Exercise physiology · Wearables · Data science

What does a wrist wearable really see during stress and exercise?

An analysis of heart rate, electrodermal activity, skin temperature and movement recorded with an Empatica E4 during acute psychological stress, aerobic cycling and anaerobic sprints. Physiology comes first, every result is computed per participant, and machine learning is only evaluated on people it has never seen.

Acute stress · 36 participants Aerobic cycling · 30 Anaerobic sprints · 31
View the code on GitHub → Notebooks Jump to findings
12,34630-s windows, quality-checked per signal
+5 bpmHR response to a mental-stress task, with the wrist stiller than at rest
3%of sprint windows with a usable heart rate from the wrist
0.83 → 0.66macro-F1 once the model is tested on unseen people
The question

Physical activity is not the same as physiological state

An accelerometer tells us that someone is moving. Heart rate, electrodermal activity and skin temperature tell us how the body is responding. The main question: how do wearable physiological signals differ between aerobic exercise, anaerobic exercise and acute stress? A follow-up runs through the whole project: how much of any difference is physiology, and how much is simply movement?

Movement

3-axis accelerometer (32 Hz). Context: pedalling rhythm, sprint bursts, stillness.

Cardiovascular

Wrist PPG (64 Hz) → heart rate and beat intervals: load, sympathovagal balance, HRV.

Autonomic & thermal

Electrodermal activity (sympathetic sudomotor drive, sweating) and skin temperature (peripheral perfusion).

Data & protocol

Three protocols, mostly the same people

PhysioNet Wearable Device Dataset from Induced Stress and Structured Exercise Sessions (v1.0.1; Hongn et al., Scientific Data 2025). 30 participants completed all three protocols, so most comparisons are within the same person. Every one of the 707 files was verified against the official checksums.

Acute stress

Seated. Baseline → Stroop (v1) → Trier Mental Challenge Test (timed arithmetic with aversive noise) → rest → opinion, opposite opinion and serial-subtraction tasks.

Aerobic

Incremental stationary cycling, 60 → 110 rpm with rising resistance (individual sessions), or a 70–95 rpm group spinning class.

Anaerobic

Wingate-style: 3–4 all-out sprints of 30–45 s, each followed by ~4 min of unloaded pedalling.

Raw signals of one participant during the stress protocol
One participant, stress protocol. HR and EDA rise during the tasks (orange) and settle during rest, while the wrist barely moves.
What Phase 1 uncovered. 25 data issues, 15 of them not listed by the authors: skin-temperature fault codes (207–235 °C), a stuck accelerometer axis, a session without usable PPG, tags out of order, and group-session timing differences. Also: no exercise session contains a true resting baseline. The only seated reference is on the stress day.
Signal quality first

Before any comparison: which windows can be trusted?

Empatica's heart rate was often physiologically implausible during cycling: flat, erratic, or locked onto the pedalling rhythm. Every 30-s window got a quality verdict per signal. The rules use signal properties only: PPG spectral quality, agreement between independent HR estimates, and cadence lock-in detected with the accelerometer. They never use the expected result.

Clean and corrupted PPG examples with spectra
Same participant, three moments. Seated, the PPG has one clean spectral peak at the heart rate. During a sprint the optical signal is swamped by grip and forearm forces.
Share of valid HR windows by state
Usable HR: ~60–67% of seated windows, 43% of aerobic stages, 3% of sprints.
Aerobic HR staircase before and after quality filtering
Independent check: quality filtering, blind to the expected result, sharpens the incremental HR staircase (median ρ 0.87 → 0.95).
Physiology

Stress and exercise both activate the body, in very different ways

All numbers are participant-level: windows are averaged within each person first, and the differences are paired within person, with bootstrap 95% CIs.

+5.0 bpm

Heart-rate response to the mental-stress task (95% CI 2.2–7.8, same day, n = 34). HR then drops below the starting baseline during recovery, which suggests the baseline itself held anticipatory arousal.

−0.006 g

Wrist movement decreases during stress. The physiological response happens with the body stiller than at rest, so it cannot be a movement artefact.

+49 bpm

At the aerobic peak, about 10× the stress response. EDA is ~11× its resting level, against ~1.6× under stress.

~60 s

After a sprint, HR keeps climbing. The cardiovascular response lags the mechanical effort, and valid HR from the wrist only returns about a minute after the sprint.

Event-locked responses to the stress task and to sprints
Event-locked responses (mean ± 95% CI across participants). Top: the stress task raises HR and EDA with no change in movement. Bottom: a sprint is a movement burst, followed by lagging HR and a sustained EDA rise.
Change from seated rest for each signal and state
Change from seated rest for every signal and state. Skin-temperature slope shows no reliable change anywhere. Exercise contrasts compare different days.
Aerobic stage-by-stage profiles
Incremental cycling: HR and movement rise steadily. EDA stays flat until ~95 rpm, then climbs steeply, consistent with a sweating threshold. Skin temperature does not rise.
Individual stress responses
Individual responses vary several-fold: 25/34 participants raise HR and 29/36 raise EDA, and the two responses are unrelated within a person (ρ ≈ 0).
Multimodal physiological profile heatmap
Multimodal profile. No single signal separates all states; their combination does. A heatmap is used instead of a radar chart because these axes have no meaningful order or area.
Intraclass correlation by feature and state
Physiology is individual: in rest, stress and aerobic states, 83–99% of HR and EDA window-to-window variance is between people. Absolute thresholds generalise poorly, and models must be validated on unseen people.
Self-report vs physiology. Within each person, blocks rated as more stressful also show higher HR (rrm = 0.40) and EDA (rrm = 0.18). The wearable tracks perceived stress, but only moderately. RMSSD fell by ~7 ms during the stress task (vagal withdrawal), but only 17–23 participants had enough clean beats, so this evidence is suggestive.
Machine learning, last

Can the three states be told apart in someone the model has never seen?

Classes: stress tasks, aerobic stages, and sprints plus their first 2 min of recovery. Validation: leave-one-subject-out. Metrics: macro-F1 and balanced accuracy, with CIs from resampling participants.

Resultmacro-F1What it means
Random window split (leaky)0.83Same people in train and test: the model recognises individuals
Leave-one-subject-out, all signals0.66 (0.61–0.71)Honest estimate for a new person
Movement only0.58Best single modality: seated vs pedalling
Physiology only (HR + EDA + temperature)0.50Above chance (0.33), weaker than movement
Movement + "HR missing" flag0.64Nearly all the gain: the PPG failing during sprints is itself a cue
Stress vs seated rest, physiology only0.65 (0.60–0.70)Where movement can't help, physiology (EDA first) does
Modality ablation
Combinations beat single signals. Movement dominates the three-class problem. Skin temperature alone is at chance level.
Confusion matrices
The hard part is aerobic vs anaerobic (both pedalling). Stress is recognised in 83% of windows.
The honest message. This is not "AI detects stress and exercise with high accuracy". A wrist wearable separates seated stress from cycling mostly by movement. It detects acute stress while sitting from physiology with modest accuracy, and it struggles to tell two kinds of cycling apart. Feature importance is reported, but never read as physiological causation.
Method

Seven notebooks, one rule: physiology before modelling

  1. Data understanding. Checksums, sampling rates verified from headers, tag-to-protocol maps, a 25-item issue registry.
  2. Signal quality. Per-window rules for PPG/HR, EDA, temperature and ACC, calibrated on clean seated data.
  3. Protocol segmentation. Cadence fidelity checked with the accelerometer; analysis states comparable across protocol versions.
  4. Exploratory physiology. Event-locked responses, stage profiles, individual trajectories, multimodal profile.
  5. Feature engineering. 14 window features with a physiological rationale; univariate effects at participant level.
  6. Statistics. Paired effect sizes with bootstrap CIs, mixed models, ICCs, repeated-measures correlation, robustness checks.
  7. Classification. LOSO validation, a leakage demonstration, modality ablation, a missingness-shortcut test, permutation importance.
Limitations

What this dataset cannot tell us

Small, confounded sample

36 young, healthy adults. Cohort is confounded with sex, age, training status and protocol version (v1: 18 men; v2: 17 women + 1 man).

No reference measurements

No ECG, power output, VO₂ or lactate. "Aerobic" and "anaerobic" are protocol labels, not verified metabolic states.

Wrist PPG during exercise

HR is unmeasurable during sprints and often corrupted while pedalling. HRV is only defensible while seated.

Different days, short tasks

The resting reference is on another day. The 30-s stress tasks are short relative to EDA and HRV dynamics.

Nothing here supports clinical inference or diagnosis.