Applications of statistical pattern recognition techniques to the analysis of ballistocardiograms
This dissertation describes a new approach to the analysis of Ballisto-cardiograms using Statistical Pattern Recognition technique, as well as the design and development of a new ballistocardiograph and its associated software. Ballistocardiograms are the result of forces exerted on the body, caused by the ejection of blood from the heart and the passage of the blood through the arterial system. The apparatus used in this study for the collection and display of the ballistocardiograms consisted of a specially designed stool with highly sensitive piezoelectric elements, which converted the forces acting on the stool into electric signals, connected to a specially built interface which converted the analogue signal into digital data. These were in turn analysed using a BBC model B microcomputer running special software which was written as part of this work. The methods of analysis developed here are based on Statistical Pattern Recognition and consist of units dealing with the preprocessing of the data, extraction of optimal features, and with their classification. By their nature, the lengths of the ballistocardiograms vary not only from person to person but there are also differences between the lengths of individual beats in the same subject. This presents a major problem for successful analysis. A novel method for non-linear standardisation of the ballistocardiogram length was developed and used in this study. This method allows the comparison of ballistocardiograms of different lengths, by projecting them into a waveform of uniform length, whilst maintaining all the information contained in the shape of the original signal. The projection is based on local cross-correlation of a template ballistocardiogram with a subset of the ballistocardiograms to be analysed. This results in a set of standard length records which in turn are used to determine the transformation. A feature extraction method based on double eigenanalysis was used to reduce the dimensionality of the data and for the extraction of features which discriminate best between the different classes analysed. Four classes of subjects were used in this study. A normal group which consisted of generally healthy and physically fit people, whose ballistocardiograms were also used to develop the new method for length adjustments; a group of subjects with mild hypertension; a group of patients with coronary artery stenosis, who were undergoing treatment at the Papworth Hospital; and a group consisting of subjects with clinical history of recent myocardial infarction. It was found that after standardisation of the length of the ballistocardiograms, and after extraction of those features which contain most of the discriminant information, the Nearest- Neighbour rule discriminated well between the group of normal subjects and the three remaining groups. The groups of subjects with mild hypertension and with coronary artery stenosis proved more difficult to separate. This can possibly be explained by the similarities in the characteristics of these two groups as far as ballistocardiograms are concerned. It was also found that the parts of the wave that have most of the discriminatory information are those corresponding to the ejection phase, for all the groups in general, and those corresponding to the last peaks of the ballistocardiograms (post ejection phase), for the group with recent myocardial infarction. Ballistocardiography is shown in this work to be a good non- invasive method for the study of the general performance of the heart. The methods described here for discrimination between groups and classification of different ballistocardiograms, by means of the analysis of their shape alone, have also proved very powerful. In particular, the new length standardisation method allows a more accurate monitoring of the heart function, than could be achieved so far. The techniques developed in this researh may be used for the prediction of various heart diseases in their early stages. This, together with the portability of the apparatus developed in this research, could turn the new ballistocardiograph into a standard clinical device.