A56-03 Temporal ECG Waveform Analysis for Predicting Out-of-hospital Recurrent Cardiac Arrest

Abstract Rationale Recurrent cardiac arrest (rearrest) is common, affecting up to half of patients who achieve return of spontaneous circulation (ROSC) after an out-of-hospital cardiac arrest (OHCA). Advance prediction of rearrest may enable early intervention and treatments with the goal of preventing rearrest. Inpatient studies have showed that heart rate, heart rate variability, and QRS prolongation are predictive of cardiac arrest. However, there are no tools to predict out-of-hospital rearrest. The aim of this study is to evaluate whether time-dependent ECG characteristics predict out-of-hospital rearrest. Methods We performed a retrospective cohort study of adult OHCA treated by Seattle Medic One who achieved ROSC between 2018-2024. We collected a 5-min ECG segment from each patient to evaluate ECG-based prediction of rearrest following ROSC. In the rearrest group, ECGs were collected from 6 to 1 minutes prior to rearrest. In the non-rearrest group, ECGs were collected based on median time to rearrest observed in the rearrest group. We calculated ECG measures across 15-second ECG windows including heart rate, standard deviation of normal-to-normal intervals (SDNN), root mean square of successive differences (RMSSD), proportion of adjacent normal-to-normal intervals differing by 50ms, and mean QRS width and height. Patients were randomized into 80%/20% training/test sets to develop a logistic model to predict rearrest for each 15-second ECG windows to determine the optimal time for analysis. Performance was quantified by area under the receiver operating characteristic curve (AUROC). We also evaluated logistic combinations of temporal trends of ECG features across the 5-min ECG segments (median, slope, and intercept). Results Of the 1,414 OHCA patients who achieved ROSC, 716 (50.6%) experienced rearrest. The median interval from ROSC to rearrest was 6.2 minutes. When evaluating each 15-second window individually, the AUROC’s were between 0.68-0.78 and were not statistically significant between each increment (figure 1). Additionally, incorporating temporal trends did not improve performance (AUROC=0.74 95% CI 0.67-0.82). The top 3 predictors for rearrest were high mean SDNN (OR = 12.9), high SDNN slope (OR = 3.5), and high mean of QRS width (OR = 1.50). Conclusion In this study, we observed that ECG features were predictive of rearrest; however, the addition of temporal analysis did not improve the model performance. Measures of increased heart rate variability and QRS width were the strongest predictors of rearrest. Future directions should evaluate deep learning and multimodal approaches to improve prediction of rearrest. This abstract is funded by: R01HL169323-03, 2T32HL007287-46A1