| Summary: | 碩士 === 中原大學 === 醫學工程學系 === 87 === By long-term monitoring ECG, which was recorded by Holter system, provides important information about the patient’s heart function. However, what the physicians concern most is how to diagnosis heart diseases, efficiently and accurately, from this large amount of data. Therefore, an automatic ECG diagnosis system that combines the wavelet transform (WT) and fuzzy neural network (FNN) is proposed and developed.
The proposed system uses the characteristic of WT; e.g. the ability of time-frequency and multi-scale filter banks analysis, to increase the reliability of extracted ECG features. Additionally, the FNN is used to classify the premature ventricular contraction (PVC) from the normal heartbeats based on the WT features.
In this study, two WT features, specifically the width of QRS duration in scale three WT and the area of QRS complex in scale four WT, and three FNN structures were studied. The results indicate that the WT features are influenced by the high frequency noise much less than low frequency drift. In addition, the simplified FNN performs best in the three structures studied.
Seven files from the MIT/BIH arrhythmia database were used to evaluate the proposed system. One of these seven files is dominated by left bundle branch block (LBBB) beats. Two test files are dictated by normal heartbeats. On the other hand, the rest of test files have large amount of PVC heartbeats. The results indicate that the combination of WT and FNN can achieve 97.18% of accuracy in the PVC detection. When the file that is dominated by LBBB is excluded, the accuracy is 99.94%.
Furthermore, four clinical Holter ECG recordings obtained in the Kaohsiung Veterans General Hospital are used to evaluate the proposed system. When compared with the dynamic-threshold classification system, the proposed system suffers much less false positive error. Thus, this study shows that the combination of WT features with FNN can be a reliable PVC detection system.
|
|---|
