Instrumentation & Measurement Magazine 25-5 - 42

Table 1 - The performance of deep learning models in previous literature studies
Author
Zhou [32]
Zhang [35]
Database
Huerta [33] CinC 2017
CinC 2011
CCDD
Zhao [34] Dynamic ECG
CinC 2011, CinC 2017 Acceptable or Unacceptable
Acceptable or Unacceptable
Acceptable or Unacceptable
Acceptable or Unacceptable
Good, Medium, or Poor
utilized to evaluate the signal quality through the extracted
feature sequence. However, the subjectivity of the selected feature
and the redundant information between these features
influenced the performance of the machine learning-based
ECG signal quality classification method [27].
With the development of Internet of Things (IOT) technology,
newly developed ECG devices usually collect ECG
through wearable devices and analyze them through cloud
servers [8], [28], [29]. Deep learning networks can extract high
level features adaptively, thereby avoiding the selection of
manual features, and have achieved excellent results in a series
of classification tasks such as medical image analysis [30] and
cardiovascular diseases detection [31]. Some deep learningbased
networks have also been developed for signal quality
assessment. Zhou et al. used a 1D convolutional neural network
(CNN) to identify low-quality ECG signals [32]. Huerta
et al. transformed ECG into scalogram by continuous wavelet
transform (CWT) and proposed a convolution neural network
to classify ECG signal quality [33]. Zhao et al. proved that the
modified frequency slice wavelet transform-based (MFSWT)
method can get better performance than the CWT based transform
method [34]. Zhang et al. developed an effective 7-layer
Long Short-Term Memory neural network for ECG signal
quality classification and obtained satisfactory classification
results [35]. These deep learning-based methods demonstrate
excellent performance in ECG quality assessment, which
means that the deep learning-based model is a feasible method
for evaluating signal quality.
However, as shown in Table 1, although these deep learning-based
methods show good performance, these models are
mainly developed or tested on a small amount of ECG segments.
The accuracy of the model proposed by Zhang [35]
decreases 5.20% when the test data changed to a larger test
database. Thus, whether these models can be applied to an
ECG analysis system requires further verification on a larger
test data. For signal quality categories, the two most common
methods were to either: classify signals into acceptable or unacceptable
categories [32], [33]; or divide the signals into three
categories good, medium, or poor quality [34], [36]. The former
division method was designed for picking out the noise
segments in the ECG, while the latter division method aimed
at further detailed division of signal quality [34], enabling the
analysis system to pick out ECG data with high signal quality
and make accurate diagnostic decisions [3].
42
Signal quality category ECG length (s)
10
10
10
10
10
Number of test
data
1,455
944
1,000
2,900
200
Accuracy (%)
94.30
91.20
97.20
92.40
86.30
Contribution of This Paper
To overcome the inability of deep learning-based model to
maintain accuracy on large-scale ECG data, we propose a
deep learning-based signal quality classification model and
apply it to dynamic monitoring. The proposed model combines
the residual and the recurrent modules. The residual
module is utilized to avoid the gradient disappearance. The
recurrent module can increase the depth of the model by sharing
weights between convolution layers. Therefore, the model
can extract fine-level feature representations with fewer training
parameters.
The proposed model was trained on dynamic ECG segments.
The three categories division method of ECG signal
quality was utilized. The dynamic ECG training data was collected
from 20 patients with different arrhythmia by wearable
device, and their signal quality was manually labeled into
one of three categories. To test whether the model can be applied
to 24 hours of continuous monitoring, ECG recordings
in the China Physiological Signal Challenge 2020 (CPSC 2020)
[37] were selected as the test database. The proposed model
showed a good performance on the test database, indicating
that the designed model can be used for evaluating the quality
of dynamic ECG signals.
Methods
Pre-processing Operations
The wearable ECG signals collected by the wearable ECG
monitor device contained baseline drift; thus, a 3-order Butterworth
high-pass filter was utilized to remove baseline drift
with a cutoff frequency of 0.5 Hz [38]. To avoid the influence
of the amplitude difference between different signals, these
signals were normalized by a zero-centered normalization
method.
Network Structure and Optimization Method
ECG signal quality classification is a challenging task due to
the individual signal differences [39] and the variety of noise
types. The overfitting and underfitting were main problems
in the deep learning network tasks. Deeper network structures
mean the network can learn deep features with better
training performance on the training data but can lead to
overfitting, while shallow network structures lead to underfitting.
To reach the best classification capability, the
IEEE Instrumentation & Measurement Magazine
August 2022
Instrumentation & Measurement Magazine 25-5

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