Abstract 364: Continuous Electrocardiogram Classification During Resuscitation Using Transfer Learning

• Published in: Circulation (New York, N.Y.) Vol. 142; no. Suppl_4
• Main Authors: Coult, Jason, Sashidhar, Diya, Rea, Thomas, Blackwood, Jennifer E, Kudenchuk, Peter, Kutz, J. Nathan, Kwok, Heemun
• Format: Journal Article
• Language: English
• Published: 17.11.2020
• ISSN: 0009-7322; 1524-4539
• DOI: 10.1161/circ.142.suppl_4.364

Abstract only Introduction: Cardiac arrest resuscitation requires CPR interruption for ECG rhythm analysis, but pausing CPR is adversely associated with survival. Ideally, automated rhythm analysis would occur agnostic of CPR state throughout resuscitation and discriminate non-shockable from shockable rhythms. Transfer learning of pre-trained deep convolutional neural networks (CNNs) may enable accurate ECG analysis when applied to time-frequency representations of the ECG. We designed and evaluated a transfer learning algorithm to identify ventricular fibrillation (VF), asystole (AS), and organized (OR) rhythms agnostic of CPR. Methods: In this observational study of out-of-hospital cardiac arrest, rhythms were manually diagnosed in continuous defibrillator ECG recordings. Non-overlapping adjacent 2-s ECG segments were extracted from the first 30 min of each case regardless of CPR during VF, AS, and OR. Each segment was represented by an intrafrequency-normalized Morlet wavelet transform from 4-40 Hz. Using a 2/3 subset of patients for training, a series of two ResNet-101 CNNs were retrained to perform a shock decision (VF vs. non-shockable) followed by a specific non-shockable prediction (AS, OR, or Indeterminate). Performance was evaluated in a 1/3 validation subset of patients using a range of probability decision thresholds to predict the class of each segment. Results: In total, 275100 segments were collected from 461 patients. Of 90962 segments from 152 validation patients, using a 0.7 probability threshold for class prediction, 21% (18930/90962) were indeterminate, shock vs. no-shock sensitivity and specificity were 90% (19702/21930) and 97% (48421/50102), and specificities among non-shockable rhythms for AS vs. OR were 84% (5032/5998) and 86% (37760/44104), respectively (Fig 1). Conclusion: Transfer learning may enable shock/no-shock and rhythm-specific ECG classification continuously throughout resuscitation regardless of CPR.