Diagnosis of shockable rhythms for automated external defibrillators using a reliable support vector machine classifier

Sudden cardiac arrest is mainly caused by ventricular fibrillation and ventricular tachycardia, which are known as shockable rhythms. In this paper, a detection algorithm of shockable rhythms including support vector machine (SVM) model uses the public electrocardiogram (ECG) databases for training and testing. The databases are the Creighton University Ventricular Tachyarrhythmia Database (CUDB) and the MIT-BIH Malignant Ventricular Arrhythmia Database (VFDB). At first, to compose a set of good features, we extend a well-known set of 2 good features such as Count2 and VF-filter Leakage Measure (Lk). We supplemented 5 more good features, selected based on a binary genetic algorithm-based feature selection, among 11 new input candidate features. All the combinations of 7 good features are estimated for their performance on the training and the testing data using the SVM models to identify 6 combinations of the final feature pool. 5-Folds cross validation is then implemented carefully to validate the performance of the SVM classifier using final feature pool on separated and entire 5s-segment databases. The final combination of 4 features, which includes Count2, Lk, Threshold Crossing Interval (TCI), and Centroid Frequency (CF), is addressed by the highest validation performance of the corresponding SVM model. The Count2 shows the proportion of the signal, which is above the mean absolute values of the output of an integer coefficient recursive bandpass filter computed for every 1 s time interval. The Lk represents the output of a narrow bandstop filter, which is applied to the ECG signal with the central frequency being the mean signal frequency. The TCI shows the average time between the fixed thresholds, which are computed for every 1s-segment using the pulses converted from the ECG signal. The CF is the frequency, which bisects vertically the area under the power spectrum. For the proposed algorithm, the average accuracy of 95.9%, sensitivity of 91.7%, and specificity of 96.8% are archived on the evaluation data of the entire database. Comparing to the performance of the SVM model using a combination of the Count2 and the Lk, we report a significant improvement for the accuracy of the SVM model using the final feature combination in average, i.e. 2.6%, 22.4%, and 2.7% on the evaluation data of the entire database, the CUDB, and the VFDB, respectively. Furthermore, existence of ventricular ectopic beats in the input data shows a negligible influence on the final performance of classification.