Abstract 182: Integrating Physiologic Signals with Machine Learning for Predicting Defibrillation Success
Current defibrillation success prediction methods suffer from low accuracies and impracticable specificities because they rely solely on identifying features and building predictive models from the electrocardiogram (ECG) signal. Features derived from ventricular fibrillation (VF) waveform analysis suffer from random effects which may vary from 73% to 189% of effect-size in the mixed effects logistic regression model. We hypothesized that other signals can help counter this observed model variance.
Capnography (PetCO2) and ECG data was parsed from records of 48 prehospital subjects in VF receiving defibrillation. ECG signals were used to build a model to predict defibrillation success in terminating VF, leading to return of spontaneous circulation (ROSC). Corresponding PetCO2 signals were then added to the ECG model.
Features that could be valuable in predicting the success of a countershock were extracted using a dual-tree decomposition algorithm for a complex wavelet transform and a novel non-linear stochastic analysis method. The latter maximizes divergence between distributions of periodicities in signals belonging to different classes (successful versus unsuccessful defibrillation).
Training and testing were performed with double/nested10-fold cross validation. A total of 6-10 features were used to build models using Support Vector Machines (with Radial Basis Functions), which have been shown to be universally consistent. Integrating PetCO2 with ECG features boosted ROC Area Under the Curve from 81% to 93.8%. Accuracy increased from 79% to 83.3%. A large ROC AUC allowed for 90% Sensitivity and 78.6% Specificity at a classifier-output threshold value of 0.25.
Within this small dataset, the addition of PetCO2 to significantly improve model performance. These findings are consistent with physiologic understanding of PetCO2, cardiac output, and coronary perfusion pressure. The burden of random effects may be countered by inclusion of multiple physiologic signals.
- © 2012 by American Heart Association, Inc.