Key result
Brief subject-specific ECG calibration achieves ~92% stress classification accuracy, matching fully personalized models.
Why the study?
ECG-derived HRV provides non-invasive stress monitoring for safety-critical human-machine interaction, yet generic population-level classifiers generalise poorly to unseen individuals due to substantial inter-subject physiological variability.
Does a balanced stratified subject-specific calibration strategy improve ECG/HRV stress detection accuracy and calibration quality compared to generic and fully person-specific models?
Does a balanced stratified subject-specific calibration strategy improve ECG/HRV stress detection accuracy and calibration quality compared to generic and fully person-specific models?
Approximately 9 minutes each of baseline and stress ECG recording per individual is sufficient to calibrate generic machine learning models to match fully person-specific stress detection performance.
Answers are grounded in this paper and cite their sources.
Per-subject MLP calibration yields high stress-detection accuracy in small samples; hypothesis-generating and requires prospective validation before clinical use.
Reliable stress detection is essential in safety-critical human-machine interaction (HMI) contexts, where stress-induced impairments can lead to errors and system failures. Electrocar-diogram (ECG)-derived heart rate variability (HRV) provides a non-invasive monitoring modality, yet generic population-level classifiers often generalise poorly to unseen individuals because of substantial inter-subject physiological variability. This study proposes a balanced stratified subject-specific calibration strategy that augments a generic Leave-One-Subject-Out (LOSO) model with equal numbers of individual baseline and stress windows, while preventing temporal leakage from overlapping sliding windows. Three modelling strategies were evaluated using support vector machine, random forest, and multilayer perceptron (MLP) classifiers with class-weighted training to address the inherent 2:1 class imbalance in the dataset. Full per-subject calibration-curve analysis identified n = 30 as a practical operating point, achieving 92.11% mean MLP accuracy within 1.1% of the population peak at n = 32. Calibration improved both discriminative performance and probabilistic calibration quality relative to the generic baseline, and the calibrated models matched fully person-specific performance across all classifiers (all p > 0.47, d < 0.26). These findings suggest that approximately 9 minutes each of baseline and stress ECG recording per individual may be sufficient to support practical subject-specific stress monitoring using classical machine learning and single-modality ECG sensing.
No takes yet. Share an insight, caveat, or question.
Jui et al. (2026) studied Physiological stress. Balanced stratified subject-specific calibration strategy for ECG/HRV stress detection vs. Generic Leave-One-Subject-Out (LOSO) baseline and fully person-specific models was evaluated on Mean multilayer perceptron (MLP) stress classification accuracy (d < 0.26, p=> 0.47). Subject-specific ECG/HRV calibration (n=30 windows) achieved 92.11% mean MLP accuracy and matched fully person-specific performance across classifiers (all p>0.47, d<0.26).
Synapse has enriched 10 related papers. Consider them for comparative context: