Abstract 304: Comparing a Novel Stochastic Integrative Machine Learning Model with AMSA for Predicting Defibrillation Success
A computational model of defibrillation success, based on physiologic signals, would be highly useful in guiding therapy during cardiac arrest. The Amplitude Spectrum Area (AMSA) measure has been explored in a number of studies as a predictor of countershock success for ventricular fibrillation (VF). We have developed a novel non-linear stochastic method for a unified predictive model of Return Of Spontaneous Circulation (ROSC) and compare this model to AMSA.
Ninety sample pre-hospital VF Electrocardiogram (ECG) signals including defibrillation were classified as successful (34) or unsuccessful (56) through 10-fold cross validation. A novel integrative model was developed with Machine Learning and features extracted from non-linear stochastic characterization of the ECG signal. This method was subsequently compared with the AMSA method. We replicated the procedure published by AMSA creators and attempted to discern a threshold. The performance of both methods in predicting post-countershock ROSC was measured. Accuracy is calculated as the correctly predicted countershock outcomes divided by all samples in the database. ROC AUC is a measure of predictive model robustness.
The novel integrative model performs real-time, short-term (7.8 second) analysis of the ECGs with prediction Accuracy and Receiver Operator Characteristic (ROC) Area Under the Curve (AUC) of 82.2% and 85%, respectively. Sensitivity was 85.3% and specificity was 80.3%. Using the AMSA methodology, no clear threshold for shock success could be identified for comparison. Employing information-gain based decision stump (or one-rule) for AMSA values yielded 64.6% accuracy and 60.9% ROC AUC. Sensitivity was 44.1% and Specificity was 77.2%.
We report multiple performance metrics for a novel integrative model as well as the previously described, Fourier Transform based, AMSA model for predicting defibrillation success. A stochastic and non-linear basis for signal characterization proves to be more adept in dealing with the known dynamic and non-stationary characteristics/morphology of the VF waveform. It significantly improves performance in analyzing the VF signal to predict countershock success.
- © 2012 by American Heart Association, Inc.