Paperdoi:10.21954/ou.ro.00112554
Personalised vs. Population-Level Baselines for Stress Detection in Young People: A SAGEXP Feasibility Study
Open Research Online (The Open University) repository item, 2026-11-16
TL;DR
Feasibility study: a population Random Forest on five per-window WESAD wrist features scored 79.7% accuracy and 83.0% F1 under leave-one-subject-out (15 subjects; stress recall 0.69; per-subject accuracy 54.4–98.3%). Retraining on each person's calibration windows improved or matched every subject, most for those the population model served worst. An int8 dense net was profiled for an ESP32-S3 wrist prototype.
Abstract
Wearable stress detection holds significant promise for supporting young people. However, traditional tools rely on users tracking their own feelings, which can be difficult before stress becomes a major problem. This study forms part of the SAGEXP (Stress Awareness and Generative Explanations) research project and investigates whether a stress classifier calibrated to an individual's physiological baseline outperforms a generalised population-trained model.
A wrist-worn prototype was designed around an ESP32-S3 with photoplethysmography, electrodermal activity, skin temperature, and accelerometer sensing. A population-level Random Forest was trained on the wrist channels of WESAD using five per-window features and evaluated with leave-one-subject-out cross-validation. A small int8-quantised dense network was trained in Edge Impulse and profiled for on-device use. Personalisation was tested by retraining the Random Forest on each subject's calibration windows and evaluating it on a time-based held-out split from the same recording. The population model achieved 79.7% mean accuracy and 83.0% mean F1 across 15 subjects. Stress recall was lower at 0.69, and per-subject accuracy ranged from 54.4% to 98.3%. Personalisation improved or matched performance for every subject, with the largest gains observed among those least well served by the population model. A separate comparison found that class weighting left overall accuracy almost unchanged (76.3% vs. 77.0%) while increasing stress recall from 38.5% to 86.5%.
Although the personalised models were evaluated offline using historical WESAD data, the results provide a strong proof of concept. Personalised, resource-constrained TinyML models offer a viable pathway to reduce the high inter-individual performance variation typical of wearable stress detection.
Verbatim from DataCite (opens in a new tab) · Licence: https://creativecommons.org/licenses/by-nc-nd/4.0/legalcode
Results on ChatPPG leaderboards
All leaderboardsNo results from this paper on ChatPPG leaderboards yet. Browse the leaderboards.
Tasks & signals
Tasks
Signals
Devices
Keywords
Datasets used
How it works
Cite this paper
Phillips, T. (2026). Personalised vs. Population-Level Baselines for Stress Detection in Young People: A SAGEXP Feasibility Study. Open Research Online repository item. https://doi.org/10.21954/ou.ro.00112554
Phillips, Temiloluwa. 2026. "Personalised vs. Population-Level Baselines for Stress Detection in Young People: A SAGEXP Feasibility Study." Open Research Online repository item. https://doi.org/10.21954/ou.ro.00112554.
Phillips, T. (2026) 'Personalised vs. Population-Level Baselines for Stress Detection in Young People: A SAGEXP Feasibility Study', Open Research Online repository item. doi: 10.21954/ou.ro.00112554.
Phillips, Temiloluwa "Personalised vs. Population-Level Baselines for Stress Detection in Young People: A SAGEXP Feasibility Study." Open Research Online repository item, 2026. https://doi.org/10.21954/ou.ro.00112554.
T. Phillips, "Personalised vs. Population-Level Baselines for Stress Detection in Young People: A SAGEXP Feasibility Study," Open Research Online repository item, 2026, doi: 10.21954/ou.ro.00112554.
@article{phillips2026personalised,
title = {{Personalised vs. Population-Level Baselines for Stress Detection in Young People: A SAGEXP Feasibility Study}},
author = {Phillips, Temiloluwa},
year = {2026},
journal = {Open Research Online repository item},
doi = {10.21954/ou.ro.00112554}
}Added to the ChatPPG research hub