Paperdoi:10.1109/icitri71999.2026.11751310
Balancing Accuracy and Size: Lightweight Stress Detection from Wearable PPG and EDA Signals for Edge Deployment
IEEE ICITRI 2026 (2026-09-01)
TL;DR
Binary stress on EmpaticaE4Stress (783 samples, 29 subjects; 8 handcrafted features: MeanNN, SDNN, RMSSD, pNN50 + 4 EDA), stratified 5-fold CV (sample-level, not subject-wise) with in-fold oversampling. RF 78.7% acc / F1 0.849 at 1,954.72 KB; MLP-Medium 77.5% / F1 0.832 at 3.25 KB (601× smaller). PPG-HRV features carried 66.8% of importance.
- Form factor: wrist (Empatica E4)
Abstract
Most work on wearable stress detection optimizes for classification accuracy alone, leaving aside the model-size constraints that edge deployment on resource-limited hardware actually imposes. We address this by evaluating five machine learning models — Support Vector Machine (SVM), Random Forest, and three MLP variants (Small, Medium, Large) — on binary stress classification using photoplethysmography (PPG) and electrodermal activity (EDA) signals from the EmpaticaE4Stress dataset. Eight handcrafted features were used: four HRV-based features from PPG (MeanNN, SDNN, RMSSD, pNN50) and four EDA-based features (mean, standard deviation, SCR count, slope). We evaluated all models with stratified 5-fold cross-validation on 783 samples (261 rest, 522 stress) from 29 subjects; minority-class oversampling was applied only within each training fold, keeping the held-out test fold free of leakage. Random Forest reached the highest accuracy (78.7%, F1: 0.849) but at a cost of 1,954.72 KB of storage — impractical for a wearable - and a closer look at its per-class recall shows this figure is inflated by a bias toward the majority stress class (56.3% rest-class recall). MLP-Medium, by contrast, reached 77.5% accuracy (F1: 0.832) at just 3.25 KB: a 601-fold reduction in size for a 1.2 percentage-point drop in accuracy relative to Random Forest. The larger MLP-Large variant (11,521 parameters), counter-intuitively, did worse than both smaller MLPs (76.1% accuracy), which points to overfitting given the limited sample size. HRV features from PPG accounted for 66.8% of total predictive importance, with RMSSD and SDNN as the strongest individual predictors. Overall, MLP-Medium offers the best trade-off for edge-deployable workplace stress monitoring.
Verbatim from Semantic Scholar (opens in a new tab) · Abstract via Semantic Scholar; © the authors/publisher
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
Abrori, S. A., & Astuti, W. (2026). Balancing Accuracy and Size: Lightweight Stress Detection from Wearable PPG and EDA Signals for Edge Deployment. IEEE ICITRI 2026. https://doi.org/10.1109/icitri71999.2026.11751310
Abrori, Syauqi Abdurrahman and Winda Astuti. 2026. "Balancing Accuracy and Size: Lightweight Stress Detection from Wearable PPG and EDA Signals for Edge Deployment." IEEE ICITRI 2026. https://doi.org/10.1109/icitri71999.2026.11751310.
Abrori, S. A. and Astuti, W. (2026) 'Balancing Accuracy and Size: Lightweight Stress Detection from Wearable PPG and EDA Signals for Edge Deployment', IEEE ICITRI 2026. doi: 10.1109/icitri71999.2026.11751310.
Abrori, Syauqi Abdurrahman, and Winda Astuti "Balancing Accuracy and Size: Lightweight Stress Detection from Wearable PPG and EDA Signals for Edge Deployment." IEEE ICITRI 2026, 2026. https://doi.org/10.1109/icitri71999.2026.11751310.
S. A. Abrori and W. Astuti, "Balancing Accuracy and Size: Lightweight Stress Detection from Wearable PPG and EDA Signals for Edge Deployment," IEEE ICITRI 2026, 2026, doi: 10.1109/icitri71999.2026.11751310.
@article{abrori2026balancing,
title = {{Balancing Accuracy and Size: Lightweight Stress Detection from Wearable PPG and EDA Signals for Edge Deployment}},
author = {Abrori, Syauqi Abdurrahman and Astuti, Winda},
year = {2026},
journal = {IEEE ICITRI 2026},
doi = {10.1109/icitri71999.2026.11751310}
}Added to the ChatPPG research hub