Neural network based analysis of the signal-averaged electrocardiogram
Martin Höher, Hans A. Kestler, Sydney W. Bauer, PETER WEISMÜLLER, G. Palm, Vinzenz Hombach · 2002
Standard time-domain late potential analysis of the signal-averaged ECG is based on the QRS duration and the terminal low-amplitude portion of the QRS. The authors evaluated the capacities of neural networks (NN) to differentiate patients with and without malignant arrhythmias based on the complete QRS data without prior parameter extraction. In 74 patients with and 116 patients without inducible ventricular tachycardia (sVT) signal-averaged ECGs were recorded. Following high-pass 40 Hz filtering and non-linear scaling (tanh), the vector-ECG was used as input to a backpropagation network with 230 inputs and 3 layers. The network was trained to discriminate between patients with and without sVT. NN classification was comparable to standard VLP analysis in terms of accuracy (66% versus 65%) specificity (72% versus 61%) and positive predictive value (56% versus 54%). Potential advantages of the NN approach are its independence from an exact QRS-offset computation and its ability to handle noisy signals.