Using within-patient correlation to improve the accuracy of shock outcome prediction for cardiac arrest

Analysis of the electrocardiogram (ECG) can to a certain extent predict if a cardiac arrest patient in ventricular fibrillation will get return of spontaneous circulation (ROSC) if defibrillated. The accuracy of such methods determines how useful it is clinically and for retrospective analysis.

We have tested the accuracy of a new shock outcome prediction algorithm that is the first to use an updating algorithm capable of learning from previous shocks within a resuscitation effort. The algorithm relies on known predictive features from the pre-shock ECG, but for each delivered shock it re-estimates the patient-dependent relationship between predictive feature value and probability of ROSC by incorporating the information from the already performed shocks. The predictive features mean-slope, median-slope, cardioversion-outcome-predictor and amplitude-spectrum-analysis originally had areas under the receiver operating characteristics curve of 0.843, 0.846, 0.837 and 0.819, respectively. The improvements in areas after applying the algorithm were (bootstrap estimate of mean improvement, 95% confidence interval in parentheses): mean-slope, 0.019 (0.00036, 0.042); median-slope, 0.024 (0.0013, 0.048); cardioversion-outcome-predictor, 0.021 (0.0010, 0.051); amplitude-spectrum-analysis, 0.026 (0.0016, 0.051). The predictions for the first shock to each patient were not included when calculating the areas, as for the first shocks the new algorithm has no previous shocks to learn from and give predictions identical to those of the original features.

It is possible to improve current shock prediction methods by using an updating algorithm capable of learning from previous shocks within a resuscitation effort.