A novel ECG signal compression method using spindle convolutional auto-encoder

In this paper, a novel family of compression algorithms is presented, which is designed to exploit the redundancy of one-dimensional (1-D) semiperiodical biomedical signals resulting from the cyclic nature of the underlying physical process. The basic idea is that a pool of past-seen cycles is maintained and cycles to be encoded can be stored as transformed versions of those residing in the pool. Conceptually, this approach is an extension of dictionary-based coding schemes used for text compression to signal patterns residing in an n-dimensional space. A cycle transformation method is introduced in order to render the pattern matching process practical and to enable cycle substitution. Based on the principles of the algorithmic family and this transformation method, an electrocardiogram (ECG) oriented algorithm is implemented and thoroughly tested. The performance of this implementation is examined theoretically and deductions about the optimal algorithm settings are made. The ECG compression algorithm is superior to the average beat subtraction algorithm as proposed by Hamilton and Tompkins (1991) in cases where high compression ratios are required.

A novel family of compression algorithms for ECG and other semiperiodical, one-dimensional, biomedical signals

This thesis seeks methods for minimal linear representation and subsequently low rate sampling of
electrocardiogram (ECG) signals. ECG, a non-invasive approach to record heart's electrical activity,
has been an ubiquitous tool for preliminary as well as complicated diagnoses of heart related issues.
The modern lifestyle of ever increasing population has elevated the rate of heart diseases. Many a
times, periodic monitoring of ECG, such as holter monitors, becomes imperative for diagnosis and
curing of heart conditions. Some of the major issues in maintaining quality of healthcare services
are low doctor to patient ratio in urban as well as resource constrained rural localities, unavailability
of trained medical professionals in remote areas, infrastructural constraints etc. In this backdrop,
personalized and mobile healthcare, such as telecardiology has been proposed.
In order to realize a resource friendly telecardiology system, several engineering aspects need
attention. This thesis focuses on a few related signal processing issues. Specifically, compact representation
and low rate sampling of ECG signals, subject to certain representation/ reconstruction
accuracy are discussed. It is observed that `sym4' and `db4' wavelets pack the energy of various
ECG signals in least number of coeficients. Further, the proposed hybrid Fourier/ wavelet method
is shown to offer even sparser representation by using Fourier approximation for the low frequency
component and wavelet approximation for the remaining part of the signals. The former contains
most of the signal energy whereas the latter accounts for key clinical information at feature points.
Next, sparsity of ECG signals is exploited to demonstrate near universality of the proposed nonuniform
sampling scheme. Recent advances in compressive sensing (CS) theory have facilitated recovery
from samples acquired in a nonuniform manner.
The evaluation of proposed methods is based on empirical studies on large ECG datasets available
publicly. This is justified as proposing a statistical model for ECG signals is difficult on account of
wide variety of such signals. Objective quality measures are used to judge the performance.

Minimal Representation of Electrocardiogram Signals: Towards Low-Cost Telecardiology

The method presented here is, conceptually, an extension of dictionary-based text coding to the semi-periodical signal domain. A pool of cycles is used as the dictionary in this case. If a pool cycle matches a signal cycle within a given error criteria, it replaces it in the reconstructed signal. If a satisfactory match cannot be achieved, the signal cycle is placed as is in the compressor's output. In order to be able to replace a signal cycle with one that occurred in the past the authors introduce a transformation method that enables them to preserve a number of preselected measurable features in the reconstructed signal. In addition to the above, the method presented exhibits some unique characteristics, unusual waveforms in the input signal (like in arrhythmia) are preserved exactly, and the reconstructed signal always 'looks' like the original, i.e. no staircase effect is present. The method is suitable for both lossless and lossy compression. >

http://ieeexplore.ieee.org/iel2/3299/9921/00470114.pdf

A novel cycle-pool based compression method for 1-dimensional semiperiodical biomedical signals [ECG/blood pressure application]

The method presented in this paper is, conceptually, an extension of dictionaly-based text coding to the semiperiodical signal domain. A pool of cycles is used as the dictionary in this case. If a pool cycle matches a signal cycle within a given error criteria, it replaces it in the reconsmcted signal. If a sati.$actory match cannot be achieved. the signal cycle is placed as is in the compressor's output. In order io be able to replace a signal cycle with one that occurred in the past we introduce a transformation method that enables us to preserve a number of preselected measurable features in the reconstructed signal. In addition to the above, the method presented, exhibits some unique characteristics: unusual waveforms in the input signal ( like in arrhythmia ) are preserved exactly, and the reconsmcted signal always 'looks' like the original. i.e. no staircase effect is present. The method is suitable for both lossless and lossy compression.

A Novel Cycle-pool Based Compression Method for 1-dimensional Semi- periodical Biomedical Signals

Proposes a new adaptive method of data compression for digital ambulatory electrocardiograms (ECGs), considering the diagnostic significance of each segment of the ECG. The R-wave is detected, followed by multi-template matching of the detected beat and judgment of the noise level; the templates are successively created during processing. The residual signal (the difference between the original ECG and the best-fit template) is approximated with the FAN data compression method SAPA2 (Scan-Along Polygonal Approximation) and then encoded. The error threshold of FAN is decreased during the P-wave segments and increased during the noise segments; the maximum error of the reconstructed signal at each time is known. This method is applied to ECGs of the AHA (American Heart Association) database and its usefulness is indicated; e.g. the bit rate is approximately 400 bps at 8% PRD (percent RMS difference) and 200 bps at 15% PRD. >

S. Tanaka

Papers

Adaptive data compression of ambulatory ECG using multi templates