Compression of ECG signals with feature extraction, classification, and browsability.

摘要

This thesis presents nonlinear analysis techniques to characterize and compress a nonstationary electrocardiogram (ECG) signal in order to alleviate some of the limitations of linear methods in biological signals. The motivation behind the study is the need to reduce the size of storage and the time of transmission and analysis of the ECG signal since a long-term ECG recording produces large amounts of data. This research has resulted in three methods for ECG data compression: (i) nonlinear iterated function systems (IFS) compression and compositional complexity partitioning, (ii) statistical feature extraction and classification to compress and browse the ECG signal, and (iii) dynamic time warping (DTW) classification and block encoding.; A nonlinear IFS (NIFS) has been developed to compress ECG data based on self-similarity of the signal. Compared to the traditional IFS, the NIFS provides more flexible modelling. It gives a compression ratio of 7.3:1, which is higher than that of 6.0:1 in orthogonal fractal technique of Oien and Narstad, under a reconstruction error of 5.8%. Furthermore, to reduce computational complexity, a variance fractal dimension trajectory (VFDT) is applied to partition the ECG signal by measuring the local compositional complexity of the signal. The segmented NIFS technique reduces computational complexity to O(N) for a time series with length N, compared to O(N2) of the IFS, and achieves a compression ratio of 5.7:1 under the same reconstruction error.; A feature extraction and classification method is also proposed to reduce the redundancy among the beats of the ECG signal. Instead of comparing the beats of the quasiperiodic ECG signal directly, statistical features with the same dimension are extracted from the beats with various lengths. First, we propose to extract moment-invariants by treating ECG beats as character images, thus preserving the shape information of ECG waveforms. A probabilistic neural network is used to classify ECG beats through such features. The classification information is applied to remove the redundancy among ECG beats and to browse the long-term recording in ECG analysis. Secondly, the Renyi multifractal dimension spectrum based on the Renyi entropy measure of the object is extracted by modelling the ECG signal through its underlying strange attractor. The investigation of the ECG strange attractor helps us recognize limitations of the low-sampling frequency (360 Hz) and noise in phase space reconstruction of the ECG signal. Therefore, denoising techniques are applied to the ECG signal to remove the noise and thus improve the convergence of the Renyi dimension spectrum. A mean absolute difference (MAD) between two Renyi dimension spectra is proposed to measure the convergence of the Renyi dimension spectrum. Experimental results show that chaos denoising improves the convergence about five times under the MAD, while wavelet denoising degrades it about two times.; Finally, a modified DTW is applied to classify ECG frames by time normalizing two frames through a warping function. Since the warping function is complicated, a segment-based registration of ECG frames is proposed to approximate the ECG frame. Partitioning of ECG frames is realized by a windowed-variance technique. Then the block encoding is proposed to compress segments of ECG frames. This compression scheme achieves a very high compression ratio of about 50:1.