Instrumentation & Measurement Magazine 25-5 - 43

Fig. 1. The network structure of the proposed ECG signal quality classification model. (a) The network structure. (b) The network structure of the RRM, where the
modules represent channel squeeze part, recurrent module, residual module, 1D convolution module and channel concatenate module, respectively.
designed model structure can be neither too shallow nor too
deep. In this work, we designed one combined model structure
(Fig. 1). ECG signals after preprocessing were input to
the model directly and were compressed to 1/2 of their original
size through a maxpooling layer. Then, three residual
recurrent modules (RRM) were used to extract features of
the input signal and these features were classified into three
classes by the softmax layer.
As shown in Fig. 1b, the RRM was composed of five main
parts: channel squeeze module, recurrent module, residual
module, 1D convolution module and channel concatenate
module. The convolution module contains one 1D convolution
layer, one batch-normalization layer, and an activity
layer with Relu activity layer. For each RRM, the channel of
the feature maps was first squeezed to 1/3 of the input channel
through a 1D convolution layer with a 1x1 kernel in the
channel squeeze part. Subsequently, the squeezed feature map
was transferred to three different modules: recurrent module,
residual module, and 1D convolution module. There were
multiple paths from the input layer to the output layer of the
recurrent module; the existence of longer paths makes it possible
for the module to learn highly complex features, while
the existence of shorter paths may help gradient backpropagation
during training [40]. The residual module was a shortcut
connection structure that enabled the model to extract deeper
features while avoiding gradient explosion and gradient disappearance
problems [41]. Three 1D convolution layers with
different kernel sizes (1x3, 1x5 and 1x9) were used to extract
feature of different receptive field. The feature maps extracted
August 2022
by the three modules were concatenated as the output of the
RRM in the channel concatenate module. Thus, the designed
RRM effectively extracted features of different receptive fields
and different depths, while avoiding gradient disappearance
or gradient explosion.
The model used a cross entropy loss function and was
trained by the stochastic gradient descent (SGD) optimization
method. To prevent overfitting, the dropout layer was added
after each 1D convolution layer. Due to the use of RRM, the
model could use fewer training parameters to achieve deeper
depth. The proposed model had a total of 60 convolutional layers
with only 237,291 trainable parameters.
Noisy ECG Generation Method
To generate noise with a specific SNR, we considered the training
data with good signal quality as the original signal without
noise. The SNR is usually calculated as the power ratio of the
original signal to the noise, which is calculated as follows:
SNR 10log 10

P
 


P
signal
noise
(1)
Assuming that the ECG signal with good signal quality was
a discrete matrix Ssignal
= [x1,x2,......,xn], its signal power was:
PSsignal

IEEE Instrumentation & Measurement Magazine
 S *S signal
length Ssignal
'
signal


(2)
43

Instrumentation & Measurement Magazine 25-5

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