|
Electrocardiogram, ECG for short, is a reflection of heart excitement and human life activity. ECG reflects how the heart is working and the changes of heart waveforms always stand for pathological changes. In modern medicine, ECG becomes a more and more important basis in diagnosing heart disease. Dynamic Electrocardiogram (Holter) records as long as 24 hours heart data- about 100,000 waveforms. As it is Impossible for a doctor to read all the ECG data, how to find out the typical waveforms for diagnosing accurately, reliably, completely and rapidly becomes a research hotspot now.Machine Learning mainly researches that how computer imitates or implements learning behavior of human being to learn more knowledge and reorganize knowledge structure in order to improve its capability constantly. Machine learning theories can be used in data mining of large scaled data. Key Technologies of machine learning mainly includes: integrated learning, Bayes network, decision trees, and statistic theory, support vector machines, Hidden Markov Model, neural network, k nearest neighbor, sequence analysis, clustering, rough set theory, the regression model, and so on.This paper mainly studies Holter waveforms, including Holter waveforms collection, preprocessing, comparing and finding a classification method and improving it. The concrete content of this paper is using wavelet transform and threshold detection together to detecting R waveforms, designing a Holter waveform automatic classification system, which combines clustering and classification together. This paper compares several classification methods, and finally takes K Nearest Neighbor as classification method with the result of clustering. By comparison of distance measurement method of similarity measurement, we use City Block distance instead of Euclidean distance and this paper introduces the concept of kernel function, which changes the linear difference of Holter waveform into non-linear difference. The experiments show that it does increase the distinction accuracy of Holter waveforms, and as a result the classification accuracy is also increased in a certain extent. We also put forward a refiltering strategy by combining SOM and Similarity Search together. It subdivides the Holter data which does not meet the requirement of classification after the first filtering stage. The result basicly fulfills the requirement of heart disease diagnosis.
|