ECG Stress Detection Pipeline

What this research found

An end-to-end pipeline for detecting physiological stress from electrocardiogram (ECG) signals: bandpass filtering, R-peak detection, eight heart rate variability (HRV) features over 60-second windows, non-parametric hypothesis testing, and four classifiers compared under leave-one-subject-out validation. Because the WESAD benchmark requires manual download and authentication, the pipeline was exercised on synthetic ECG generated to match WESAD's chest-sensor specification — three simulated subjects at 700 Hz — rather than on the real recordings. On that synthetic cohort, Random Forest reached 98.81% accuracy and a single feature, pNN50, carried roughly 78% of the model's predictive weight.

• The numbers come from synthetic signals, not real recordings. With WESAD gated behind manual download and authentication, a generator produced three simulated subjects of 30 minutes each at 700 Hz, with realistic P-QRS-T complexes plus baseline wander and high-frequency noise added to mimic recording conditions.
• One feature does nearly all the work. Across validation folds, pNN50 — the share of successive heartbeat intervals differing by more than 50 milliseconds — carried 0.7801 plus or minus 0.0418 of the gradient-boosted model's feature importance, with mean interbeat interval at 0.1029 and RMSSD at 0.0497, leaving the frequency-domain features near zero.
• All eight HRV features separated the two conditions after Bonferroni correction across 84 baseline and 84 stress windows, six of them at p < 0.001. pNN50 showed the largest effect (75.88% down to 52.88%, Cohen's d = 3.33), followed by mean interbeat interval (624.09 to 520.54 milliseconds, d = 1.52), sample entropy (1.55 to 1.23, d = 0.69), and the LF/HF ratio (1.53 to 0.61, d = 0.51).
• Model choice barely mattered. Under leave-one-subject-out validation, Random Forest reached 98.81% plus or minus 0.84% accuracy with 2 windows misclassified, SVM 98.21% with 3, logistic regression 97.62% with 4, and XGBoost 97.02% with 5. The linear model landing within about a percentage point suggests the classes are largely linearly separable in the eight-feature space.
• Two features moved against textbook expectation: high-frequency power rose under stress (23,691 to 35,446, d = -0.25), the only feature to increase, and the LF/HF ratio fell rather than rising as conventional sympathovagal interpretation would predict.

How it was done

ECG sampled at 700 Hz was bandpass filtered between 0.5 and 40 Hz with a 4th-order zero-phase Butterworth filter, improving signal-to-noise ratio by 10.81 decibels. R-peaks were detected with a Pan-Tompkins implementation, and interbeat intervals were cut into 60-second windows stepped every 30 seconds and never crossing a condition boundary, giving 168 windows split evenly between baseline and stress. Eight features were computed per window: four time-domain measures (mean interbeat interval, SDNN, RMSSD, pNN50), three frequency-domain measures from Welch power spectral density, and sample entropy. Shapiro-Wilk tests indicated non-normal distributions, so comparisons used Mann-Whitney U tests with Bonferroni correction over the eight features plus Cohen's d for effect size. Random Forest, SVM, logistic regression, and XGBoost were then compared under leave-one-subject-out cross-validation, with standardisation applied inside the pipeline for the two scale-sensitive models.

Data sources

• Synthetic ECG generated to the WESAD chest-sensor specification — 3 simulated subjects, 700 Hz, 30 minutes each, 168 sixty-second windows
• WESAD (Wearable Stress and Affect Detection), UCI Machine Learning Repository — the intended source, not retrieved during the run
• Schmidt et al., ICMI (2018) — WESAD dataset description
• Task Force of the European Society of Cardiology (1996) — heart rate variability measurement standards