Can Your Wearable Tell You're Getting Sick? It Can Spot the Change, but Most Alerts Will Be Wrong

You wake up feeling fine. Your watch disagrees: two of your overnight vitals are outside your normal range. Apple’s Vitals app sends exactly this notice, and it lists illness among the possible causes [6]. By lunchtime you either have a sore throat or you remember the three glasses of wine.

Both outcomes are common, and the science explains why. Wearables really can detect the body’s response to an infection a day or more before symptoms. Research models built on that response catch most infections. But the same models, run across a whole user base in a winter week, fire far more often for people who are not getting sick than for people who are. For anyone building on health data, the second fact matters more than the first.

What changes before you feel ill?

An infection triggers the immune system before it causes symptoms, and the immune response changes things a wearable measures [1][2]:

• Resting heart rate rises. In a small Stanford subset with Fitbit data, COVID-19 diagnosis was associated with a median increase of 7 beats per minute in the days before symptoms [1].
• Heart rate variability falls, and skin temperature and breathing rate rise [3].
• Activity drops and sleep lengthens, often before people consciously feel unwell [1].

The research on how early and how often this shows up is consistent:

Most of this work was done on COVID-19 between 2020 and 2022, when large cohorts were wearing devices and getting tested. The physiology is not specific to COVID-19, and the same changes appear with flu and other respiratory infections [2][5].

What wearables do with this today

Two of the most widely used features show how carefully the industry handles it:

• Apple Watch Vitals. The app builds a typical overnight range for heart rate, respiratory rate, blood oxygen, wrist temperature and sleep duration, which takes about a week of nightly wear. It notifies the user when at least two metrics fall outside that range, and lists medications, elevation changes and illness as possible factors [6].
• Oura Symptom Radar. It compares heart rate variability, temperature, breathing rate and heart rate with the user’s baseline each morning, needs seven nights of data within 14 days to start, and reports no signs, minor signs or major signs. Oura calls it an early warning, not a diagnosis, and it cannot tell a cold from flu or COVID-19 [7].

Both are built to say “your body is under unusual strain”, not “you are getting sick”. The reason becomes obvious once you count the alerts.

Count the alerts, not the catches

Sensitivity tells you how many sick people get flagged. What a user experiences is the opposite question: when an alert fires, how often is it right? That depends on how many people are actually getting sick, and in any single week it is a small share.

Adults average two to three colds a year, concentrated in winter [8]. Take a winter week in which 2 in 100 adults develop a respiratory infection, and apply the TemPredict model’s accuracy to them:

• Main model (82% sensitivity, 63% specificity): about 38 alerts per 100 people, 2 of them real. About 5% of alerts are right [3].
• Validation group (90% sensitivity, 80% specificity): about 22 alerts, 2 of them real. About 9% are right. That group had only 10 confirmed cases, so treat it as a best case [3].

This is the base rate problem, and no model escapes it. A product can only improve the odds by raising the threshold, which is what Apple does by requiring two metrics and Oura does with its “major signs” level. That cuts false alarms and misses more real infections. There is no setting that catches nearly everyone and is usually right.

Why the alert fires when you are not sick

The body’s response to infection overlaps with its response to many ordinary things. In the Stanford real-time study, alerts were also triggered by other infections, stress, alcohol and travel, although at a lower rate than by COVID-19 itself: 1.15 alert days per person against 3.42 for infected people [2]. The common false triggers:

• Alcohol, which raises resting heart rate and lowers heart rate variability overnight.
• Stress, poor sleep and travel, including time zone changes [2].
• Hard training, particularly an unusual session.
• Vaccination, which produces a short immune response much like a mild infection.
• The menstrual cycle. Temperature and resting heart rate rise in the luteal phase, so a temperature-led alert needs to know where the user is in her cycle.
• Medication. GLP-1 drugs raise resting heart rate by about 3 beats per minute on average, as we covered in GLP-1 pills and what changes for programs.

Every one of these can be known or asked about, which is the most useful thing a product can do with them.

What products should do with an illness signal

1. Report strain, not illness. “Your resting heart rate and temperature are higher than usual” is accurate. “You may be getting sick” will be wrong for most people who see it.
2. Require more than one signal, for more than one night. Agreement between heart rate, heart rate variability and temperature, sustained over two nights, trades a little speed for fewer false alarms. Apple’s two-metric rule is the minimum [6].
3. Ask a question before you assert anything. A one-tap check-in on symptoms, alcohol, travel or a hard workout turns a vague alert into a useful one. DETECT found that combining sensor data with symptoms beat symptoms alone [4], and the check-in is how a product gets the symptoms.
4. Make the response cheap to be wrong about. Suggesting rest, an easier workout or an earlier night costs nothing if the alert is false. Telling someone to test, isolate or see a doctor does.
5. Know the user’s context. Cycle phase, medications and recent vaccination should change the threshold, not just the explanation.
6. For employers and insurers, aggregate. At the level of a population, raised resting heart rate across many users tracks flu activity well [5]. At the level of an individual employee or member, most signals are false, and acting on them invites both error and a privacy problem.

Wearables do something genuinely remarkable: they see the body fighting an infection before the person does. But the same sensitivity that catches the infection also catches the wine, the deadline, the flight and the hard run. In any given week most people are not getting sick, so most alerts are about something else. The products that use this well treat it as a signal of strain worth responding to gently, not a diagnosis worth announcing.