Modelling T-end in Holter ECGs by 2-layer perceptrons

Automated detection of T-end in high precision is required for ECG safety assessment of new chemical entities (NCEs). This task may be effectively accomplished by neural networks (NNs). Since it is scarcely known which configuration of NNs to choose for obtaining an optimal prediction, we explore a variety of layouts for a 2-layer perceptron. Our training reference is the Physionet QT database with expert T-end annotations. The filtered and re-sampled signal from both channels spanning a variable time interval that contains the main part of the T wave is our model input. We investigate model variations by number of sampling points and hidden units. We train these models using Bayesian techniques and compare their properties by the evidence parameter cross validation error goodness of fit and the estimated prediction error. While evidence and cross validation error favor medium sized models, residual standard deviation decreases to approximately 12 ms, whereas the estimated prediction error increases under growing model size. A medium sized 2-layer perceptron (15 sampling points over the T wave on individual channels, 15 hidden units) is suitable to describe expert annotations of T-end with a residual standard deviation of 15 ms. This configuration has promising generalization capabilities and can handle all T morphologies found in the training data set.