Wearables Catch Stress About Three Times in Four

Not reliably. Wearable AI built to detect stress in students had a pooled mean accuracy of 0.856, with sensitivity of 0.755 and specificity of 0.744. The authors call that suboptimal. In a 40-person lab test of a research wristband, wrist skin conductance showed neither correlation nor visual resemblance with a finger sensor. Consumer stress scores themselves have little published validation, and none of the manufacturers in one review disclosed their formulas.

What signals does a wearable stress score use?

Mostly heart rate variability, sometimes skin conductance. A systematic review of devices that detect stress and anxiety found that the included studies unanimously used heart rate variability parameters. Stress was detected better with HRV and EEG together.

Electrodermal activity was used in recent studies. The review says it had high accuracy for stress detection but questionable reliability. It also found that average heart rate, which many commercial devices use, was less accurate for stress and anxiety than HRV, electrodermal activity and possibly respiratory rate.

The scores themselves are opaque. A review of composite health scores, including stress monitors, found that none of the manufacturers disclosed exact algorithms. Few provided empirical validation.

How accurate is wearable stress detection?

Moderate at best in the pooled data. A 2024 meta-analysis reviewed wearable AI for detecting stress in students. It included 19 of 327 retrieved studies. The pooled analysis used 37 accuracy estimates from 6 of them.

Pooled mean accuracy was 0.856 (95% CI 0.70 to 0.93). Sensitivity averaged 0.755, specificity 0.744 and F1-score 0.759. The authors concluded that wearable AI shows promise but currently has suboptimal performance.

Accuracy varied with the number of stress classes, the type and location of the device, the data set size and the ground truth used. Each of those choices changes what counts as stress in the test.

Mental effort produces a similar response. In 892 people in a driving simulator, added cognitive load raised heart rate and lowered RMSSD, with heart rate the better discriminator over 30-second windows (Scientific Reports, 2024), and a meta-analysis of cardiac workload measures found HRV sensitive to task load, event rate and task duration (Human Factors, 2020).

Can a wrist sensor pick up stress signals reliably?

It depends on the signal. The Empatica E4 research wristband was tested against ECG and a finger skin conductance sensor in 40 healthy participants. The conditions included seated rest, paced breathing, a Stroop task, a speech task, slow walking and typing.

Mean heart rate was the most accurate measure across all conditions. HRV was satisfactory at seated rest, paced breathing and recovery, but not in dynamic conditions such as speaking. Accuracy fell with wrist movement, cognitive and emotional stress, and larger wrist circumference.

Wrist skin conductance showed neither correlation nor visual resemblance with the finger signal. The authors suggest the two sites may reflect different phenomena. The stress conditions that a score aims to detect were the same conditions that lowered accuracy.

For a related question, see whether a wearable can tell excitement from anxiety.

The signal itself is not fully understood. A methods review notes that the mechanisms generating electrodermal activity are still not completely explained and that a quality gap remains between wearable EDA and standardized laboratory recording (Physiological Measurement, 2022). The field’s own publication standards treat ambulatory recording as a separate case with its own problems (Psychophysiology, 2012).

Do heat and exertion look like stress to the sensor?

They can. Thermoregulatory mechanisms change electrodermal activity independently of emotion. A 2026 study used the WESAD dataset to correct EDA for temperature with two modeling approaches.

The corrected features improved separation between stress-related activity and heat-related activity. Temperature-derived parameters were reliable contributors to classification. The abstract reports no accuracy figure. The authors present temperature compensation as an essential preprocessing step.

That is a research result on a lab dataset. It shows the confound is real. It does not show that any consumer score applies such a correction.

Breathing, movement and speech, electrode problems, and the environment, including temperature, humidity, noise, air pressure and altitude, all shift electrodermal readings, and wearable instruments are the most exposed to them (Physiological Measurement, 2026).

What does this not settle?

The pooled accuracy comes from student samples and research-grade or prototype setups. It does not describe a specific consumer score. The E4 test covered 40 healthy people in lab conditions and slow walking.

Nothing here tests whether a consumer stress number tracks self-reported stress or cortisol across an ordinary week. Skin conductance is a well-established index of sympathetic arousal. Arousal has more than one source. A stress label is an inference on top of it.

Common questions

KnowYourPrime Evidence Registry: Finding KYP-ER-0006: Work Stress and HRV · Finding KYP-ER-0005: Readiness and Recovery Score Validity