Title: How to teach a machine to read an electrocardiograph
Authors: Andreas Kryger Jensen - University of Copenhagen (Denmark)
Thomas Alexander Gerds - University of Copenhagen (Denmark) [presenting]
Abstract: The statistical challenge to train a computer for the task of interpreting a digital electrocardiograph (ECG) is considered. The aim is to predict the current health state of the patient's heart and the development of future heart diseases. To capture the signal in these highly correlated data we develop a two-step approach. The first step consists of a series of unsupervised functional data analyses performed separately for each of the 12 leads of the ECG. The second step uses the health outcome to combine the lead specific patterns into a personalized prediction. For each outcome separately, we train the model regarding the sparse multivariate principal modes of association in order to achieve optimal predictive performance invalidation data.