Research Of ECG High-level Features For Beats Recognition

Frequencies and categories of heart beats are the important signs for diagnosing cardiovascular diseases. Automatic analysis systems classify beats according to their frequency component, mathematical abstractions of morphology, index of key points, and so on, while doctors make the judgement by semantic concept understanding. The difference between human being and computer in understanding signals causes “semantic gaps”. High-level features extraction can solve this problem and thus boost the performance of ECG beats recognition systems.In this paper, we focus on ECG high-level features for beats recognition and introduce the following achievements:1. We propose a high-level feature, Beats Vector Quantization(BVQ), for beats classification. The vector quantization feature has shown its superiority in the field of discrete-time series recognition, but existing vector quantization features are short of wave-diagnostic, which makes these features inappropriate for beats recognition. In order to solve this problem, we propose a wave normalization method, which detects components of the beat and normalizes them. The proposed method can boost the wave-diagnostic of vector quantization. Based on the method, the BVQ feature and the BVQ feature based beats recognition system are proposed. We compared our feature with traditional time-domain features and frequency-domain features on QTDB by feature-replaced experiment, and the results show that the proposed feature outperforms contrastive features in both accuracy and feature dimensionality, which suggests that BVQ feature is capable of increasing the accuracy and real-time performance of beats recognition systems.2. We propose an efficient dictionary for beats recognition, and improve the principle of dictionary construction. While dictionaries are decisive factor of recognition performance of the systems based on high-level features, the existing dictionary learning methods employ combination among dictionary training samples, making the dictionaries very sensitive to “dirty data”. For solving this problem, we propose an efficient dictionary, the construction principle of which is that the dictionary components are constructed by selecting representative samples. The proposed dictionary can avoid the interference from the dirty data, and thus increase the accuracy of systems. In order to test the effectiveness of the dictionary, we applied it in BVQ feature, employed QTDB as data source, built feature-replaced experiment, and compared the BVQ feature based on efficient dictionary with low level features such as time-domain features and frequency-domain features, as well as the BVQ feature based on traditional dictionary. The results show that BVQ feature based on efficient dictionary outperforms low-level features in both accuracy and feature dimensionality, which suggests that the BVQ feature based on efficient dictionary is capable of increasing the accuracy and real-time performance of beats recognition systems; the BVQ feature based on efficient dictionary outperforms the BVQ feature based on traditional dictionary in accuracy, which suggests that the proposed dictionary is capable of boosting the effectiveness of the high-level feature.3. We propose an object-attribution high-level feature and improve the principle of high-level feature extraction. Most of the existing high-level features are encoding features, the principle of which is extracting high-level information from low-level information using data mining. However, features based on this principle are not capable of representing position information and morphological information at the same time. In order to solve this problem, we propose an object-attribution high-level feature, which is based on a novel principle that employs object as attribution for representing high-level information, then directly extracts feature by modeling the high-level information. Under this principle, the object-attribution feature is capable of representing position information and morphological information at the same time. In order to test the effectiveness of the proposed feature, we employed QTDB as data source, built feature-replaced experiment, and compared it with low level features such as time-domain features and frequency-domain features, as well as the high-level encoding features such as BVQ features. The results show that proposed feature significantly outperforms contrastive features in accuracy, which suggest it is capable of increasing the accuracy of beats recognition systems.