Shock Advisory Neural Network for Continuous Detection of Ventricular Fibrillation, Organized Rhythm and Asystole During Cardiopulmonary Resuscitation

This study presents a shock-advisory system (SAS) for continuous rhythm analysis during cardiopulmonary resuscitation (CPR) aiming to align with different resuscitation actions in out-of-hospital cardiac arrest (OHCA) for ventricular fibrillation VF (shock), organized rhythm OR (pulse check) and asystole ASYS (continue chest compressions). Two stages of convolutional neural network with raw ECG input (10s, 125Hz) were combined to generate a ternary rhythm decision (VF, OR, ASYS). In a total of 2838 OHCA interventions from automated external defibrillators (AED), we extracted episodes with VF (training/test: 409/393), OR (2151/2025), ASYS (4613/3979), around the standard AED analysis, following 2-minutes of CPR. The algorithm was applied in 1-second sliding mode, giving 21 decisions during CPR (CPR-ECG) and 10 decisions in presence of artifact-free ECG (CleanECG). The true positive rate of SAS on the test-set (CPRECG, Clean-ECG) is: VF (93.1%, 98.1%), OR (72.1%, 87.5%), ASYS (82.1%, 91.1%). Major OR false negatives (24.5%, 11.5%) are detected as ASYS, and inversely, major ASYS false negatives (4.5%, 12.6%) are detected as OR.These errors do not affect SAS specificity. With VF sensitivity (93.1%, 98.1%), OR specificity (96.6%, 99%) and ASYS specificity (92.9%, 99.2%), the presented SAS algorithm provides competitive performance during CPR.