On the use of artificial neural networks in a commercial holter algorithm

The Medilog ADAPT analysis algorithm uses multilayer perceptron artificial neural networks to estimate the Bayesian a posteriori probability that each detected QRS complex is supraventricular or ventricular given measurements such as the QRS width and height, etc. It is shown that the paradigm used in Medilog ADAPT improves on the standard paradigm for neural network classifiers in two important aspects. To solve the problem of missing data the features presented to the neural network´s inputs are first converted to probability estimates, so that when a feature cannot be measured it is assigned a value of 0.5. Then during neural network training the optimum neural network is the one that generates an output probability of 0.5 when its inputs are all assigned a value of 0.5. It is shown that using this neural network Medilog ADAPT is 99.9% accurate when tested against a database of eight 3-channel 24-hour ECG recordings.