Abstract 124: A Machine-Learning Based Shock Advice Algorithm for Reliable Rhythm Diagnosis During Out-Of-Hospital Cardiac Arrest

Abstract

Introduction: In out-of-hospital cardiac arrest (OHCA) a reliable rhythm diagnosis by the automated external defibrillator (AED) allows for the delivery of correct therapy, a prompt defibrillation for shockable rhythms and resuming cardiopulmonary resuscitation for non-shockable rhythms. The aim of this study was to develop a machine-learning (ML) based shock advice algorithm (SAA) for reliable rhythm diagnosis during OHCA. Materials and methods: The study analyzed data from 853 OHCA patients treated with AEDs by the basic life support personnel in the Basque Health service (Osakidetza, Basque Country, Spain) between 2013 and 2015. The datased used in the study contained 4212 5-s ECG segments, 489 shockable and 3723 non-shockable, annotated by consensus between six clinicians. The dataset was split patient-wise into training (60%) and test (40%) sets. Each ECG segment was preprocessed and 15 well-known waveform features were computed. The SAA was composed of two blocks. First, a low electrical activity (LEA) detector based on the power of the ECG for a prompt diagnosis of asystole. Second, a shock/no-shock ML algorithm based on a hidden Markov model that made the shock/no-shock classification of those segments not detected as asystole by the LEA detector. The training set was used to select the most discriminative features and develop/optimize the SAA. The test set was used to measure the performance of the method in terms of sensitivity (SE) and specificity (SP), and to compare it with the performance of a commercial AED. This procedure was repeated 50 times to estimate the distributions of the performance metrics. Results: The method showed a mean (SD) SE and SP of 97.7% (1.0) and 99.1% (0.4), respectively. While the commercial AED presented a SE and SP of 94.2% (1.3) and 99.8% (0.1). Conclusions: Both methods were compliant with the American Heart Association’s requirements (SE>90% and SP>95%). However, our ML approach outperformed the SAA of the commercial AED, increasing the SE in 3.5-points with a decrease in SP of 0.7-points. Therefore, a ML based SAA algorithm can accurately make shock/no-shock diagnoses during OHCA and might improve the performance of current algorithms.