Gains:
- Ability to explain the opportunities and pitfalls in processing wearable sensor data (PPG, accelerometer, skin temperature) with artificial intelligence
- Ability to evaluate motion artifact, skin tone bias, and non-medical device limits
- Ability to manage false alarm burden and need for clinical verification of continuous monitoring alerts
Wearable devices (smart watches, wristbands, patches, rings) have brought physiological measurement outside the hospital and into daily life. These devices continuously measure heart rate, oxygen saturation, activity, sleep, and sometimes ECG. For the biomedical engineer, this is an unprecedented source of data: not just a few minutes in the laboratory, but weeks of continuous monitoring in real life. But this data is far from the controlled conditions of the clinical environment. In this unit, we will see the opportunities, the most important pitfalls (motion artifact, skin tone bias, device class) and false alarm burden in processing wearable sensor data with artificial intelligence.
Let's make a distinction up front: most consumer wearable devices are not medical devices; It is for "healthy living/wellness" purposes. Some features (certain ECG or atrial fibrillation notifications) may have received regulatory approval, but even these are not diagnostic but refer to a physician. Artificial intelligence generates trends and awareness in this data; The diagnosis belongs to the doctor.
Sensor Types and What They Measure
PPG (photoplethysmography): Sends green/infrared light to the skin and estimates pulse rate and indirectly oxygen saturation from the change in reflection. It is cheap and common, but very sensitive to movement.
Accelerometer: Measures movement and position; It is used for step counting, fall detection and sleep/activity classification.
Skin temperature and electrodermal activity: Temperature trends and conductivity due to perspiration; It gives indirect clues in stress and cycle monitoring.
Single-lead ECG: Some devices take a short-term, single-lead ECG with finger contact. This is not a replacement for the 12-lead diagnostic ECG; It is a limited scanning tool.
The Three Most Important Traps
Motion artifact. PPG signal is distorted while walking, waving or driving. The model may mistake motion noise for a heartbeat. So in wearable AI, accelerometer data makes PPG question “is it trustworthy?” (sensor fusion).
Skin tone and physical bias. Optical sensors may be less accurate with darker skin tones, tattoos, and low perfusion (cold, poor circulation). If a model is trained with mostly light-skinned, young, healthy users, it will make systematic errors across different groups. This is a matter of justice and patient safety.
Device class and context. Consumer device data is not as accurate as clinical monitor data. Labels such as "sleep stage" or "stress score" are often rough estimates and have not been validated for clinical diagnosis.
Three Mini Cases: By the Numbers
Case 1 — False alarm load. An atrial fibrillation notification feature generated thousands of notifications per month in the active user group; A small proportion of the reports were genuine when confirmed by a confirmatory medical ECG. The fact that many healthy users were burdened with unnecessary anxiety and unnecessary admission to the clinic reminded us that with low pre-test probability, many positive reports may turn out to be false. The notification was a screening sign, not a diagnosis.
Case 2 — Pulse disturbed by movement. In a running study, the PPG heart rate locked onto the arm swing frequency, showing 130 instead of the actual 160. When accelerometer fusion was added, the heart rate reliability flag was lowered during movement-intensive periods and erroneous values were not reported. Sensor fusion significantly increased accuracy.
Case 3 — Trend catching. In a post-ICU discharge follow-up, the wearable device's slow increase in resting heart rate over several days was noted before the development of an infection, allowing early evaluation of the patient. The value here was not a single measurement, but a trend relative to one's own baseline; However, the decision was made based on physician evaluation.
Weak Prompt / Strong Prompt
Weak prompt:
Look at the data of my watch, am I sick?[data]
Powerful prompt:
Your role: You are a wearable data analysis assistant (YOU ARE NOT DIAGNOSIS, not a medical device). Evaluate the following anonymous wearable measurements:- Interpret the values as TREND relative to the person's own baseline, attributing diagnosis to a single measurement.- Mark possible periods of motion artifact and low signal quality.- Indicate in each printout that the device has not been validated for medical diagnosis.- List separately the "physician consultation recommended" flags at the end. Anonymous measurements:[HR trend, SpO2, activity, sleep, signal quality]
Four Copiable Templates
1) Trend analysis:
Summarize the following 14-day resting heart rate series as a trend according to the person's baseline; Mark significant deviations. Don't diagnose. Series: [data]
2) Signal quality flagging:
In this PPG heart rate series, mark as “low confidence” periods of possible motion artifact based on accompanying accelerometer data. Data: [serial]
3) False alarm evaluation:
For a notification feature, draft a briefing note explaining how many positive notifications may be false positives at a low pre-test probability.
4) User information text:
Write on-screen text that presents the wearable device output: "This is not a medical diagnosis; if you have symptoms, consult a physician" in plain language.
Model's Role: By Task Type
Task type
AI contribution
criticality
verification
Activity/step classification
high
low
spot check
Resting HR trend
high
medium
Baseline comparison
Sleep stage estimation
medium
low
Rude, not clinical
Rhythm notification (scan)
medium
high
Medical ECG + physician
diagnostic decision
None
very high
Clinical evaluation
Tip: Look for a trend in wearable data over the person's own history, not a single snapshot. The most valuable signals are a person's deviation from his baseline; The population mean may mislead individual interpretation.
Caution: Just because a wearable device has a “certified feature” does not mean that all of its outputs are clinically accurate. Approval is valid only for a specific function and use; The remaining metrics are mostly estimates for wellness purposes.
Common mistakes
- Mistaking consumer device data for clinical data. Most devices are not validated for diagnosis.
- Not filtering out motion artifact. Without accelerometer fusion, PPG heart rate would be misleading.
- Ignoring skin tone bias. Optical sensors operate with different accuracy in different groups.
- Not taking into account false alarm burden. At low pretest probability, many of the positives may be false.
- Attributing diagnosis to a single measurement. The value should be interpreted as a trend and the decision should be left to the physician.
In summary
- Wearable devices provide continuous, real-life data, but most are not medical devices.
- PPG is prone to motion artifact; sensor fusion (with accelerometer) improves accuracy.
- Skin tone bias in optical sensors is a fairness and safety issue.
- At low pretest probability, many of the reports may be false positives; false alarm burden must be managed.
- The largest value is in the trend; Diagnosis is always made by physician evaluation.
Application task
Create a 14-day resting heart rate series (includes a slow increase from baseline) from a wearable device. Using the powerful prompt, have the model interpret this series as a trend and see if it diagnoses a single day. Then describe the clinical limits of the device and the “consult physician” threshold in a paragraph.
checklist
- [ ] I know common wearable sensors (PPG, accelerometer, temperature, single channel ECG) and their limits.
- [ ] I understood motion artifact and the role of sensor fusion.
- [ ] I understand that skin tone bias is a fairness and safety issue.
- [ ] I can evaluate the false alarm burden and the impact of low pre-test probability.
- [ ] I internalized that I interpreted the wearable data as a trend and left the diagnosis to the physician.