Research On Representation Methods And Similarity Measures Of Ecg Time Series

For a long time, heart disease has been the first killer of human health. ECG provides important information on the functional activity of the heart, can help diagnose some common heart diseases. By detecting the QRS complex, P wave, T wave, conventional ECG recognition algorithm extracts characteristics from each waveform, and then classifies ECG samples based on these characteristics. At present, the detection of the QRS complex is relatively mature,but it still cannot effectively detect the P wave and T wave.With the advance of medical information, hospital and community have accumulated vast amounts of ECG data. Time series similarity search method does not need to detect specific waveform. By checking out the most similar sample from the ECG database, it can determine the type of the test sample. The main topic of this thesis is ECG time series similarity search. This thesis focuses on the study of representation of ECG time series and similarity measure. The following are the main research work and achievements of this thesis:ECG time series is a high-dimensional data. In order to facilitate storage and computation, it needs to compress and represent ECG time series. According to characteristics of ECG time series, this thesis proposes a method of piecewise linear regression based on residual balance and boundary constraints(RBBC_PLR). At the same compression ratio, RBBC_PLR maintains a smaller fitting error, while effectively removing noise.After piecewise linear representation, ECG time series turns into a set of segments from observations. In order to measure the "distance" between two segments, this thesis defines the concept of "conversion cost between two segments". Based on the "conversion cost", this thesis proposes a new similarity measure: dynamic time warping based on conversion cost(Cost_DTW). Experiments show that this similarity measure can more quickly and accurately search out similar ECG samples from the ECG database.RBBC_PLR, Cost_DTW and k NN constitute the ECG time series similarity search system. This system is able to achieve ECG automatic identification effectively.