Submitted:
18 November 2026
Posted:
19 November 2026
You are already at the latest version
Abstract
Wrist-worn stress and affect classifiers are usually developed on laboratory recordings with fixed condition schedules, so a window's position can reveal its condition even for held-out participants. We hypothesize there may be a window position effect. Using public datasets, we quantified protocol information with a position-only classifier on the Wearable Stress and Affect Detection (WESAD) dataset and three further scheduled datasets (N=104) to test this effect. Results showed that, on WESAD, position alone reached a four-class macro-averaged F1 score of 70.9%, against 52.2% for six wrist pulse-variability and skin-conductance features using that classifier, and elapsed time alone exceeded the published wrist results of the WESAD benchmark. Position also outperformed physiological features on the three further datasets. Residualizing features on position could not show whether a classifier uses the schedule, because it penalized feature copies shuffled within conditions as much. Across WESAD's counterbalanced protocol versions, the six features' stress–amusement separation transferred, in a test that detected only a strong added linear drift, whereas the gain from wrist temperature couldn't be separated from the schedule. Participant-wise accuracy therefore doesn't alone show what a wrist-worn classifier relies on; scheduled-protocol evaluations should report a position-only control and, where possible, a counterbalanced comparison.
Keywords:
wearable devices
; window position effect
; stress detection
; participant-wise evaluation
; temporal confounding
; electrodermal activity
; photoplethysmography
; WESAD
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.