Session P79.4

Detection of Ventricular Fibrillation by Sequential Hypothesis Testing of Binary Sequences

J Pardey*

Huntleigh Healthcare
Surrey, UK

The detection of ventricular fibrillation (VF) by sequential hypothesis testing of measurements taken from the surface ECG has been reported in a number of papers. Two measurements which demonstrate some utility for this purpose are the average interval between threshold crossings (TCI) and the complexity of the ECG as quantified by the normalised Lempel-Ziv complexity measure (CM). Both measurements work by converting the ECG to a binary sequence, and the published results show good performance on small, well-defined training and test sets. However, the ANSI/AAMI national standards EC38:1998 and EC57:1998 require that VF detection algorithms are tested on three standard databases, namely the MIT-BIH and AHA databases and the Creighton University Sustained Ventricular Arrhythmia database. This paper presents the results of using TCI and CM in conjunction with sequential hypothesis testing on these databases. The results reveal that TCI outperforms CM on two of the three databases, but the numbers of false positive and false negative VF detections are still too high for the method to be deployed clinically. To address this a new method for converting the ECG to a binary sequence is presented which takes into account the polarity of the ECG and yields greatly improved results for both TCI and CM.
Further analysis highlights two other weaknesses with the method. The first is the reliance on a single measurement, TCI or CM, to discriminate VF from non-VF rather than multiple independent, or quasi-independent, measurements. The second is the reliance of sequential hypothesis testing on a priori Gaussian approximations which, it is shown, do not generalise for new data. To address this the problem is reformulated in terms of the well-known pattern recognition paradigm of feature extraction and classification. It is then shown that using both TCI and CM as input features to a non-linear classifier is a promising alternative to sequential hypothesis testing of either measurement separately.

(Abstract Control Number: 42)