, which is a liability rather than a benefit.

The Privacy Gap

Health data collected by a consumer device frequently sits outside the regulations that protect data held by a hospital. The protections that apply to medical records generally do not extend to a fitness company's servers, and the terms governing that data are set by a commercial agreement rather than by health law.

Reasonable questions before committing to a device:

1. Who holds the data, and in which country?
2. Can it be shared with or sold to third parties, including insurers and employers?
3. What happens to it if the company is acquired or ceases trading?
4. Can it be exported in a usable format, and deleted on request?
5. Does the device function if the company's service shuts down?

That last question is not hypothetical. Several health devices have been discontinued, leaving users with hardware that no longer works.

Using One Sensibly

1. Decide what question you are answering before buying. A device bought without one produces data without a purpose.
2. Read trends, not single readings. Individual values fluctuate for reasons that have nothing to do with health.
3. Know which metrics are estimates. Weight them accordingly.
4. Do not self-diagnose from a notification. Take it to a clinician as one piece of information.
5. Turn off what you do not need. Fewer, more meaningful alerts.
6. Notice the effect on you. If the tracking is increasing anxiety, that is a reason to stop, not to optimise harder.

The Balanced View

For a person managing a diagnosed condition, with a clinician who receives the data and a plan for what to do with it, wearable AI devices can substantially improve care. That case is well supported.

For a healthy person monitoring themselves continuously without clinical context, the picture is far less clear — a genuine chance of catching something important, alongside a substantial chance of finding something that was never going to matter and paying for it in worry and investigation.

Both of those are real. Which one you experience depends less on the device than on whether there is a clinician at the other end and a plan for what happens next.