Using a translation-invariant neural network to diagnose heart arrhythmia

The use of a neural network to classify ECG signals directly, without parametrization, is discussed. The input to such a network must be translation-invariant, since the distinctive features of the ECG may appear anywhere in the input window. The input must also be insensitive to the episode-to-episode and patient-to-patient variability in the rhythm pattern. A simple transformation of the ECG time-series input that meets both criteria is considered. A set of internally recorded, transcardiac ECG signals obtained from 54 patients was used to train and test the network. For each type of rhythm presented to the network, the network could output normal sinus rhythm (NSR), ventricular tachycardia (VT), and ventricular fibrillation (VF) (for networks trained to discriminate between VT and VF), or ambiguous (AMB)