04-05-2012, 04:06 PM
A Total Least Squares Approach for Data Reduction of Longterm ECG Recordings
[attachment=21301]
INTRODUCTION
Compression of digitized ECG data is essential to many
aspects of computerized electrocardiography including efficient
storage of longterm ECG's for subsequent processing
and evaluation. Existing ECG data reduction techniques
can be classified into two major categories: 1) direct data
(time domain) methods and 2) transform domain (frequency
domain) methods [5], [71, [SI, [9], [17], [IS]. Both methods
attempt to remove signal redundancies by exploiting the
correlation features embedded in an ECG data sequence.
The direct time domain data methods are generally used for
implementation of real-time rhythm analysis systems. The
most popular of these include the amplitude zone time epoch
coding (AZTEC) [I] and the turning point (TP) reduction
algorithm [21.
Illustralion of the procedure on Real ECG
Initially, we tested the sensitivity of the procedure to
correlated and uncorrelated noise. Likewise, the effect of
signal enhancement was assessed for data reduction. First we
used an ECG signal with the presence of correlated noise.
The parameters of the pole-zero model were first determined
without signal enhancement, by solving equation (8) using
the noisy data. It was observed that a straight forward TLS
solution required an extremely large model order (prediction
size) for representing the relevant morphological features in
the ECG, due to the presence of correlated noise. The prediction
size was decided by the Akaike information criterion
(AIC) applied to the DCT coefficients of the ECG strip. To be
on the safe side, the prediction size was chosen to correspond
to the region where the AIC measure became sufficiently flat.
[attachment=21301]
INTRODUCTION
Compression of digitized ECG data is essential to many
aspects of computerized electrocardiography including efficient
storage of longterm ECG's for subsequent processing
and evaluation. Existing ECG data reduction techniques
can be classified into two major categories: 1) direct data
(time domain) methods and 2) transform domain (frequency
domain) methods [5], [71, [SI, [9], [17], [IS]. Both methods
attempt to remove signal redundancies by exploiting the
correlation features embedded in an ECG data sequence.
The direct time domain data methods are generally used for
implementation of real-time rhythm analysis systems. The
most popular of these include the amplitude zone time epoch
coding (AZTEC) [I] and the turning point (TP) reduction
algorithm [21.
Illustralion of the procedure on Real ECG
Initially, we tested the sensitivity of the procedure to
correlated and uncorrelated noise. Likewise, the effect of
signal enhancement was assessed for data reduction. First we
used an ECG signal with the presence of correlated noise.
The parameters of the pole-zero model were first determined
without signal enhancement, by solving equation (8) using
the noisy data. It was observed that a straight forward TLS
solution required an extremely large model order (prediction
size) for representing the relevant morphological features in
the ECG, due to the presence of correlated noise. The prediction
size was decided by the Akaike information criterion
(AIC) applied to the DCT coefficients of the ECG strip. To be
on the safe side, the prediction size was chosen to correspond
to the region where the AIC measure became sufficiently flat.