2.5 Evaluation of Distortion Measures
2.5.5 Comparison of the Distortion Measures
To test the performance of the distortion measures under filtered signal (background noise removed) con- dition, the mita record 123 is chosen and the noise present in the signal is filtered. The original and the filtered signals are shown in Fig 2.10. P and R wave of the filtered signal is distorted at different distortion levels for testing purposes. The reconstructed signals with P and R wave distortions are shown shown in Fig. 2.10. These distortions are measured using different error measures employed in literature and these results are shown in Table 2.5. From the table and the figure, it can be observed that the non-diagnostic dis- tortion measures such as RMSE, PRD1, MAX, NMAX, SNR and CC provides the quantity of error. It does not provide which diagnostic features are distorted. This can be observed in WWPRD if the local errors of the WWPRD measure are provided for verifications. The results shown in the table and figure demonstrate that the distortion measure in wavelet domain provides information on distortion of the diagnostic features whereas other time domain distortion measures such as RMSE, PRD1, SNR, MAX, NMAX and NCC only quantifies the error. The WWPRD measure provides the errors between the distorted P wave/R wave and
Table 2.5: Performance of non-diagnostic and diagnostic error measures
Distortion Non-diagnostic Measure WWPRD Measure
RMSE PRD1 SNR MAX NMAX NCC A5 D5 D4 D3 D2 D1 Total
Experiment 1 : Figs. 2.10 (a), (b), (c) and (d)
P-wave 16.1 6.32 23.98 106.5 36.6 0.9981 3.5786 1.5442 0.3938 0.2848 0.1232 0.1444 6.069 R-wave 15.7 6.14 24.22 231.8 79.7 0.9981 0.7668 0.2166 1.0535 0.9826 1.6733 0.5577 5.251 with noise 15.9 6.22 24.11 57.5 19.8 0.9981 1.1291 0.5051 0.7228 0.9218 1.1415 1.693 6.113
Experiment 2 : Figs. 2.10 (a), (e), (f) and (g)
P-wave 27.3 10.70 19.40 116.8 40.1 0.9943 7.5221 1.5844 0.33 0.2416 0.103 0.1226 9.904 R-wave 27.4 10.72 19.39 432 148.5 0.9943 0.1208 0.228 0.5424 1.9927 3.1595 1.5806 7.624 with noise 28 10.97 19.18 96.7 33.2 0.9943 2.6213 0.7263 1.2061 1.5552 1.85 3.171 11.13
Experiment 2 : Figs. 2.10 (a), (h), (i) and (j)
P wave 38.8 15.19 16.36 0.126 43.5 0.9891 10.6235 1.4974 0.2525 0.1832 0.0766 0.0878 12.7209 R wave 37.4 14.66 16.67 0.692 238 0.9891 0.3464 0.3653 0.5423 1.9759 4.5813 2.2818 10.0929 with noise 38.1 14.95 16.51 0.114 39.3 0.9891 4.6187 1.133 1.4266 2.2192 2.78 4.2031 16.3806
100 200 300 400 500
−0.5 0 0.5 1 1.5 2
sample number
amplitude
100 200 300 400 500
−0.5 0 0.5 1 1.5 2
100 200 300 400 500
−0.5 0 0.5 1 1.5 2
100 200 300 400 500
−0.5 0 0.5 1 1.5 2
100 200 300 400 500
−0.5 0 0.5 1 1.5 2
100 200 300 400 500
−0.5 0 0.5 1 1.5 2
sample number
100 200 300 400 500
−0.5 0 0.5 1 1.5 2
sample number
100 200 300 400 500
−0.5 0 0.5 1 1.5 2
sample number
amplitude
100 200 300 400 500
−0.5 0 0.5 1 1.5 2
100 200 300 400 500
−0.5 0 0.5 1 1.5 2
100 200 300 400 500
−0.5 0 0.5 1 1.5 2
sample number Filtered signal: mita record 123
P wave distorted: (b), (e) & (h) R wave distorted: (c), (f) & (i) (b)
(f) (g) (e)
(c) (d)
(a) Original signal:
mita record 123
(h) (i) (j)
signal + noise: (d), (g) & (j)
Figure 2.10: Performance of the distortion measures. (a) Filtered signal of the original signal taken from mitarecord 123. (b,e,h) P-wave distorted signals. (c,f,i) R-Wave distorted signals. (d,g,j) noisy signals
the original P wave/R wave at the frequency bands of P wave and R wave, respectively.
However, the inclusion of insignificant error is more in the case of WWPRD measure. In most cases,
Table 2.6: Performance of the ECG distortion measures
Characteristics of distortion/ Objective quality or distortion measures
quality measures RMSE PRD1 PRD2 PRD3 SNR NCC MAX NMAX WWPRD
global error measure (GEM) √ √ √ √ √ √
sensitive to mean (S2M) √ √
sensitive to baseline (S2B) √
local EM (LEM) √ √
diagnostic distortion measure (DDM) √
sensitive to noise (S2N) √ √ √ √ √ √ √ √ √
the resulting WWPRD by the band D1 and D2 are high due to the presence of noise. In [190], the total error is calculated by adding all the weighted errors which are contributed by the band A5and from D5 to D1. This may mislead the judgement of the signal quality when the signal contains more noise if the noise elimination algorithm is not employed in the compression method. It can be observed that the reported WWPRD measure is more sensitive to smoothing of low level background noise. The performance of all objective distortion measures are summarized in Table 2.6.
Many wavelet based ECG compression methods are reported and the tests are carried out using the noisy records from themitadatabase and the percentage root mean square difference (PRD) criterion in the liter- ature. A major design goal of any compression method is to obtain the best clinical quality with the highest compression ratio (CR) using the optimal coding parameters such as threshold or/and quantization bit ob- tained for a quality or distortion specification. But the measurement of distortion in the compressed signal is difficult because the distortion introduced by different types of compressors are very diverse. The effect of noise filtering is one of the features using the wavelet transform for compression and it is demonstrated in various compression results reported in the literature. In this case, the magnitude of insignificant errors may not be of much relevance from the point of view of clinical quality of the compressed signal. The ef- fects of noise on the rate-distortion performance of the proposed methods and the SPIHT based methods are demonstrated in the previous chapter. Although PRD does not exactly correspond to the result of a clinical subjective test, it is easy to calculate and compare, so it is widely used in the ECG compression literature.
Common disadvantages of PRD and WWPRD criteria are that a smoothing of low-level background noises of the ECG causes a large PRD value but no clinical feature distortion and, conversely, a small average dis- tortion can severely deteriorate a signal clinical performance if all the error is concentrated in a significant feature region. The weighted diagnostic distortion (WDD) measure correlates well with subjective test but it suffers from high computational complexity, and there is no standard protocol for its implementation.
Thus, in order to introduce closed loop CR or quality control, one needs an adequate diagnostic distortion measure for the compressed signal. Moreover, the choice of which distortion measure must be used for compressed signal is of critical importance when noise suppression and signal compression is established
simultaneously. In the area of ECG signal compression, little attention has been paid towards the evaluation of distortion of clinical information. A suitable objective distortion measure can help proper evaluation of the well-designed ECG compression methods under noisy environments. Otherwise, the quality of the compressed signal has to be evaluated by subjective test, visual inspection of the clinical features. However, performing subjective test is difficult task in closed loop rate or quality control method in which the optimal coding parameters are adaptively chosen to compress time-varying signal characteristics effectively. Thus, the compression system typically involves tradeoff between the rate and quality of the output. Undoubtedly, there is a need for an objective quality measure for local and global assessment. Thus assessment of distorted signal quality is an open problem today.