Abstract 308: Automated Detection of Patients with Return of Spontaneous Circulation in the Retrospective Analysis of Resuscitation Episodes

Abstract

Introduction: Retrospective analysis of out-of-hospital cardiac arrest episodes is important to evaluate the performance of resuscitation teams, and to advance research. Large numbers of electronic files are compiled in cardiac arrest registries and identifying patients with return of spontaneous circulation (ROSC) is expensive and time-consuming. The aim of this study was to analyze the feasibility of automatically annotating ROSC using the signals recorded by defibrillators. Materials and methods: A set of 893 patients (261 with ROSC, of which 127 were transient) recorded by the Dallas-Fort Worth Center for Resuscitation Research using Philips HeartStart MRx devices were included. Every record contained the ECG and the thoracic impedance (TI) signals. ROSC was automatically identified as follows: 1) chest compression pauses longer than 60 s were identified using either the TI or the compression depth signals. 2) Organized ECG rhythms were identified using a comercial defibrillator's algorithm. 3) Pulsatile rhythms were identified using a published machine learning algorithm based on the ECG or the ECG and TI signals. The model uses up to 9 features, extracted from both the ECG (6) and the TI (3) signals. ROSC was identified when a 10s-interval was classified as pulsatile and the ROSC onset was set at the beginning of the pause. The performance was assessed in terms of Sensitivity (Se, ROSC patients), Specificity (Sp, no ROSC patients) and overall F1-score. Error in the ROSC onset was calculated comparing the automatically detected onset to the reviewed time of the clinical annotation of ROSC. Results: The algorithm based exclusively on ECG showed Se, Sp and F1 scores of 90.0%, 92.7% and 90.5%, respectively. The algorithm that included both ECG and TI improved scores to 91.2%, 94.3% and 92.1%. For the set of patients with transient ROSC the algorithm showed a Se of 81.9% (95.6% for permanent ROSC) using the ECG and TI signals. The median (IQR) absolute error on the time of first ROSC onset was 0.8 (0.6-8.2) s. Conclusions: An accurate algorithm to automatically identify patients with ROSC is demonstrated. This technique could be useful for retrospective analysis of cardiac arrest registries.