Research Of Clustering Strategies For Dynamic Electrocardiogram Waveform

Dynamic electrocardiogram (Holter) had been applied in clinic extensively with the properties of simple, convenient and hurtless after its appearance. The primary value of the Holter was to present evidences for clinical diagnosis and treatment by finding and recording electrocardiogram (ECG) waveform changes, which were difficult to catch in normal ECG examination. Since there were 100,000 waveforms in one period Holter recording (24 hours), study points were focused on how to find out the individual variation waveforms quickly and accurately. The paper used clustering to pick out born waveforms in order to reduce the Holter waveforms for deep diagnosis. According to this method, the individual variation waveforms were clear and the diagnosis efficiency was improved.Based on analyzing the properties of Holter waveform and the examination features, clustering strategy in machine learning was presented to do the first analysis of the Holter waveforms, and result was used for the latter waveform selection analysis. Pretreatments were need for Holter analysis, because it was a weak signal, and contained strong noisy and strong randomicity. Through study, the wavelet transform (WT) was usually used in analyzing the time-varying signal uniquely because of its excellent time-frequency localization characteristic. But it can not be applied to real-time analysis and treatment due to its higher computational complexity. Another threshold detection method had the properties of high real-time analysis but low precision. By applying above two methods synchronously we got a better result in detecting R waveform in the Holter.Normally, there were two kinds of clustering strategy: one needed to know the number of categories k in advance, the other kind didn't. Since the value of k had a big influence on the clustering result, it can not be ascertained accurately without actual data. The method that didn't need k had a low computational accuracy or generates more categories than the actual class number. According to the properties of clustering strategies and the features Holter waveform, the paper presented two clustering strategies on waveform shape selection. In the SOM_AGG_k-means strategy, SOM neural network was used to do a rough clustering at first. The result of that can give us a fuzzy class number. And then, agglomerative clustering algorithm was used to obtain accurate class number. At last k-means was used to get the satisfied clustering result. In the Max-min Distance_k-means strategy, max-min distance clustering algorithm was used first to calculate the exact clustering number k and obtain the relevant representational clustering center. Then k-means algorithm was used to obtain the clustering samples for Holter data. The result showed that both strategies can reach the target of clustering similar waveforms together. From the view of class validity, the Max-min Distance_k-means strategy gave more accurate result than the first one.