Instrumentation & Measurement Magazine 25-5 - 46

Table 2 - Test accuracy of the proposed model for
each patient with three categories division method
Patient
number
01
03
05
07
09
Accuracy
(%)
92.25
92.89
86.28
94.34
92.65
Patient
number
02
04
06
08
10
Accuracy
(%)
87.93
96.46
89.93
94.88
95.51
The model was tested on the CSPC 2020 database, and the confusion
matrix of the test result is shown in Fig. 3.
The two-category signal quality classification method (acceptable
and unacceptable) was adopted, which means that
both good-quality and medium-quality ECG segments were
regard as acceptable while the poor-quality ECG segments
were regarded as unacceptable. The accuracy of the two-category
signal quality classification division method was 98.72%.
The recognition sensitivity and precision for acceptable ECG
segments were 99.73% and 98.89%, respectively. The recognition
sensitivity and the precision for unacceptable ECG
segments were 86.90% and 96.50%. The F1-score of the twocategory
signal quality classification division method was
95.40%.
If the three-category signal quality classification method
were adopted, the accuracy of the proposed model on the
87,888 test segments was 92.31%. The recognition sensitivity,
specificity, and the precision for ECG segments with good
quality were 97.00%, 88.20% and 91.13%. The classification
sensitivity, specificity, and the precision for ECG segments
with medium quality were 86.33%, 96.54% and 93.49%. The
classification sensitivity, specificity, and the precision for ECG
segments with poor quality were 86.99%, 99.73% and 96.50%.
Table 2 lists the classification accuracy of each patient's ECG
data, where the accuracy range was 86.28% to 96.46%.
Noise Stress Test
To test if there were severe overfitting in the proposed model
and determine the sensitivity of the model to different noises,
we designed a noise stress test. Noise with different SNR
was added to the training data, and the classification results
predicted by the test model were recorded. A total of 2,000
segments of good quality were selected from the training database.
The SNR was gradually changed from 30 dB to -10 dB by
1 dB each time. Three different noises: white noise, electrode
motion artifact, and muscle artifact, were selected for the noise
test. The electrode motion artifact and muscle artifact were extracted
from the MIT-BIH Noise Stress Test Database [42]. The
frequency of muscle artifact was 1-500 Hz, and the white noise
existed in the whole frequency band. The electrode motion artifact
occurred mainly in the range of 1-10 Hz overlapping the
ECG spectrum [38].
White Noise Stress Test: Fig. 4a shows the classification results
of the white noise stress test. For the classification result of each
SNR, the red, blue, and black bars represent the proportions of
good, medium, and poor signal qualities determined by the
proposed model, respectively. The proportions of good signal
Fig. 4. Test results of the noise stress test. (a) White noise; (b) Muscle artifact; and (c) Electrode motion artifact. The first row in the figure represents the ECG
segments with different SNR noise, and the second row represents the percentage of different predict categories of different SNR.
46
IEEE Instrumentation & Measurement Magazine
August 2022

Instrumentation & Measurement Magazine 25-5

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