Generative Adversarial Network Data Augmentation Strategy for Wavelet-Based Electrocardiogram Anxiety Detection
Results in Engineering · 2026-11 · 114134
DOI: 10.1016/j.rineng.2026.114134 ↗Kurz gefasst
Fachartikel in der Zeitschrift Results in Engineering, 2026-11. Verfasst von Eli Gabriel Avina-Bravo und 5 weiteren Autor:innen. Die Arbeit nennt 65 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0).
Abstract (Originalsprache)
Anxiety is a prevalent mental health condition with measurable physiological correlates, yet the scarcity of labeled biosignal data remains a critical bottleneck for developing robust automated detection systems, especially those based on machine learning. This study assesses a class specific Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) for electrocardiogram (ECG) data augmentation, applied to binary anxiety classification (low vs. moderate) labelled with a six item short form of the State-Trait Anxiety Inventory (STAI) state scale. ECG signals from 21 subjects were preprocessed with a filter-first pipeline and dynamic wavelet decomposition, then segmented into overlapping windows and sequences for a CNN-LSTM classifier evaluated under subject-independent (grouped) five-fold cross-validation. With generators trained once on all 21 subjects (first analysis), 1:2 augmentation raised mean ROC AUC from 70.9% ± 12.6% to 89.6% ± 7.8% and mean sensitivity from 45.1% to 87.3%. Because the held out subjects contributed to generator training in that design, the pipeline was repeated with both generators retrained inside each fold on the training subjects only. Under this protocol mean ROC AUC was 76.2% ± 9.3% for the baseline and 67.6% ± 32.4% for the augmented classifier (one sided Wilcoxon p = 0.406 ), and sensitivity rose from 36.9% to 59.3% while specificity fell from 81.8% to 69.8%. At recording level the ROC AUC difference was 3.6 points lower for the augmented classifier (95% interval − 20.9 to + 11.4 ), and a leave one subject out analysis gave 2.5 points lower ( − 15.9 to + 12.2 ). Similarity metrics of the synthetic windows were computed in sample and are descriptive. In this cohort, the benefit observed with shared generators was not reproduced when the generators saw only training subjects, which indicates that generators should be trained inside each fold before a gain is attributed to augmentation. The evaluation is limited by cohort size, short self reported labels and one seed.
Zitieren
APA
Avina-Bravo, E. G., Garcia, U. D., Mireles, C., Contreras-Domínguez, D. E., Chairez, I., & Alfaro-Ponce, M. (2026). Generative Adversarial Network Data Augmentation Strategy for Wavelet-Based Electrocardiogram Anxiety Detection. Results in Engineering, 114134. https://doi.org/10.1016/j.rineng.2026.114134
Vancouver
Avina-Bravo EG, Garcia UD, Mireles C, Contreras-Domínguez DE, Chairez I, Alfaro-Ponce M. Generative Adversarial Network Data Augmentation Strategy for Wavelet-Based Electrocardiogram Anxiety Detection. Results in Engineering. 2026:114134. doi:10.1016/j.rineng.2026.114134
BibTeX
@article{AvinaBravo2026,
author = {Avina-Bravo, Eli Gabriel and Garcia, Ulises D. and Mireles, Caridad and Contreras-Domínguez, Diego Ernesto and Chairez, Isaac and Alfaro-Ponce, Mariel},
title = {Generative Adversarial Network Data Augmentation Strategy for Wavelet-Based Electrocardiogram Anxiety Detection},
journal = {Results in Engineering},
year = {2026},
pages = {114134},
doi = {10.1016/j.rineng.2026.114134}
}
EUVIMED