Transfer learning based deep network for signal restoration and rhythm analysis during cardiopulmonary resuscitation using only the ECG waveform.

Minimizing the interruption of cardiopulmonary resuscitation (CPR) is an important technique to improve the survival of out-of-hospital cardiac arrest (OHCA) patients. Recent studies have adopted deep convolutional neural networks (CNNs) to analyze corrupted ECGs during CPR for shockable rhythm detection, however, none of them provided restored ECG signals using only the ECG waveform. In this study, a self-supervised UNet deep learning network is designed to restore the underlying ECG signals during chest compressions without additional reference signals. The UNet is pretrained using a large dataset created by combining 12-lead clean ECG signals with simulated CPR artifacts generated by a mathematical model. Transfer learning is applied to fine-tune the UNet parameters based on combined ECG signals from OHCA patients. Finally, an independent dataset extracted from OHCA patients is used to test the model. Compared with the artifact corrupted ECG signals, the overall signal-to-noise ratio (SNR) is improved from −5.3 (-11.2, 0.8) dB to 1.9 (0.7, 3.9) dB after the restoration, and the accuracy of the rhythm detection is improved from 65.8% to 90.8%. The proposed method not only effectively restores the underlying ECG signals but also provides accurate shock advice without interrupting CPR using only the ECG. [ABSTRACT FROM AUTHOR]