Modular connectionist architectures for multi-patient ECG recognition

Arrhythmia recognition in implantable cardioverter-defibrillators is based on timing information derived from the electrocardiogram. It has been shown that it is possible to achieve improved recognition of arrhythmias by using multilayer perceptron based classifiers. Neural network based classifiers trained on multiple patients have also exhibited a limited degree of patient independence. Patient independence is difficult to achieve, however, since the pattern vectors derived from the electrocardiograms of each patient have unique statistical characteristics. It is possible that patient independent classifiers can be made more robust if the database used for training can be partitioned into independent homogeneous subgroups of patients, where the distribution of data from each patient can be represented by similarly parameterised statistical models. This paper is concerned with the application of a modular connectionist architecture, that combines associative and competitive learning, to identify homogeneous subgroups of patients within a larger patient population. The study is confined to the classification of sinus tachycardia and ventricular tachycardia on the basis of electrocardiogram morphology