A Deep Learning Based Approach for Detecting Fetal Stress Using CTG Signal

The modern society is also facing the ever-growing demands of health needs and the ever-growing standards of living, and the same can be said of the future generations. The population of individuals who would desire quality services for their children during pregnancy and as delivery babies would increase. To facilitate healthy growth of the foetus and detection of potential problems at their earlier stages of development, parents will need to have real-time physiological records of the foetus, its heart rate, and the intensity of its contractions during the development [1]. Psychological or physiological stress could cause significant variance in the physiological parameters of some hospitalized patients that are not present in their natural conditions.

Recent research has noted that out of every three individuals who are provided with home blood pressure equipment, their measurements are expected to be in accordance with their clinical diagnosis, particularly in individuals who already have high blood pressure and those who are at risk of developing heart disease conditions [2]. The fact that physiological indicators of blood pressure, blood oxygen, and EKG of pregnant women and foetuses were measured was important due to the fact that Professor Johnson, through his remote fetal monitoring, had established that the determination of the physiological indicators was essential. The techniques would also be applicable in determining the position of the baby adequately and providing a comfortable space for the mother. Therefore, the application of remote fetal monitoring technology would be applicable at home during the perinatal stage effectively to improve the quality of care during the perinatal period.

The paper will capture real-time fetal health and abnormal fetal development based on a remote fetal heart monitoring system, which involves neural networks to classify the data of fetal heart rate monitoring. The data collection is long and cumbersome, and the qualifications of professional medical specialists are required due to the professional specificity that can be offered in classifying the data on classical fetal heart rate (FHR) as the process is time-consuming. In practice, the investigation of these FHR attributes is time- and labour-intensive, and therefore it cannot be handy to provide immediate medical outcomes. This is retrogressive in medical diagnosis. However, the recently developed neural network models have the ability to identify valuable information in fetal heart rate data in the form of the corresponding algorithms automatically and generate successful classifications with the actual results. In this way, we research the problem of data prediction of fetal cardiac monitoring through neural networks.

However, the problems encountered in the use of fetal heart rate monitoring data are excessive noise, large amounts of information, and high error rates, which will eventually not result in positive outcomes of fetal heart rate classification. The issue that ought to be addressed in this study is the manner in which to preprocess data and come up with a model of a neural network that would suit the data. In this study, different neural network models of the FHR data categorisation are applied. It is the most applicable contribution to the identification of the most appropriate model of FHR data classification because it compares and contrasts the performance of a large number of models, summarises the findings, and, based on the optimal model, informs medical practice.

3.2 Pre-processing

The FHR signals are pre-processed to eliminate noise. Here, pre-processing is performed using a wavelet transform. Then the de-noised signal was smoothed by a smoothing filter. The pre-processed signals are then analysed and converted into matrices. The converted matrices are transformed to form an image. The transformed images are fed into the feature extraction block. Wavelet denoising was performed using the Daubechies (db4) wavelet at level 3 decomposition with adaptive soft-thresholding (σ = 2 log n) to remove high-frequency noise and preserve FHR morphology. After the wavelet reconstruction, the baseline drift was further stabilised by using Gaussian smoothing (kernel size = 5, 1.5). The hybrid method saved more than 95 per cent of the signal energy and raised the signal-to-noise ratio (SNR) to 28.3 dB, compared with 19.4 dB. A comparative analysis using Empirical Mode Decomposition (EMD) and the Short-Time Fourier Transform (STFT) revealed that wavelet-Gaussian preprocessing achieved the best morphology preservation, and it was therefore used in all subsequent analyses.

3.3 Feature extraction

A pre-processed data is then formed into an image to be used as a feature. Machine learning feature selection uses a first set of measured data to generate derived values or features that are supposed to be informative and non-redundant. This eases the stages of later learning and generalisation and in certain cases results in superior human interpretations. Dimensionality reduction is connected with feature selection.

The input data to an algorithm may be considered unnecessary or too large to manipulate, which may be narrowed down to a few values, or a feature vector (e.g., the precise value in feet and meters, or the repetition of the image in form of pixels). The stage of choosing the first characteristics is referred to as feature selection. The selected features are expected to contain the information that is relevant in the input data to complete the task one wants to achieve but with this reduced representation instead of using the whole original data. In this case, deep learning is applied to extract features. Two distinct deep learning methods are being used: GoogleNet and AlexNet that are both grounded on CNNs.

The GoogleNet Architecture has a total of 27 layers of pooling that form 22 levels. It has 9 linearly stacked inception modules. The inception modules have their terminus points linked to the global average pooling layer. One of the biggest contributions to neural network research that was made, especially in CNNs, was the Inception Network. Inception Networks has now been made available in three versions Inversion versions 1, 2 and 3. This original edition of the Inception network is known as GoogleNet. Overfitting can occur when there are several deep layers in constructing the network. We proposed GoogleNet design, which is a solution to this problem, provided on the basis of the idea of filters that can be of various sizes. This notion is simply extending the network and not enriching it.

The CNN-based AlexNet model was used in this work. An architecture of convolutional neural network is called AlexNet. AlexNet was made up of eight networks five convolutional layers in the first stage with max pooling layers in certain cases and three fully connected layers in the final stages. It also worked well in training as compared to tanh and sigmoid in the case of the non-saturating ReLU activation function. AlexNet is undoubtedly one of the simplest methods of learning the principles and methods of deep learning. It is not a complicated architecture, by comparison with some of the state-of-the-art CNN designs of recent times. Features were extracted using GoogleNet and AlexNet and redundant features eliminated. This extracted feature is the input to the classification model then.

3.4 Classification using machine learning and proposed enhanced VGGNet

The proposed enhanced VGGNet architecture was used in this process and the findings were compared to three machine learning algorithms namely Decision Tree, K-Nearest Neighbours (KNN) and Support Vector machine (SVM). In a higher dimensional space, SVM defines hyperplanes as defined by the set of points at which the dot product with a particular vector is a constant; this space of vectors is orthogonal (and, thus, minimal) space defining a hyperplane. The distance that is perpendicular to the closest data points to the hyperplane is called the margin (both sides). The support vectors are the points that have the least margin distance. The process of grouping is used to compute margin value. In this case, margin is applied as an operator.

The Pattern recognition KNN is a method of classification that is based on the nearest neighbours of a new sample whose classification is based on a training data set and the nearest k neighbours. In KNN classification, a data is classified according to the closest data of the data. This proximity is achieved by giving the value of K which is put on the model. In case of k = 1 it would be moved over to an nearest class to 1. The categorization performed by K-NN is categorized as a local approximation and all the computations are delayed until a time when the user intends to test the function. The same is that in this technique distance is used to cluster the available data, this is to say that when the scale of the features utilized is radically different or when the features utilized in the procedure are not of the same physical quantities the normalization of the training information can greatly enhance the quality of the method.

The decision trees depend on various methods of establishing whether or not a node should be sub-divided into two or more sub-nodes. Production of the sub- nodes increases homogeneity of the physical sub- node. That is, it can claim that the purity of the node is improving in reference to the target variable. Results of the feature selection algorithms are used in order to create ten classes. This further divides into these ten groups which are further subdivided into three; Pathological, Suspicious and Normal. The chosen feature is then feed forwarded to the improved VGGNet.

There are four layers in the architecture of the proposed skip network of VGGNet as illustrated in Figure 2. It is a layered architecture. Figure 2(a) shows the created broadened skip network of broad semantic biomedical signal training. Figure 2(b) is the description of the properties of the network which were obtained at every output stage. Additional skipming module: The amount of feature map layouts of the final stage of production once all the other multi-scale filter coefficients are done is kept at 12 although the amount of feature maps produced by producing 1 × 1 convolution will also give the same number of feature maps as the input provided the number of channels on the input channel is Cin. The overall output characteristic maps of the skip module are therefore (Cin + 12). Figure 2(c) depicts the upsampling and concatenation and convolution with a specific number of the number of output feature maps and combining them to merge the generated feature maps of the extended skip module and the decoder. The algorithm of a combination of the skip module and feature maps is depicted.

Figure 2. (a) Proposed skip network architecture, (b) Additional skip module, (c) Combined procedures for integrating feature maps

One of the several VGG Net-like encoders is made up of a series of 3 × 3 convolution, batch normalisation, and a nonlinear activation. The encoder has been configured to apply max pooling to the spatial resolution of the feature maps right after every VGG-block (except the final one). The last two convoluted blocks (big kernels) of the encoder are designed to capture a bigger context, then there is batch normalisation, nonlinear activation and two additional convoluted blocks. In order to compute these huge convolutions, separable kernels are computed to minimize the number of computations. The activation used in the network is nonlinear and is referred to as the rectified linear unit (ReLU).

Both the encoder and the decoder are equally designed but the latter has fewer feature maps which minimize compute and memory. Each block of the decoders is preceded by a repeating format that consists of upsampling, 3 × 3 deconvolution that is repeated and then batch normalisation by the nonlinear activation functions. The size of feature mappings in all the stages of the decoder is maintained constant except that the size of feature mapping is equal to the size of the target classes in the final one. To achieve the network, long skip connections of the feature maps of the encoder and the feature maps of the decoder are done. The long skip module of the form of a bank of multiscale filters is presented in Figure 2(b). As can be seen, the feature maps of the corresponding decoder are joined to the output feature maps of the extended skip module. The proposed Net is constructed of and links the constituents of three separate DNNs: a multi-scale filter bank, or incidence module; skip layers to bring out finer spatial information; and bigger convolutional kernels to introduce bigger context to the image. The work shall attempt to generalize the architecture training the such relations with multi-scale convolution and no layers of fixed identity links (copy-and-concatenate). The convolutional kernels are huge in order to make it expand the effective receptive field of the network so that it can be able to learn semantic graphics. The big kernels are however utilized in the encoder only and bypass in the skip layers.

Finally, the huge convolutional kernels of the skip layers have been criticized in Figure 2(b) and it is assembled into a multi-scale filter bank to inject the relevant context of the inputs in the entire learning process without fine-tuning to select the specific size of the kernel. This is not an effective learning method compared to the numerous scales of opportunities that this module provides because the only scale of the environment one can choose.

In order to train the DNN networks in both tasks, the loss should be the cross-entropy loss as follows:

$\begin{aligned} & N \quad L \\ & \mathrm{CE}_{\text {los }}=-1 / \mathrm{N} \sum \sum 1 y i \in L c \log p[y i \in L c] \\ & i=1 \, c=1\end{aligned}$    (1)

N denotes the total number of signals, L the number of semantic categories and 1$y i \in L c$ is a binary indicator function.

Where category c is the ground truth label of the ith observation, p[$y i \in L c$] is the predicted probability of the model of the classification. Both datasets are evaluated against the trained model using the intersection over union (IoU) metric.

$I o U=\frac{T P}{T U P}$      (2)

where, P is the anticipated category and T is the target.

The algorithms for the skip layer and the integrating module are given below. After the classification, the performance of each method is analysed.

The Proposed Enhanced-VGGNet architecture assumes the stages of the structural modifications that distinguish it against the original VGG-16/19 systems. Firstly, the standard sequential convolutional stack has been substituted with multi-scale convolutional blocks with parallel 1 × 1, 3 × 3, and 5 × 5 kernels. This allows the network to discover both local and global contextual information in the CTG signal, which is comparable to the Inception module but modified to one-dimensional biomedical time-series data. Second, an encoder/decoder connection is performed with long skip connections between the encoder and decoder layers, and enables the incorporation of low-level and high-level features and mitigates the issue of gradient vanishing. Thirdly, the deeper blocks have been augmented with deeply separable convolutions to radically reduce the number of trainable parameters and the cost of computation and improve the efficiency of inference. Other aspects that qualify to be redundant include dropout layers and batch normalisation, which can effectively boost convergence stability and decrease overfitting. Together, these design innovations make the product lightweight but expressive that may sufficiently detect fetal stress. A change alone yields a 1220 per cent increase in validation accuracy, and the Enhanced-VGGNet is completely accurate, 96.6 per cent, compared to the base VGG and the hybrid CNN model with identical training. (Figure 2).

This paper introduced an automated fetus stress monitor. Women who have long-term stress and those with diabetes mellitus or who have high blood pressure, are more prone to complications during pregnancy and labor. These factors contribute significantly to elevated preterm births and death of babies. The current fetal monitoring systems have numerous pitfalls. The use of a simple wearable automated system, that would aid in constant monitoring of the fetal vitalities will also be suggested, and the healthcare professionals would be able to treat the infants in a seamless manner. The methods of FHR extraction and classification have introduced an organized method of interpreting CTG. They also give correct results that allow instant detection and diagnosis of abnormalities. The suggested Enhanced VGGNet and AlexNet as the feature extractors give a sensitivity, specificity and accuracy of 95.8, 91.2 and 96.6 percent respectively that demonstrates its superiority over other techniques. To implement the Enhanced-VGGNet in the real world, it was tested on embedded software (NVIDIA Jetson Nano 4 GB). The mean latency of inference was 42 ms per CTG sample with memory access less than 2 GB, which validates the ability to perform near real-time monitoring of fetal-stress. The power usage was kept low (under 5 W) and it was possible to integrate this power into a wearable or portable device used in monitoring. This type of deployment will enable 24/7 bedside or home-based monitoring, which is consistent with tele-health and smart-maternal-care projects. The next objective is the quantized model compression to reduce latency even more in the future.

Our future work depends on changing this automated system with fully connected deep learning methodology with convolutional neural networks with reduced number of layers.

All the source codes and related pictures will be available from the corresponding author upon request. The dataset utilized in this study, including CTG signals from laboring women between 38 and 41 weeks of gestation, is not publicly available due to patient privacy and confidentiality considerations. However, upon reasonable request and with appropriate ethical approvals, access to anonymized data may be provided by the corresponding author for research purposes. Any additional datasets generated or analyzed during the current study are available from the corresponding author upon reasonable request.