Towards self-improving NN based ECG classifiers

Presents a method that allows to develop performant neural network (NN) classifiers by using undocumented databases to improve the learning process. A total of 1220 unvalidated cases was used in this study to enrich a small, however well documented ECG database containing 118 normals, 52 myocardial infarction and 75 ventricular hypertrophy patients randomly split into a learning set of 125 cases and an independent test set of 120 cases. The learning set was used to train a feedforward neural network that was in turn used to classify the undocumented database. These newly categorized cases were then merged with the initial learning set to form a new learning set that was again used to train the neural nets. The improvement of total accuracy obtained after a few iterations was >4% with final results comparable to those obtained by cardiologists.