Predicting defibrillation outcomes by combining ventricular fibrillation and defibrillation waveforms: a retrospective clinical study

AIM: To validate whether combining ventricular fibrillation (VF) and defibrillation (DF) waveforms could improve the prediction accuracy of DF outcomes in a retrospective cardiac arrest cohort. METHODS: Electrocardiographic waveforms were recorded via defibrillators for patients who experienced VF and DF. DF waveforms were modelled on the basis of reported energy and transthoracic impedance and assessed by related errors between modelled and delivered waveforms. The uncorrupted preshock VF waveform and the modelled DF waveform were combined using a convolutional neural network. The data were randomized into training and testing sets at a ratio of 4:1. The termination of ventricular fibrillation (TOVF), return of organized rhythm (ROOR), and return of potentially perfusing rhythm (RPPR) after each shock were used as DF outcomes. The performance was evaluated by comparing the area under the receiver operating characteristic curve (AUC) with that of the amplitude spectrum area (AMSA). RESULTS: Related errors for the modelled DF waveform ranged from -1.10% to 1.31%. Compared with those of AMSA, AUC values were significantly greater for TOVF (0.627 vs. 0.578; p=0.010), ROOR (0.854 vs. 0.804; p<0.001), and RPPR (0.873 vs. 0.836; p=0.004) when VF and DF waveforms were combined. The most notable improvement occurred in shocks with AMSA values ranging from 4.6-15.0 mVHz, which demonstrated AUC increases of 9.0, 7.3, and 5.1 points for ToVF, ROOR, and RPPR, respectively. CONCLUSIONS: The combination of VF and DF waveforms using a deep learning-based approach significantly improved the prediction accuracy of DF outcomes regardless of the criteria used to define DF success.