Published August 29, 2026
| Version 3
Software
Open
Reproducibility artifact: On-device cognitive-load estimation from wrist PPG and accelerometry (cross-dataset study and consumer-smartwatch deployment)
Authors/Creators
- 1. TNU - University of Information and Communication Technology, Thai Nguyen, Vietnam
Description
Source code and aggregated statistical results supporting a manuscript on wrist-worn (Empatica E4 / Google Pixel Watch 2) cognitive-load estimation from PPG-derived HRV and accelerometry, evaluated on WESAD (stress-proxy pilot) and WAUC (genuine mental-workload labels).
Version 1 covered the base pipeline: LOSO handcrafted-feature (SVM/MLP) and raw-signal 1D-CNN baselines, test-time personalization via T3A, bootstrap-CI and subject-structure-preserving permutation testing, per-fold session-structure confound residualization, a nonparametric within-subject matched-session control, and the on-device calibration comparison evidencing a cross-device sensor-unit-conversion fix for a real consumer-smartwatch deployment.
Version 2 added the audit and reviewer-response layer (experiments 08-14): a subgroup performance-disparity and T3A personalization-fairness audit across sex / activity / age / BMI axes with Benjamini-Hochberg correction (plus an intersectional sex x activity variant); a session-block permutation test and a nested sensor-vs-metadata comparison; per-subject ROC-AUC and a Monte-Carlo power curve for the subgroup gap; physical-workload-stratified and per-CNN matched-session tests; a 44-subject ACC-only sensitivity analysis that reconstructs ACC features without the PPG signal-quality gate and re-runs the unchanged LOSO pipeline; and an off-device Kotlin-vs-NumPy numerical-parity check for the deployed model. All Version 2 additions are aggregated result tables only.
Version 3 makes no analysis or result changes: it removes the earlier double-blind-review anonymization of the author-facing text (README wording and the Android package name in the experiments/13 parity check).
Raw per-subject physiological recordings, per-window feature tables, per-fold/per-candidate training artifacts and model checkpoints, and the manuscript itself are excluded; this is a curated subset of a larger research repository. Code released under MIT, result files under CC BY 4.0.