Fast Detection of Ventricular Tachycardia and Fibrillation in 1-Lead ECG from 3-Second Blocks

Filip Plesinger1, Petr Andrla1, Ivo Viscor1, Josef Halamek2, Pavel Jurak1
1Institute of Scientific Instruments of the CAS, 2Institute of Scientific Instruments, CAS, CZ


Abstract

Background: Ventricular tachycardia (VT) may lead to heart failure or death. Therefore, Holter ECG devices are used to report the occurrence of these VT events. The problem is that regular (non-hospitalized) patient activity during the daytime may lead to a large number of false-positive VT events. Here, we present a method detecting VT and ventricular fibrillation (VF) events, suitable for real-time application on continuously incoming ECG data.

Method: We designed a method for detection of VT/VF events in short-time (3 s), 1-lead ECG blocks. Five features are extracted from this block using analysis of ECG spectra, derivatives, amplitude measures and auto-correlation. The extracted features are put into a logistic regression model showing the probability of a VT/VF event. The model was trained on the public PhysioNet CUDB dataset consisting of 393 automatically selected blocks.

Results: The model (AUC 0.99) showed a sensitivity and specificity of 95% and 97%, respectively (5-fold cross validation). The model was also tested on the public PhysioNet VFDB dataset, showing specificity and sensitivity of 95% and 83%, respectively. Both the feature extraction code (Matlab format) and the model are publicly accessible and easy implementation of the logistic regression model predetermines it for real-time applications.