Instrumentation & Measurement Magazine 25-5 - 49

Table 4 - The performance of different signal quality assessment models
Author
Zhou [32]
Zhang [35]
Proposed
Database
Huerta [33] CinC 2017
CinC 2011
CCDD
Signal quality category
Method
CinC 2011, 2017 Acceptable or Unacceptable 1DCNN
Acceptable or Unacceptable CWT+CNN
Acceptable or Unacceptable
Acceptable or Unacceptable
LSTM
CPSC 2020
Acceptable or Unacceptable RRM
Redmond [36] dynamic ECG Good, Medium, or Poor
Zhao [34]
Proposed
dynamic ECG Good, Medium, or Poor
Good, Medium, or Poor
CPSC 2020
to extract the overfitting features. Therefore, the model does
not have serious overfitting.
Three different noises were utilized in the noise test. The
model shows better sensitivity to the white noise and muscle
artifact. The distribution ranges of white noise and muscle
artifact interference are relatively similar, and both belong to
high-frequency noiseand are manifested as a visible change of
signal amplitude in a short period. Therefore, the model more
easily identifies the waveform changes in the segments to accurately
classify their signal quality. Compared with the other
two noises, the model is less sensitive to electrode movement
artifacts. This is mainly because the electrode motion artifacts
are mainly distributed in the range of 1-10 Hz [38], resulting in
small short-term waveform amplitude fluctuations. The classification
model failed to respond to its slight amplitude change
in a short period when the SNR was high. Moreover, the frequency
band of electrode motion artifacts overlaps with the
ECG signal. Last but not least, all of the ECG segments in the
training data are allowed to contain baseline drift noise, and
even ECG segments with good signal quality are also allowed
to contain slight baseline drift, which may also cause the model
to be insensitive to low-frequency noise.
Model's Classification of Dynamic ECGs
In this paper, we proposed one deep learning-based ECG
signal quality classification model for dynamic ECG. The proposed
model had an average validation accuracy of 90.12% by
the 10-fold cross validation and an accuracy of 92.31% on the
24-hour dynamic ECG recordings in CSPC 2020. As listed in
Table 4, compared with the deep learning models proposed
by Zhou [32], Huerta [33] and Zhang [35], the proposed model
uses a larger amount of test data and achieves the best classification
accuracy with the two-category signal quality division
methods (acceptable and unacceptable). Although there was
a decrease of accuracy when the three-category signal quality
division method was adopted, the proposed model also
showed a better performance than the feature-based method
proposed by Redmond [36] and the deep learning-based classification
model proposed by Zhao [34]. Compared with deep
learning methods that use ECG segments as input directly,
August 2022
Time-domain
features
ECG length
(s)
10
10
10
10
10
-
MFSWT+CNN 10
RRM
10
Number of
test data
1,455
944
1,000
2,900
87,888
300
200
87,888
Accuracy
(%)
94.30
91.20
97.20
92.40
98.72
78.70
86.30
92.31
models using time-frequency spectrogram as input [33], [34],
usually require much more pre-processing time to generate the
spectrogram of the ECG segments. The 1D CNN models [32]
with a single receptive field may not be sensitive to the noise of
a specific frequency and its multiples.
As shown in Fig. 3, there was a better accuracy on the test
database than the validation result, which was mainly caused
by the imbalance between different signal quality categories
in the test database. The model shows a good performance on
picking out the good- and poor-quality segments. There were
48,811 of 87,888 segments with good signal quality in the test
database while the amount of the ECG segments between different
categories of the validation data were the same. Thus,
the test accuracy was better than the validation result. The proposed
model also showed a much better performance when
the signal quality was classified to two categories. The specificity
of both the acceptable and unacceptable categories were
more than 96%, and the sensitivity of the acceptable category
was more than 99%. This means the proposed model can effectively
pick out the acceptable segments, and the segments
predicted as poor signal quality by the proposed model were
basically signals with poor quality. Table 2 shows that the
model has certain deviations in the classification accuracy of
ECGs from different patients. This is mainly due to the differences
in the overall signal quality of patients, leading to
differences in the accuracy of the final classification. However,
the minimum accuracy rate exceeds 86%. Due to the imbalance
between the different signal quality groups, we also used
F1-score for further evaluation. The F1-score of the proposed
model on CPSC 2020 was 91.74% with three-category signal
quality classification method and 95.40% with the two-category
signal quality classification method. This means that the
accuracy of the model is indeed affected by the imbalance of
data. However, the F1-score of the proposed model also indicates
that the model can be a viable signal quality assessment
method for dynamic ECGs.
As shown in Fig. 5 and Fig. 6, it can be concluded that the
signal quality results of the proposed model were in good
agreement with the artificially marked tags when tested on 24hour
wearable ECGs. Segments with arrhythmia, such as PVC,
IEEE Instrumentation & Measurement Magazine
49
Instrumentation & Measurement Magazine 25-5

Table of Contents for the Digital Edition of Instrumentation & Measurement Magazine 25-5
