Electrocardiogram-based pulse prediction during cardiopulmonary resuscitation.

ABSTRACT

OBJECTIVE: Resuscitation requires CPR interruptions every 2 min to assess rhythm and pulse status. We developed a method to predict real-time pulse status in organized rhythm ECG segments with and without CPR artifact. METHODS: The study cohort included patients who received attempted resuscitation following ventricular fibrillation arrest. Using audio-supplemented defibrillator recordings, we annotated CPR, rhythm, and pulse status at each two-minute rhythm/pulse check. Paired ECG segments with and without CPR were extracted at each rhythm/pulse check. Using one-third of cases for training and two-thirds for validation, we developed three wavelet-based ECG features and combined them with a logistic model to predict pulse status. Predictive performances of each individual ECG feature and the combined logistic model were measured by the area under the receiver operator characteristic curve (AUC) in the validation cases with and without CPR. RESULTS: There were 238 cases and 911 ECG segment pairs. Among 319 organized rhythm segments in the validation set, AUC for pulse prediction during CPR ranged from 0.67 to 0.79 for the individual ECG features. The logistic model was more predictive than any individual feature (AUC 0.84, 95% CI 0.80-0.89, p < 0.05 for each comparison) and performed similarly regardless of CPR (p = 0.2 for difference). CONCLUSION: ECG features extracted by wavelet analysis predicted pulse status with moderate accuracy among organized rhythm segments with and without CPR. Further study is required to understand how real-time pulse prediction during CPR could help direct care while limiting CPR interruption.