Abstract 242: Prediction of Shock Outcome Using Signal Processing and Machine Learning
Introduction: Chances of successful defibrillation, and that of subsequent return of spontaneous circulation (ROSC), worsen rapidly with passage of time during cardiac arrest. Each shock delivered with intent to achieve ROSC, if unsuccessful, causes electrical damage to cardiac tissue, in addition to adding to the time lost. The predominant arrhythmia of cardiac arrest is ventricular fibrillation (VF). The Electrocardiogram (ECG) signal of VF has been analyzed for certain characteristics which may be predictive of successful defibrillation. To date, no analytical technique has been widely accepted. We propose a new method to analyze the chaotic nature of VF using multiple feature extraction and machine learning techniques.
Methods: In an IRB approved protocol, cardiac arrest was electrically induced in 6 swine subjects for 12 minutes. Animals then underwent 5 minutes of CPR, followed by defibrillation countershocks(CS), CPR, drug therapy and additional CS as needed. Butterworth filter was used to remove low-frequency noise and breathing artifact from the ECG. Devising feature extraction from the ECG signal involved considering VF as a chaotic rhythm that represents an overall dysfunction, and comparing it to the more regular rhythms. Features of the signal in time and frequency domains, supplemented by the use of Wavelet transform (for decomposition), were extracted. Daubechies 4 wavelet, previously shown to provide better time-frequency resolution with the ECG, was used at 4 levels. A total of 22 features were utilized. Machine learning methods were applied for classification of shocks as either successful or unsuccessful.
Results: A total of 26 CS were analyzed. They were best classified as successful or unsuccessful by employing the Random Tree method, with number of random attributes set to 5. Received Operator Characteristic area under the curve for the model was 70.6%. A sensitivity of 73.1%, along with a specificity of 72.7%, was achieved. Classification could be performed on ECG tracings of 40 seconds.
Conclusions: Real-time, short-term analysis of ECG, through signal-processing and machine-learning techniques, may be valuable in determining CS success. As other features of resuscitation are analyzed, improved performance is anticipated.
- © 2010 by American Heart Association, Inc.