Compression of the electrocardiogram (ECG) using a library of orthonormal wavelet basis functions

ECG signals are recorded for diagnostic purposes in many clinical situations. In order to permit good clinical interpretation, data is acquired at high resolutions and sampling rates. The substantial volume of data generated warrants the use of reliable and efficient data compression techniques. Compression is required to reduce (a) storage requirements of hospital databases and ambulatory ECGs, and (b) the time to transmit data over telephone lines, when remote interpretation is required. Due to the large data storage requirements, particularly in the case of ambulatory ECG monitoring (AECGM), there is a strong need for enhanced compression. Existing techniques suffer from relatively poor compression ratios (CR), limited robustness in the presence of noise and have limited the usable signal bandwidths and sampling rates. This dissertation develops a data compression scheme that produces high CR, robustly handles noisy signals, readily lends itself to real time implementation, and reproduces the clinically significant features of the data with high fidelity, while allowing higher bandwidths and sampling frequencies to be used.; The described approach is a transform technique, using the recently introduced discrete wavelet transform (DWT). The wavelet transform uses multiresolution analysis (MRA), where slowly varying components of the ECG like the P and T waves are analyzed by wavelets with larger scales, while bursty portions like the QRS complex are analyzed by wavelets with smaller scales. Segmented ECG data is analyzed by a library of orthonormal basis functions that are translated and dilated versions of the Daubechies prototype wavelets. The expansion of the signal in these basis functions is performed by the DWT which is implemented using conjugate quadrature mirror filters (CQMF). Due to the time-frequency localization property of the wavelets, the energy of the input signal is efficiently packed into a small set of concentrated wavelet coefficients. The best basis for the given signal segment is chosen as the one whose expansion coefficients have the lowest total entropy. Wavelet coefficients are chosen from the best basis to either give a fixed CR or fixed percentage r.m.s. distortion (PRD), the latter technique a novelty, that is achieved by establishing a relationship between the cumulative energy of the chosen coefficients and the energy in the current input segment. The quantized and encoded wavelet coefficient sequence is stored and the original signal can be reconstructed by the inverse DWT which is implemented using CQMF.; The algorithm was empirically tested on selected records from the MIT-BIH Database, a standardized database of ambulatory ECGs, and yielded average CR of about 16:1 and 13:1 at corresponding PRD of 10% and 7.5%, respectively. Due to the denoising capabilities of the basis functions, distortion introduced in the reconstructed signal, as measured by the PRD, is primarily due to the noise reduction achieved by the algorithm, rather than from the distortion of clinically significant features. The reconstructed signals have been subjectively evaluated by Cardiologists, and their fidelity has been found to be adequate for ambulatory ECG analysis.