Abstract P2001: Generative Models for Synthesizing Cardiac Arrest Electrocardiograms With and Without Chest Compression Artifact
Generative Modelle zur Synthese von Herzstillstands-EKGs mit und ohne CPR-Artefakt
Circulation · 02.11.2026 · 154 (Suppl_1)
DOI: 10.1161/circ.154.suppl_1.p2001 ↗Auf einen Blick
Generative KI-Modelle erzeugten OHCA-EKGs mit und ohne CPR-Artefakt. Mit synthetischen Daten trainierte Klassifikatoren erreichten eine gute bis moderate Leistung, lagen aber etwas unter Modellen mit echten Daten.
Fragestellung
Können mit synthetischen OHCA-EKGs trainierte Klassifikatoren reale EKGs während und ohne CPR zuverlässig klassifizieren, und wie schneiden sie im Vergleich zu mit echten Daten trainierten Modellen ab?
Methode
Retrospektive Studie zu erwachsenen VF-OHCA von 2006-2021 in einem metropolitanen EMS-System; 80 %/20 % Randomisierung für Training/Test; GAN und Diffusion Models zur EKG-Synthese; Vergleich von 1-D- und 2-D-CNNs.
Ergebnisse
2011 Patient:innen mit 1,03 Millionen EKGs; 60 % mit CPR. Im Testdatensatz 213664 EKGs von 402 Patient:innen. Synthetisch trainierte Modelle: AUC 0,95-0,99, F1 0,78-0,87; 3-Klassen-BAC 62-85 %. Echte Daten waren im Mittel etwas besser.
Schlussfolgerung der Autor:innen
Generative Modelle konnten EKGs mit und ohne CPR erfolgreich synthetisieren. Mit synthetischen Daten trainierte Klassifikatoren zeigten etwas geringere Leistung als mit echten Daten trainierte Modelle.
Kernaussagen
- 2011 Patient:innen; 1,03 Millionen EKGs; 60 % mit CPR
- Synthetische Trainingsdaten lieferten gute binäre Klassifikation
- Leistung mit echten Daten im Mittel etwas höher als mit synthetischen
Deutsche Zusammenfassung: KI-gestützt aus dem Original-Abstract erstellt; alle Zahlen wurden automatisch mit dem Original abgeglichen. Maßgeblich ist der Original-Abstract.
Kurz gefasst
Fachartikel in der Zeitschrift Circulation, erschienen am 02.11.2026. Verfasst von Xinyuan Zhang und 9 weiteren Autor:innen.
Themen: Reanimation
Abstract (Originalsprache)
Background
ECG rhythm assessment during resuscitation of ventricular fibrillation/tachycardia (VF/VT) out-of-hospital cardiac arrest (OHCA) requires CPR interruption because CPR causes ECG artifacts. Ideally defibrillators would use artificial intelligence (AI) for binary (non-shockable vs. shockable) and 3-class (asystole [ASY] vs. organized [ORG] vs. shockable [VF/VT]) ECG classification during CPR to guide diagnosis and treatment. However, algorithm development is challenged by a paucity of OHCA ECGs, particularly during CPR. Goal: We built generative AI models to synthesize OHCA ECGs with and without CPR artifact to (1) determine whether classifiers trained with synthetic (artificial) ECGs can classify real (clinical) ECGs, and (2) compare performance between classifiers trained with real vs. synthetic ECGs.
Methods
We performed a retrospective study of adult VF-OHCA from 2006-2021 in a metropolitan EMS system. Patients were randomized 80%/20% for training/test. Defibrillator pad ECGs clipped into 2.5-s segments were manually labeled by CPR and rhythm (ASY, ORG, VF, VT) by expert reviewers. Using training ECGs, we first developed generative adversarial networks (GAN) and diffusion models (DM) to synthesize ASY, ORG, and VF/VT ECGs with and without CPR. Then using (1) the real training ECGs, (2) GAN-generated ECGs, and (3) DM-generated ECGs, classifiers were trained for binary (ASY/ORG vs. VF/VT) and 3-class (ASY vs. ORG vs. VF/VT) rhythm prediction. Classifiers included 1-D convolutional neural networks (CNNs) with ECG voltages as inputs and 2-D CNNs with ECG scalograms as inputs. To evaluate synthetic data utility, we compared classifiers trained with real vs. synthetic data in the test dataset using binary (area under the receiver operating characteristic curve [AUC], F1 score) and 3-class (balanced accuracy [BAC]) metrics.
Results
The cohort had 2011 patients with 1.03 million ECGs (60% with CPR). Classifiers were trained on real, GAN-generated, and DM-generated datasets of 816134 ECGs each (Figure). In the test dataset of 213664 ECGs from 402 patients, classifiers trained with synthetic data had good binary performance without CPR (AUC=0.97-0.99, F1=0.78-0.87) and with CPR (AUC=0.95-0.96, F1=0.84-0.86) (Table 1), and moderate 3-class performance without CPR (BAC=65-85%) and with CPR (BAC=62-84%) (Table 2). On average, classifiers trained with real data outperformed those trained with synthetic data by an AUC of 0.019 (vs. GAN) and 0.016 (vs. DM), F1 of 0.084 (vs. GAN) and 0.081 (vs. DM), and BAC of 4.4% (vs. GAN) and 6.1% (vs. DM).
Conclusions
Generative models successfully synthesized ECGs with and without CPR. Classifiers trained on synthetic data had slightly reduced performance versus those trained on real data. The results suggest potential for on-demand synthesis of OHCA ECGs, enabling AI sandboxing and development of next-generation ECG algorithms.
Zitieren
APA
Zhang, X., Zeng, A., Tsang, E., Steffen, B., King, J. A., Boyle, P. M., Johnson, N. J., Kudenchuk, P. J., Rea, T., & Coult, J. (2026). Abstract P2001: Generative Models for Synthesizing Cardiac Arrest Electrocardiograms With and Without Chest Compression Artifact. Circulation, 154(Suppl_1). https://doi.org/10.1161/circ.154.suppl_1.p2001
Vancouver
Zhang X, Zeng A, Tsang E, Steffen B, King JA, Boyle PM, et al. Abstract P2001: Generative Models for Synthesizing Cardiac Arrest Electrocardiograms With and Without Chest Compression Artifact. Circulation. 2026;154(Suppl_1). doi:10.1161/circ.154.suppl_1.p2001
BibTeX
@article{Zhang2026,
author = {Zhang, Xinyuan and Zeng, Alan and Tsang, Edric and Steffen, Brandon and King, Julia A and Boyle, Patrick M and Johnson, Nicholas J and Kudenchuk, Peter J and Rea, Thomas and Coult, Jason},
title = {Abstract P2001: Generative Models for Synthesizing Cardiac Arrest Electrocardiograms With and Without Chest Compression Artifact},
journal = {Circulation},
year = {2026},
volume = {154},
number = {Suppl_1},
doi = {10.1161/circ.154.suppl_1.p2001}
}
EUVIMED