Machine Learning and Cardiology

Project Overview

1.1 Problem statement

In this project we will study the separation of foetal and maternal ECGs, based on ECG recordings during different stages of a (human) pregnancy. From a set of real ECG recordings, the challenge is to extract the separate foetal and maternal components, and interpret the ECGs individually. This involves signal analysis using Fourier and wavelet decomposition, pattern recognition and data mining. Moreover, the hypothetical aim of this project is a portable automatic monitoring device that can also communicate autonomic - if necessary - to a gynaecologist by sending relevant data automatically through mobile phone or internet. This involves telecommunication as the signals must be compressed, stored selectively, sent, and decompressed.

All these aspects relate to some of the courses offered in the blocks during the project. The final product of this project is a software tool that is capable to separate foetal ECGs with a high degree of specificity and sensitivity, and indicate (recognize) some pre-defined pathological conditions in either foetus or mother.

 

1.2 Bio-medical Background

A relevant problem in Biomedical Engineering related to Signal Processing is the extraction of foetal electrocardiograms (ECGs). This problem arises for instance when the foetal heart condition is desired to be monitored. So-called invasive techniques have been shown to be very accurate but require access to the foetal scalp and therefore are only feasible during delivery. An earlier diagnosis during pregnancy using non-invasive techniques is consequently essential. These usually involve measurements from electrodes attached to different points of the mother's skin.As a result the recordings pick up a mixture of foetal ECG (FECG) and maternal ECG (MECG) contributions. In addition, other random disturbances must be considered which become important due to the low level of the FECG signals, such as the maternal electromyogram (EMG), interference of thermal noise from the electrodes, and other electronic equipment, etc. Under normal conditions some of these disturbances can be eliminated. For instance a persistent EMG interference usually indicates an uneasy position of the mother. Moreover depending on the stage of pregnancy EMG interference produced by contraction of the uterus may be observed as well as baseline wandering during foetal movements. Nevertheless these disturbances are rather occasional and most of the time the main one is due to thermal noise which is an apparently white-noise signal. Hence it is necessary to extract the wanted foetal contributions from the rest of non-desired components that corrupt the cutaneous recordings. This task of obtaining the FECG from skin electrode signals is another example of the so-called inverse problems which can be coped by using different techniques. Since the mother's and foetus' heartbeats show different rates, the first approach one may come up with is a simple filtering in the frequency domain.However, there exists a spectral overlap between the maternal and the foetal ECG signals which does not make this solution very successful. Furthermore, the heart rate is not constant but exhibits a random variation with time (the so-called heart rate variability, HRV) which also does not recommend the use of frequency filters to tackle the problem. A more parametric approach seems to be unsuitable. In the first place, the parameters of the model would be subject to a large uncertainty, owing to differences from patient to patient. In the second place, the parameters of a single patient would not be time invariant, for instance because of the foetal growth.Finally and most fundamentally, because in order to be applicable to medical diagnosis and treatment the procedure should be able to deal with unexpected ECG patterns. All these considerations suggest that the FECG extraction has to be formulated as a so-called blind identification problem.

 

1.3 Project Coordination & Members

The examiners of this project are Ronald Westra and Jos Uiterwijk. Jan Paredis is coordinating the project and tutoring the groups. Every week he will have a meeting with each project group.

The members of this project are:

 

Index | Next >>>