週期性時間序列分群並應用於心電圖分析

A nonparametric model for clustering ECG data

葉琮銘(TUNG-MING YEH)

指導教授 ： 林財川

國立臺北大學/商學院/統計學系/碩士(2008年)

摘要

時間序列資料普遍地存在日常生活中，而時間序列資料本身兼有隨著時間週而復始的週期特性時，又可將其細分為週期性時間序列。舉例來說，人類晝夜節律基因、人體的體溫以及心電圖。近期在週期性時間序列資料分群研究上有越來越廣泛的趨勢。為克服參數化模型在線性、定態及對稱資料上使用的限制，本論文決定從非參數函數週期模型的角度出發。透過非參數函數週期模型的序列特徵萃取的特性來進行週期性序列的分群。
本文，主要針對心電圖週期性時間序列資料進行研究及剖析。就心電圖資料來說，其是由許多單一週期的心電圖所組成，而單一週期心電圖又可視作幾個特定的波形所組成，而且各個波形都有其對應上的特徵，這些特徵又是做為心律不整判斷上的依據。因此本篇論文企圖建立心電圖各個波形總和的模型，並希望透過模型的結果來表達波形的特徵，進而利用這些特徵來進行各單一週期心電圖的病症分群。過程之中，由於心電圖的波形相加效應以及波形結構的非對稱性，因此利用非參數估計中之廣義可加性模型和非參數函數週期模型，把單一循環週期下心電圖資料藉心電圖可加模型來表示。
根據模型能有效萃取波形特徵的特性，可將單一循環週期下心電圖各特徵波形的平均、振幅與相位潛在訊息估計出來。透過兩種以參數估計為基礎的分群方法，分別是參數差異性檢定分群方法以及參數距離分群方法，將不同標的心電圖序列資料予以分群。在 MIT-BIH 心律不整資料庫編號 201 心電圖記錄資料實證應用上，將 25 筆單一週期心電圖資料分為 5 群的分群相似度可達 0.757215，顯示所提出的分群方法具有一定的分群能力。

關鍵字

週期性時間序列 ； 模型基礎分群 ； 非參數模型 ； 心電圖

並列摘要

Time series data generally exists in daily life. When time series data have the periodic that goes round and begins again along with time, can subdivide into it again periodic for time sequence. Give examples to say, human circadian rhythms gene and body temperature and electrocardiogram. In the near future at clustering periodic time series data study up have more and more extensive trend. For overcoming parameter model the restriction of usage is on the linearly and stationary and symmetry data. The paper decides to set out from the angle of the nonparametric function periodic model.
This text mainly aims at the data of the electrocardiogram periodic time series to carry on research and analysis. In regard to electrocardiogram data, it by many single electrocardiograms of period constitute, and single period electrocardiogram can see to make a few particular wave forms to constitute again, and each wave forms all have its characteristic in shoulding ascend, these characteristic again is be used as arrhythmias judgment up of basis. Therefore this paper attempt to build up the model of the electrocardiogram each waveform total, and hope to express a characteristic of form through the result of model, Then make use of these characteristics to carry on the disease of each single period electrocardiogram for clustering. Process in, because the wave of the electrocardiogram form mutually adds an effect and wave form structure not symmetry. Therefore make use of nonparametric estimate in of the general additive model and the nonparametric function period model. We can indicate single circulation period electrocardiogram data model with ECG addive model.
According to the model, we can effectively extract a characteristic of waveform, and can single circulation period descend each characteristic of electrocardiogram waveform of mean, amplitude and phase the latent message estimate come out. Through two kinds of parameter-estimate-based clustering the method is the parameter differences test clustering method and parameter distance clustering method, give different object electrocardiogram series data clustering. Applied in 201 electrocardiogram record data from MIT-BIH arrhythmias database, 25 single period electrocardiogram data be divided into 5 clusters, and similarity can reach to 0.757215. Show put forth of the clustering method has certain centses ability.

並列關鍵字

periodic time series ； model-based clustering ； nonparametric models ； ECG

參考文獻

Ceylan, R. and Y. Özbay (2007), “Comparison of FCM, PCA and WT techniques for classification ECG arrhythmias using artificial neural network,” Expert Systems with Applications;33:286-295.

Cleveland, W.S. (1979), “Robust locally-weighted regression and smoothing scatterplots,” J.Am. Statist. Assoc;74:829-836.

De Boor, C. (1978), “A Practical Guide to Splines,” Springer.

Dempster, A.P., N. Laird, and D.B. Rubin (1976), “Maximum likelihood estimation from incomplete data using the EM algorithm (with discussion),” J. Roy. Statist. Soc. Series B;39:1–38.

Dokur, Z. and T. Olmez (2001), “ECG beat classification by a hybrid neural network,” Comp Meth Prog Biomed;66:167-181.