Classifying ischemic events using a Bayesian inference Multilayer Perceptron and input variable evaluation using automatic relevance determination

Abstract

In this paper we present a Bayesian inference Multilayer Perceptron (MLP) which was used to classify the events of the Long Term ST Database (LTSTDB) as ischaemic or non-ischaemic episodes IntroductionMyocardial ischaemia is one of the most common fatal diseases of the western industrial world. It is a heart problem which is caused by the lack of oxygen and nutrients to the contractile cells (muscles) and leads to dangerous arrhythmias and myocardial infractions. The methods which are employed to detect myocardial ischaemia are based on the measurement of blood flow and oxygen supply of the heart. Two of these methods are coronary angiography and exercise test which are either very expensive or very exhaustive for the patients . These are the reasons why these methods are applied only to high risk patients.Myocardial ischaemia can also be detected from the abnormalities that are depicted in the ST segment of the electrocardiogram (ECG) despite the fact it does not contain any information about the blood flow and the oxygen supply of the heart [2]. This method is cheaper than coronary angiography and demands less effort from the patient than the exercise test. Nevertheless, we should bare in mind that apart from myocardial ischaemia the abnormalities that are observed in the ST segment of an electrocardiogram can be also the result of many other factors such as changes in the heart rate, the position of the subject, noise in ECG.The development of a classifier that will be able to identify whether the changes in the ST segment are caused from ishaemia or from other reasons, was the challenge of Physionet and Computers in Cardiology of 2003.The Long term ST Database (LTSTDB) was used for the purposes of this challenge. LTSTDB contains 86 records of 19-26 hour ECG, 43 of these records are available from Physionet as a training set for the algorithms which were developed to detect ischaemia.In each one of these 43 records the significant ST episodes were specified. The procedure which was used in order to determine these episodes is [3]: a) an episode begins when the ST deviation exceeds the 50µV b) the ST deviations must reach a threshold value V min for a period of time T min c) the episode ends when the ST deviation becomes less than 50µV continuously in the following 30 seconds Three different protocols were used to define whether these events were ischaemic or non-ischaemic according to different combinations of the values V min and T min . The protocol which was used for this paper was the protocol b. The values of V min and T min for that protocol were 100 µV and 30 seconds respectively. Using these values of V min and T min we located 1772 episodes in the LTSTDB, 1369 of which were non-ischaemic and the rest 403 were ischaemic.In this work we also used the Physionet annotation files in order to be determined the J point of the beginning of each episode, the position of the R peak and the values of the ST deviation for each episode.The remainder of this paper is set out as follows: The next section c...

Plane-Gaussian artificial neural network

“…MLPs and RBFNs, two types of classical multilayer feedforward neural networks, have been widely studied and got many successful applications such as in face recognition, disease diagnosis, risk investment, and so on [1][2][3][4][5], here just name a few. It had early been proved theoretically that they both are capable of approximating any continuous function or a mapping from an input space to an output space to arbitrary precision only if provided with sufficient hidden neurons [6][7][8][9][10].…”