Filling in the gap: A general method using neural networks

When a set of medical signals has redundant information, it is sometimes possible to recover one signal, from its past and the information provided by the other signals. In this work, we present a general method to realize that task. It has been known for a long time that multilayered networks are universal approximators, but, even with the backprop algorithm, it was not possible to train such a network, to realize complex real life tasks. In the last years, Geoffrey Hinton presented a training strategy that allows to overcome the previous difficulties. We describe a way of adapting Hinton´s strategy to our task. An example of a situation considered here, consists on training a Multilayered perceptron to take ECG leads II and I as input and produce as output missing lead V. This method got the best scores among participants in the Physionet/Computing in Cardiology Challenge 2010.