Abstract 18877: An Accurate Method for Automated Rhythm Classification During CPR

Abstract

Introduction: Accurate identification of the cardiac rhythm during CPR could facilitate appropriate defibrillation and guide medication treatments without interrupting chest compressions for rhythm analysis. Hypothesis: A rhythm classification method which “reads-through” CPR artifact would identify shockable rhythms with high sensitivity and non-shockable rhythms with high specificity during CPR. Methods: Two physicians annotated rhythms as shockable (VF or other shockable), non-shockable (asystole or organized) or indeterminate, throughout defibrillator recordings from a convenience sample of 185 treated OHCA cases. CPR was determined from impedance and/or accelerometer signals. Each ECG tracing was divided into consecutive 4-second frames, and only frames during CPR were included. The “read-through” method was derived from a development set (N=95 cases) and uses features generated by cross-correlation with a series of QRS-like wavelets and a hidden Markov model to integrate sequential data. In the validation set (N=90 cases), frame-by-frame classifications (shockable vs. non-shockable) were compared to the gold standard of physician assessment. Sensitivity and specificity were estimated using a multilevel model to account for clustering of frames within cases. Results: The validation set included a median 456 (IQR 179, 756) seconds of ECG data from each case. Overall, there were 13,851 4-second frames: 10% shockable, 88% non-shockable and 2% indeterminate. Mean sensitivity for VF was 0.97 (95% CI 0.84, 0.99), and specificities for an organized rhythm and asystole were 0.99 (95% CI 0.99, 1.0) and 0.99 (95% CI 0.97, 1.0), respectively. There was moderate variability in VF sensitivity across cases (IQR 0.77, 1.0). Conclusion: A “read-through” CPR method classified the shockable versus non-shockable rhythms with high overall sensitivity and specificity, though with some variability across individual patients.