Recurrent Neural Networks to Predict the Outcome of Subsequent Defibrillation Shocks in Cardiac Arrest

Accurate prediction of defibrillation success could help to optimize the treatment of out-of-hospital cardiac arrest (OHCA). On top of the classical predictors derived from the ventricular fibrillation (VF) waveform, the outcome of the preceding shock can also be used to enhance prediction. This work introduces a recurrent neural network (RNN) to predict the outcome of subsequent shocks, using the full history of previous outcomes and VF measures. Data from 957 OHCA patients were analyzed, comprising 3159 shocks of which 2202 were subsequent shocks. Shocks were labeled as successful/unsuccessful based on their ability to restore an organized rhythm. Each patient data were modeled as a time sequence, one shock per timestep, and fed to a RNN. Individual shocks were characterized using a single classical predictor (computed on a 2s pre-shock VF window) and the outcome of the previous shock. Seven predictors were independently analyzed. Model performance was assessed through 10-fold cross-validation (CV) over 50 different CV data partitions. Balanced accuracy (BAC) was chosen as target performance metric. Mean Slope was the best predictor, with median (interquartile range, IQR) BAC of 84.2 (84.0 – 84.4)%, 3.6 points above that of simple thresholding with no previous shock information. RNNs could improve subsequent shock-outcome prediction using previous shock history.