A Convolutional Neural Network Approach for Classifying Cardiac Arrest Rhythms Corrupted by Chest Compression Artifacts

Trygve Eftestøl1, Anna Kaczyńska2, Ali Bahrami Rad3, Elisabete Aramendi4, Iraia Isasi4, Andoni Elola5, Lars Wik6, Jo Kramer-Johansen7
1University of Stavanger, 2Lodz University of Technology, 3University of North Dakota, 4UPV/EHU, 5University of the Basque Country, 6Norwegian National Advisory Unit for Prehospital Emergency Care (NAKOS), Oslo University Hospital; Norwegian Air ambulance dept., Division of Prehospital Care, 7Norwegian National Advisory Unit for Prehospital Emergency Care (NAKOS), Oslo University Hospital; Division of Prehospital Care, Institute of Clinical Medicine, Faculty of Medicine, University of OsloNorwegian Air ambulance dept., Division of Prehospital Care

Abstract

Aims: Patients undergoing cardiopulmonary resuscitation (CPR) may exhibit transitions among several ECG rhythm categories and clinical states relevant to resuscitation management, including ventricular fibrillation (VF), ventricular tachycardia (VT), asystole (ASY), pulseless electrical activity (PEA), and pulse-generating rhythm (PGR). Accurate rhythm recognition is vital for effective CPR, but reliability during chest compressions remains highly challenging. We therefore trained deep neural networks to classify the cardiac rhythm on both electrocardiograms (ECG) influenced by chest compression artifacts or not.

Methods: The data set consisted of ECG segments from patients experiencing out-of-hospital cardiac arrest and receiving CPR. Three types of four-second ECG segments were analyzed: 2883 artifact-free segments from 100 patients and 9957 artifact-corrupted segments before and after filtering from 272 patients. An adaptive filtering scheme was applied to remove the artifacts. A series of convolutional neural networks (CNN) were implemented for the five-class classifier with 1-5 layers and variations on a set of hyperparameters in a 10-fold cross-validation scheme. The CNNs were trained with approximately 90% of the datasets, respectively, and tested with separate sets (10%) in a stratified manner. Separate models were trained and tested for each of the three datasets.

Results: A balanced accuracy for the best performing models trained on the clean, corrupted, and filtered datasets were 90%, 62%, and 71%, respectively. The best models had four to five layers.

Conclusion: We proposed a deep learning approach to automatically recognize the five common cardiac arrest rhythms during resuscitation. Chest compression artifacts substantially degrade the performance of ECG-based classification during CPR (by 28 percentage points). While deep learning models trained on corrupted signals exhibit reduced accuracy, preprocessing with adaptive filtering partially mitigates this effect (an increase of 9 percentage points). These findings underscore the importance of artifact suppression for reliable automated analysis in real-time resuscitation settings.