2004 LEEE International Workshop on Biomedlcal Circuits & Systems
A HIGH PERFORMANCE COMPRESSION ALGORITHM FOR ECG
WITH IRREGULAR PERIODS
Hsiao-Hsuan Chou’,
Ying-Jui
Chen
’,
Yu-Chien
Shiuu’,
’,
and Te-Son Kuo
1 Department of Electrical Engineering, NTU, Taipei, Taiwan, R.O.C. 2 MIT, Cambridge, MA 02139, USA
3 Department of Nuclear Medicine,
Far
Eastern Memorial Hospital, Panchiao, Taipei, Taiwan, R.O.C. 4 Institute of Biomedicine Engineering, NTU, Taipei, Taiwan, R.O.C.ABSTRACT
Electrocardiogram (ECG) signals have both intra- and inter-beat correlations, which can be exploited for compression by arranging the ECG signals into appropriate two-dimensional (2-D) representations. In this paper, we propose a novel approach that maps 1-D ECG
signals to 2-D arrays effectively and then compresses the 2-D arrays with eMcient image compression algorithms. Compared to existing 2-D ECG compression methods, the proposed algorithm is unique in that it reveals much more intra- and inter-beat correlation characteristic of ECG
signals, Therefore, the image Compression algorithms can achieve enhanced performance, Furtbermore, unlike existing 2-D ECG compression methods, the proposed .
algorithm works well f o r both regular and irregular ECG
signals with extremely varying periods. In particular, its performance is insensitive to QRS miss detection cases.
1. INTRODUCTION
Modern
ECG monitoring devices generate vast &outsof data and require huge storage capacity. In order to process, transmit, and store the data efficiently, many
ECG compression methods were proposed and could be
classified into three major categories [l]:
1) Parameter extraction techniques: such as prediction [2] and vector quantization methods 2) Transform-domain techniques: such as 2-D
discrete cosine transform @CT) [l], singular value decomposition
(SVD)
[4], and wavelet transforms [5][6][7].3) Direct time-domain techniques: such as
AZTEC
[PI,
Scan-along polygonal approximation(SAPA)
[9], and fan algorithm [lo]. 131.
Among tbe categories listed -above, most of tbe inethods adopt 1-D representations-for I ~ D
ECG
signals. However, since theECG
signals have‘ both sample-to-sample (intra-beat) , and beat-to-beat (inter-beat) correlations, some 2-D compression approaches have been proposed for better compression performances, For example, Lee [I]used
“cut anddim
beats approach.and 2-D discrete cosine transform’’ to getpretty good compression results in regular
ECG.
Wei [4]used
truncated singulac value decomposition algorithm to compress 2-DECG
arrays. Recently, wavelets arewidely used for both 1-D and 2-D ECG compression [5][6][7]. Their results are summarized in Table 1. Most of the papers showed good ECG compression performances for regular ECG cases. However, their compression performance dropped in irregular
ECG
signals. In order for the 2-D ECG compression algorithms to accommodate irregular ECG signals,
we
propose below a novel procedure that converts (1-D) ECG signals to easier-to-compress 2-D representations. As for the codec, PEG2000 is chosen because it is the latest international standard for static image compression with better performance than others [11 J and it bas been adopted to compress ECG with high efficiency [ 121 [ 131. The following section will describe the 1-D to 2-D process in detail including the QRS detection and alignment, length limitation, period sorting, and mean extension steps.2. NOVEL 2-D ECG PROCESS
Before describing the algorithm, two measures are fmt introduced to compare the compression performance
with other algorithms: the compression ratio (CR) and the percent root mean squared difference (I”). The CR is calculated as the number of bits in the original
1-D
ECG
signal over the number of bits in the compressed signal, and PRD is given by (I)where
(’
is the original signal of length N, and(’
is the reconstructed si@.2.1. QRS detection and alignment
To map 1-D ECG signals to 2-D arrays, the peaks of QRS waves should be detected first to identify each heartbeat. Many
QRS
detection algorithms [14] have been proposed and we choose “A simple real-timeQRS
detection algorithm” [lS] for its high detection accuracy (99.5%). Others with high enough detection accuracy would also work. After each
QRS
peak of heartbeatsegments is identified, the originax 1-D
ECG
signals are cut at every 130th sample before each QRS peak (a 0-7S03-8665-5/04/$20.00 02004 JEEE130-sample shiR fiom the QRS peak, a smooth region estimated fiom human physiology [16]). Note that we choose not to use the
QRS
peaks to &Limit each heartbeat segment because we want to avoid large boundary values, which tend to result in discontinuities and make the 2-D arrays hard to compress.2.2. Length limitation
Because the periods of hearfbeats are not all the same, the lengths of the heartbeat segments
can
be different. This is more severe if the ECG is abnormal or theQRS
detection step misses some QRS peaks such that the identified “heartbeat” is very long, resulting in a 2-D m y with huge size. To deal with this issue, we need to
limit the length of the detected heartbeats to form a 2-D ECG representation of reasonable sizes. We choose the Iength limit to be 512 pixels. Heartbeats with longer periods than the length limit will be cut into two or more segments. If the length limit is not set to 512 but 1024 or some larger value, i.e., the 2-D arrays are extended to a larger size by more dummy values such as zero or the mean value o f the arrays, this
w
i
l
l
not affect the CR and PRD results because the PEG2000 codec can process dummy values very well. The purpose of the length limit step is just to avoid unbounded array sizes ifQRS
cannot be detected correctly for a long time.23. Period sorting
The 2-D array resulting fiom QRS detection and aIignment already exhibits the inter-beat correlation of the original ECG signals. However, the period irregularity presents a challenge to 2-D compression algorithms. To exploit the inter-beat correlation and simultaneously obtain a superior performance, we propose an additional period sorting step which sorts the period of each heart beat segment. This is a novel and powerful method for irregular ECG compression because it reduces the period differences among the adjacent heartbeats effectively, resulting in improved CR and PRD.
Existing 2-D ECG compression methods [1][4][7] show good results for normal ECG signals but not for irregular ones with extremely varying periods. This implies that these algorithms have poorer PRD on irregular
ECG
signals. However, irregular signals are more significant for clinic diagnosis than normal ones. It is very desirable if a compression algorithm can process abnormal ECG signals very well. This effective period sorting step addresses this issue and better exposes the2-D correlation structure to be exploited by the following extension method described in the next section to form a 2-D array.
2.4. Length equskation with mean extension
segment by repeating its last sample, and mean extensions which pads short segments with the mean of the whole signals of the last samples of heartbeat segments. In summary, this step equalizes the length of each heartbeat segment to form a proper 2-D array. Experiments indicate that mean extension with mean of the last samples of heartbeat segments performs better than the other three methods.
3. Experiments1 results
The proposed algorithm was applied to 100, 117, 119, and 232 in MIT-BIH arrhythma database E171 with
1 1-bit resolution and 360 Hz sampling rate. Record 119 with extremely varying periods is taken for an example. First, the QRS peaks are detected, cut and aligned. Fig. 1 (a) would be the 2-D greyscale array if we were to cut the ECG at
QRS
peaks. Note the undesirable large values near the segment boundaries as pointed out earlier. By taking the proposed QRS detection and alignment method with zero extension (but no period sorting), we obtain Fig. 1 @). Then, by the proposed period sodng step, Fig. 1 (c) is formed. Lastly, by mean extensionWith
the mean of the last samples of all heartbeat segments, Fig. 1 (d) is obtained. Note that in Fig. 1, the 2-D arrays become smoother and smoother as the algorithm progresses. These 2-D representations are to be compressed by JPEG2000. The compression performances are shown in Fig. 2. It also shows that CR and PRD results are improved step by step. The comparison results of various 2-D ECG compression afgorithms are summarized in Table 1. It is clear that the p r o p s 4 algorithm performs better than others, especially in irregular ECG record 119. An example of the original signal, reconstructed signal, and mors are shown in Fig. 3. The errors in this case are small and evenly distributed. 20 6 4 0 4 6 0 3 8 0 I 100 120 loo 200 300 400
Several extension methods have been considered, including zero extension which pads the short segments with zeros, zero-order extension which extends a
1MK)
5w
0
IOW
0
Fig. 1. Step-by-step illustrations of the
2-D
greyscalearrays resulting Erom the proposed algorithm. (a) Cutting at QRS peaks followed by zero extension.
@) Cutting with 130-sample shifts *om the
QRS
peaks (c) Period sorting applied to @).(d) Period sorting applied to (b)
with
mean extension. followed by zero extension.Fig. 2. The CR and PRD pairs of the various 2-D arrays.
.
: Fig. 1 (a) *: Fig. 1 @)+:
Fig. 1 (c)A:
Fig. 1 (d) (cl Fig. 3. Record 119, CR=13, PRD=0.4%. (a) Original ECG signal.0)
ReconstructedECG
signal. (c) Difference between (a) and (b).4. Discussion
4.1. Overhead of odginal ordering and heartbeat lengths
Although the original ordering and lengths of heartbeat segments need to be recorded losslessly in the compressed data for signal reconstruction, the corresponding overhead is negligible, for example, U250 of the original data size of record 11 9 mentioned above. Thus the CR varies fiom 23.8 to 21.8, but
PRD
is improved fiom 5.0% to O.Sl%, which justifies the slight reduction in compression ratio.4.2, Sensitivity to accuracy of QRS detection
An interesting question to ask is what will happen when QRS peak is not detected correctly or when the irregular ECG signals do not have obvious QRS peaks in some periods. The following test case demonstrates how this approach performs with
QRS.
miss-detections thathappen fkequently in abnormal ECG cases. Again, record
119 is taken for an example for comparison. The QRS detection misses 30% on purpose to make a large 2-D array without length limit (Fig. 4). The compression result is summarized in Table I, too. Its CR and PRD pair obtained by this algorithm are 22.7 and 1.18% respectively, still much better than the other algorithms even with 30% QRS miss-detections.
Table 1. PRD results of various 2-D
ECG
compression algorithms.Algorithm
A. Bilgh et. Al[13] Proposed Algorithm
Wei et. AI [4] A. Bilgin et.
AI
1131 Proposed AlgorithmLee et. A1 [I] Ali Bilgh et. AI [13]
Proposed Algorithm Proposed Algorithm Missdetections with 30% QRS S2:4 1 1 119 ~ 121.6:l
I
1.29I
-
I
P
8
8 1 1decomposition”, TEEE Trans. Biomed Eng., vol. 5, pp.
290-299,2001.
[5] B. A. Rajoub, “An efficieht coding algorithm for the
compression of ECG signals using the wavelet transform”,
IEEE Trans. Biomed. Eng., vol. 49, pp. 355362,2002.
[6] M. L. Hilton, ‘Wavelet and wavelet packet compression of
electrocardiograms”, IEEE Trans. Biomed Eng,, vol. 44, pp. 384402,1997.
[7] A. R A. Mogbddam and K. Nayebi, “A two dimensional
HW) wavelet packet approach for ECG compression”, Signal
Processing and its AppIkutions, the 6th International, Symposium, pp. 226229,2001.
[SI J. R. Cox, F. M. Noile, H. A. Fozzard, and G. C. Olover, “AZTEC: pre processing program for real-time ECG rhythm
‘analysis”, IEEE T r m . Biomed E n s , vol. 15, pp. 128-129, Fig. 4. 2-D greyscale array of I19 with 30% QRS miss
detections.
1M)o
I
200 400 6w em io00 1ZW tlDO lB00 18M Moo(C)
Fig. 5. Record 119 with 30% QRS miss detection (CR=22.7, PRD=l. 1
S%).
(a) Original ECG signal.
@) Reconstructed ECG signal. (c) Difference between (a) and (b).
5. CONCLUSION
The proposed algorithm effectively rearranges the I-D ECG signals to smooth 2-D images and makes it easy for JPEG2000 to enhance the compression performance. AU
the experimental results show how the proposed algorithm improves
PRD,
especially in irregularECG
and even in
QRS
missdetection test cases. They also show superior performance compared to other methods in the literature,REFERENCES
[I] H. Lee and K. M. Buckley, “ECG data compression using cut and align beats approach and 2-D transforms”, IfiEE Trans.
Biomed Eng., vol. 46, pp. 556-565, 1999.
[2] G. Nave and A. Cohen, “ECG compression using long-term
prediction”, IEEE Trans. Biomed. EHg., vol. 40, pp. 877-885, 1993.
[3] B. Wang and G. Yuan, ”Compression of ECG data by vector quantization”, IEEE T r m . Biomed Eng., vol. 40, pp. [4] J. I. Wei, C. J. Cbang, N. K Chou, and G. J. Jan, “ECG
data compression using truncated singular value
23-26, 1997.
1968.
[9] M. Ishijima, S. B. Shin, G. H. Hostetter, and J. Sklansky, “Scan-along polygonal approximation for data compression of
electrocardiograms, ” IEEE Tram. Biomed Ens., vol. 30, 11,
[IO] R C. Barr, “Adaptive sampling of cardiac waveforms”, J.
Elechpcard, 21, pp, 5740,1988.
[l 11 David S. Taubman and Michael W. Marcellin, JPEG2000:
image compression finmdamentals, siandmclF. and practice,
Kluwer Academic Publishers, Boston, 2002
[I21 A. Bilgin, M. W. Marcellin, and M. I. Altbach, “Compression of electrocardiogram signals using JPEG2000”,
IEEE Tran. Consumer Electronics, vol. 49, pp.833-840, NOV. 2003.
[13] A. Bilgin, M. W. Marcellin, and M. I. Altbach, ‘Wavelet compression of ECG signals by JPEG2000”, Con$ Datu Compression, DCC2004, pp. 527 - 527, Mar. 2004.
[I41 B. U. Kohle, C. Hennig, and R OrgImeister, “The principles of software QRS detection”, E E E Biomed Eng.
Mug., pp 42-57, Jan.-Feb. 2002.
[ 151 1. Lee, IC. Jeong, J. Yoon, and M. Lee, “A simple real-time
QRS detection algorithm”, IEEE h o c . Biomed. Eng., vol. 4, pp. 1396-1398,1996.
[I61 J. Vander, J. H. Sherman, and D, S. Luciano, Humun Physiology3 McGraw-Hall, 6” ed, chap 14, pp 393472,1994,
[LA
“MIT-BM Arrhythmia Database CD-ROM”, 2nd Ed.,Hmard-MIT Division of Health Sciences and Technology,
Aug. 1992. pp. 723-729, 1983.