Year: 2026
Type: preprint
Abstract: Wrist research and consumer pipelines often treat elevated heart rate or skin conductance as evidence of psychological strain, a practice documented across ambulatory stress surveys and laboratory affect studies. That habit confuses constructs psychophysiology has long kept apart: seated cognitive/psychosocial demand versus physical exertion. This paper asks what a measurement model must clear if the claim is specifically cognitive. We audit a public PhysioNet induced-stress and structured-exercise feature matrix (100 sessions; 41 participants; STRESS, AEROBIC, ANAEROBIC), complement it with WESAD and electrodermal/physiological-computing standards, and define three quantities with explicit pass criteria: a temporal eligibility ratio R_t/s, an exertion-control effect-size ratio R_d on named features with 95% confidence intervals, and an exploratory person-wise leave-one-participant-out (LOPO) check. Under a frozen Random Forest and keyword contract, temporal-keyword features hold 22.9% of top-30 importance (R_t/s=0.30). For heart-rate trend slope, stress-versus-aerobic d=-1.53 (95% CI [-2.08,-0.99]) while aerobic-versus-anaerobic d=-0.26 (95% CI [-0.76,0.23]), so point R_d=5.80 (session bootstrap 95% CI [1.75,106], very wide). Pooled LOPO accuracy is 0.76 (participant-bootstrap 95% CI [0.67,0.84]) versus stratified session 5-fold 0.76 (SD 0.07). We drop cross-corpus mean-EDA "sign flip" comparisons from the gate because WESAD stress-versus-rest and PhysioNet stress-versus-exercise are not matched contrasts. A cognitive wearable claim that cannot meet the stated criteria remains an arousal claim.
Authors: Eleanor Finch, Shreyanka Pokhrel, Tom Hughes
Language: English
Topic: Emotion and Mood Recognition
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7235990252
