Wearable stress-detection models trained to distinguish rest from stress labelled 82.6% of held-out exercise-session windows as stress, demonstrating that exercise-associated activation can produce substantial false-stress responses even when challenge data are excluded from model development.
Key Findings
Results
The best-performing classifier (XGBoost) achieved moderate rest-versus-stress discrimination but misclassified the large majority of exercise-session windows as stress.
XGBoost achieved a balanced accuracy of 0.703, stress detection (sensitivity) of 0.604, and rest specificity of 0.801 on the rest-vs-stress task.
Despite reasonable rest-vs-stress performance, 82.6% of held-out exercise-session windows were labelled as stress.
Exercise-session data were entirely excluded from model development, including preprocessing, hyperparameter selection, and probability-threshold calibration.
Evaluation used nested leave-one-participant-out cross-validation across 29 participants.
Results
The participant-level exercise-session false-stress rate was substantially higher than the rest false-stress rate.
The exercise-session false-stress rate was 0.840 [95% CI: 0.775, 0.897] at the participant level.
The rest false-stress rate was 0.221 [95% CI: 0.151, 0.301].
The paired difference between exercise-session and rest false-stress rates was 0.619 [95% CI: 0.537, 0.699].
This comparison was made within the same dataset and the same model, isolating the effect of physiological activation type.
Results
Protocol-defined stress blocks were associated with higher self-reported stress than rest blocks, partially validating the stress-induction manipulation.
The mean difference in self-reported stress between stress and rest blocks was 1.22 [95% CI: 0.89, 1.53] on a 1–10 scale.
27 out of 29 participants reported higher stress during stress blocks than rest blocks.
Statistical comparison used a one-sided Wilcoxon test.
This finding confirms that the protocol-defined stress blocks did produce subjectively experienced stress above resting levels.
Results
Adding time-domain heart-rate-variability features did not materially change the false-stress rate for exercise sessions.
Inclusion of time-domain HRV features was tested as a potential way to improve stress specificity.
The high exercise-session false-stress rate persisted regardless of whether HRV features were included.
This suggests the confounding by physiological arousal during exercise is robust to common feature-engineering choices.
Results
A cross-dataset transfer analysis from a wearable stress and affect detection dataset to PhysioNet performed poorly, providing secondary evidence of the generalization problem.
The authors performed a wearable stress and affect detection-to-PhysioNet transfer analysis.
Transfer performance was described as 'poor.'
The authors treated this as secondary evidence only, because it also contains dataset shift and protocol shift confounds in addition to the arousal-specificity issue.
The primary evidence for exercise confounding came from the matched within-dataset experiment.
Methods
The study design used a matched within-dataset stress-specificity experiment as the primary test, distinguishing it from typical rest-vs-stress validation approaches.
PhysioNet event tags were used to reconstruct protocol-defined rest and stress-induction blocks.
Participants had matched aerobic and anaerobic exercise recordings.
The authors explicitly framed the rest-vs-stress separation as insufficient to establish specificity to stress rather than broader physiological activation.
Background
Wearable stress-detection models are commonly validated by separating rest from stress, which does not establish that model outputs are specific to stress rather than broader physiological activation.
The authors identify this as a systematic gap in validation methodology for wearable stress detection.
The study was motivated by the observation that physiological signals during exercise and during stress may be similar enough to confuse classifiers trained only on rest vs. stress.
Four feature-based classifiers were tested under this critique: the best was XGBoost.
The study re-analysed two public wearable datasets to test this hypothesis.
What This Means
This research suggests that wearable devices marketed for stress detection may not actually be measuring stress specifically — they may instead be detecting any kind of physical or mental activation that elevates heart rate and other physiological signals. The researchers tested four machine-learning models trained on data from 29 people during rest and stress tasks, then checked what those models said when the same people exercised. Even though exercise data were never used during model training, the best model (XGBoost) incorrectly flagged exercise as 'stress' 82.6% of the time. By contrast, it only misclassified ordinary rest as stress about 22% of the time. The gap between these two error rates — roughly 62 percentage points — shows that exercise-related physical arousal looks a lot like stress to these algorithms.
Importantly, the stress-induction tasks used in the study did produce genuine subjective stress: 27 out of 29 participants reported feeling more stressed during those tasks than at rest, confirming the experiment was working as intended. The problem is not that the stress tasks failed, but that exercise activates the body in ways that are physiologically similar enough to stress that classifiers cannot tell them apart. Adding additional heart-rate-variability features did not fix this problem.
This matters because millions of people use consumer wearables that claim to track stress throughout the day, including during and after workouts. This research suggests those readings during physical activity may be largely meaningless as stress indicators, and that the standard way researchers have been testing these models — comparing rest to stress — does not catch this flaw. Better validation methods that include exercise challenges are needed before wearable stress scores can be considered reliable.
Check Your Own Numbers
Upload your bloodwork. We'll cross-reference your results against this study and 4,700 others.
Aydoğan Y, Povina F. (2026). Stress or arousal? Exercise confounding in wearable stress detection.. Medical engineering & physics. https://doi.org/10.1088/1873-4030/aea41e