How wearables measure stress: what PPG research shows

Stress is one of the most popular features on consumer wearables and one of the least settled in research. The sensor does not measure stress. It measures the pulse, and stress has to be inferred from how the pulse behaves.

On our map of wearable PPG research, stress, emotion and cognitive load form a cluster of ten wearable studies. Most are lab studies with small groups. This review reads those studies alongside the wider heart rate variability literature and explains what a research team can reasonably expect from PPG-based stress measures. For a vendor-neutral look at the same question, see whether a smartwatch can detect stress.

What happens in the body under stress

When a person faces a challenge, such as a deadline, a hard conversation or a mental arithmetic test, the autonomic nervous system shifts. Sympathetic activity rises and the parasympathetic (vagal) brake on the heart eases. Heart rate tends to go up, beat-to-beat variation tends to shrink, small arteries constrict, and breathing changes.

A meta-analysis of 37 publications found that heart rate variability changed under stress in most studies. The most common pattern was lower parasympathetic activity, seen as less high-frequency variability 1. The standard definitions of these measures go back to a 1996 task force report 7, and a later overview explains how each metric relates to the nervous system and why norms vary so much between people 8.

For more on the parasympathetic side, see what vagal tone means.

What PPG can see

PPG captures several stress-related changes at once.

Pulse rate and pulse rate variability. The time between pulse peaks stands in for the time between heartbeats. At rest, pulse rate variability tracks heart rate variability well. During movement, agreement drops 9.

Pulse shape. Vessel constriction and changes in heart output alter the shape of each pulse. A numerical modeling study simulated pulse waves at six levels of stress and found 17 waveform features that trended with stress at one or more sites. Three changed at every site: time from pulse onset to peak, time from the dicrotic notch to the end of the pulse, and pulse rate. The authors stressed that in vivo studies were needed to confirm these findings 3.

Breathing. Breathing rate can be estimated from PPG and often changes with stress. See the wearable sleep apnea research review for how PPG picks up breathing.

Not every heart rate variability metric survives the move from ECG to PPG. A study of an ear-worn PPG sensor during rest and a mental stress task found moderate to good agreement with ECG for average interbeat interval and some slower variability measures, but poor agreement for fast beat-to-beat metrics such as RMSSD and high-frequency power. Most PPG metrics still showed a significant change during the stress task 10. The fast metrics are the ones most tied to parasympathetic activity, so this gap is a real limit.

What the stress studies found

The ten studies in the map cluster and the reviews around them show a consistent picture.

Lab protocols dominate. A scoping review of PPG wearables worn on the wrist or ear found 23 eligible studies. Four were done in real life, eighteen in clinical or lab settings and one in both 2. The review concluded that much more research was needed before wearable stress monitoring is reliable.

Strong lab results, narrow settings. A machine learning model predicted state anxiety scores from PPG features recorded at the arm with a custom system, with a correlation of 0.81 in a leave-one-out test 11. That is a promising signal from a small lab study, not proof that a wrist device can track anxiety.

Personal models beat generic ones. One study turned interbeat intervals from PPG into images and trained a neural network to classify stressed and non-stressed states. Person-specific models performed far better than generic ones. Adding a small amount of each person's own data to a generic model raised test accuracy to about 88 percent 4. The authors' conclusion was that stress shows up differently in each person.

Other contexts. Studies in the cluster also looked at driver anxiety using several biosignals together 12 and at cognitive load during game playing with PPG 13. These widen the use cases but use small samples and set tasks.

Combined signals help. A systematic review of smart devices for stress and mental health found that studies of stress and anxiety used heart rate variability, and that average heart rate, which many commercial devices rely on, was less accurate than heart rate variability or electrodermal activity 5.

Datasets that shaped the field

Much of the stress algorithm work uses a few public datasets. WESAD recorded wrist and chest sensors, including the PPG-based blood volume pulse, in 15 people during a lab study with neutral, stress and amusement conditions 6. K-EmoCon collected physiological signals during naturalistic conversations with continuous emotion annotations 14. Shared datasets make methods comparable, but they are small, and models tuned on them often lose accuracy on new people and new settings.

For a technical overview of feature and model choices, see this review of PPG stress detection methods and this summary of the stress evidence.

Why most results stay in the lab

Three problems keep stress detection from moving cleanly into daily life.

Ground truth. In a lab, the stress task is known. In daily life, researchers rely on self-reports, which are sparse and subjective. Without a solid label, it is hard to know whether a model is right.

Confounders. Exercise, caffeine, alcohol, heat, illness, posture and poor sleep all change pulse rate and variability in ways that look like stress. A model that does not account for them will misread a brisk walk as a stressful meeting.

Individual differences. Resting variability differs widely between people and changes with age 8. A drop that is large for one person is normal for another, which is why personal baselines work better than population models 4.

Questions to ask of any stress study

When a paper or product claims to detect stress from PPG, a few questions sort strong evidence from weak.

• Where was it tested? A seated lab task with a known start and end is much easier than a commute, a workday or a night of poor sleep.
• What was the ground truth? A validated questionnaire, a cortisol sample or a scripted task are stronger labels than a single end-of-day rating.
• How many people, and who? Ten young volunteers do not represent older adults, shift workers or people with chronic conditions.
• Which signal did the work? If the model mostly used average heart rate, exercise and caffeine will fool it 5.
• Was the model tested on new people? Results on the same people a model was trained on are usually far better than results on strangers 4.
• What happened during movement? Many studies drop noisy segments. Ask how much data was excluded and why.

Few studies answer all six well. That is not a reason to ignore the field. It is a reason to treat stress estimates as hypotheses to test in your own population.

How to design a stress study with wearable PPG

Start with a known task. Include a standardized lab stressor for each participant to anchor their personal response before free-living monitoring.

Collect labels in daily life. Use short, frequent self-reports or event markers, timed so they can be matched to the signal.

Log confounders. Activity, posture, caffeine, sleep and temperature all matter. Motion data help separate physical effort from mental stress.

Keep beat-level data. Fast variability metrics are sensitive to missed or extra beats. Raw waveforms let you inspect beats, reject noise and choose metrics that survive at the wrist. As an example, Sensor Bio V2 records raw green, red and infrared PPG at 25, 50 or 100 Hz and derives heart rate and heart rate variability, and partner teams can export the waveforms for their own analysis. V2 is a general wellness device and is not FDA cleared, and it is not intended to assess any mental health condition.

Model within people. Compare each participant with their own baseline. Report generic and personal model results separately.

Think about the intervention. Many studies pair stress tracking with a response such as paced breathing. See HRV biofeedback and how researchers use continuous HRV.

How these signals are measured: the science of PPG. For every research area, return to the wearable PPG research map.

For research teams

Put the raw signal in your study or product

Building a stress or resilience study? We can help you plan beat-level PPG capture and export.

Talk to Sensor Bio

Limits and open questions

• Most studies are small and lab-based, with short, artificial stressors.
• Daily-life studies lack reliable ground truth.
• Fast heart rate variability metrics are the least reliable from wrist PPG and the most tied to parasympathetic activity.
• Generic models perform poorly across people, and personal models need calibration data.
• Stress, anxiety and emotion are different constructs, and studies often blur them.