1D-CNN: Classification of normal delivery and cesarean section types using cardiotocography time-series signals

3.1 Dataset

The current study utilized the CTU-UHB intrapartum CTG dataset, freely available at Physionet [32,33]. The dataset comprised 552 intrapartum CTG samples carefully obtained from singleton pregnancies without known congenital disabilities or intrauterine growth restrictions. The significant attributes of the sample and attribute distributions are depicted in Table 1. The dataset consists of 506 recordings delivered vaginally, considered normal delivery cases (control group) in the current study, whereas 46 recordings were delivered by cesarean section (case group). Each record also contains FHR in bpm and UC in mmHg time-series data with a 4 Hz sampling frequency. The recordings are 90 min at maximum and are started at most 90 min prior to delivery. The second phase of labor took up to 30 min. The FHR signal served as the input layer for the 1D-CNN model in the current work under all data balance conditions discussed in Section 5. The present study aims to classify FHR signals in normal delivery and cesarean section records. Each record’s delivery-type parameter is considered the gold standard for the classification. The FHR pre-processing used in the current study is explained in Section 3.2.

3.2 Dataset pre-processing

The movement of pregnant women and fetus, as well as incorrect sensor placement, can introduce noise in the FHR signal. The classification efficiency is highly influenced by input quality; therefore, the pre-processing phase is of utmost importance. The pre-processing used in this article is based on [24], where missing values were linearly interpolated, and missing values that lasted longer than 15 s were removed directly. If the absolute difference of two adjacent values was greater than 25 bpm, then interpolation was carried out between the initial sampling point and the first point of the next stable part. In addition, FHR values larger than 200 bpm or less than 50 bpm were considered extreme points and eliminated. The linear interpolation was used to fill up the corresponding segment.

3.3 Input normalization

The normalization method changes the features to a similar scale, which is essential to improve the classification model’s performance and training stability. This work employed the Z -score normalization method after an empirical analysis of other normalization strategies. The Z -score changes the input such that a new population with a mean of zero and a standard deviation of 1 is obtained. This method subtracted the FHR mean from each FHR value before dividing the difference by the standard deviation. The Z -score normalization approach is depicted by equation (1):

In equation (1), μ is the mean of the population, σ is the standard deviation of the population, FHR ( i ) is the input FHR value, and FHR ( i ) ′ is the normalized FHR. Due to the varying duration of labor among patients, the FHR input signals vary in length. In order to obtain records of similar length, FHR signals are padded with 0 values. Performing Z -score normalization by ignoring the zero entries in the FHR improved the overall performance of classification.

3.4 1D-CNN

Due to the unique property of combining feature mining and labeling procedures within one learning frame, 1D-CNNs have recently gained popularity in a variety of signal-processing applications. Compared to traditional multilayer perceptron (MLP) networks, it can process inputs with high efficacy. CNN has high fault tolerance and robustness and is simple to train and optimize [13].

The core CNN design contains an input layer, a convolution layer, a pooling layer, a fully linked layer, and an output layer. Figure 1 depicts the model employed in this research. A pair of convolution and max pooling layers were employed, succeeded by a layer that flattened the 1D-CNN to facilitate the creation of the fully connected MLP. The MLP comprised two dense layers consisting of 10 neurons and one neuron each, which were fully connected. Concerning binary classification, sigmoid activation was utilized in the output layer to classify the FHR signal input into normal and abnormal classes. At the convolution layer, rectified linear unit (ReLU) activation was applied. The parameter tuning was carried out for the convolution layer’s two essential parameters: the number of filters (denoted by variable n_filters) and the kernel size (denoted by variable kernel_s). The values of those parameters were determined empirically, as discussed in Section 5. The He initialization method was employed to initialize the kernel weights. The batch normalization enhanced the efficiency of the deep neural network concerning the training period and overall stability. The layer-wise parameter settings for the proposed 1D-CNN model are shown in Table 2. The detailed parameter tuning for n_filters and kernel_s used in the 1D-CNN layer is discussed in Section 5.2.

Figure 1

Architecture diagram for 1D-CNN model.

The Adam optimization algorithm was utilized with an initial learning rate of 0.005. The problem of local minima was resolved by implementing the exponential learning rate decay approach, which regulates the change in learning rate across all layers. The hyperparameter settings used during the training phase are shown in Table 3.

The equations (2)–(4) summarize the 1D-CNN model’s forward propagation operation. Below is a description of the convolution operation in the convolution layer:

where z j l is the intermediate output from l − 1 th convolutional layer, b j l is the bias for j th neuron at l th layer, x i l − 1 is the input to the convolutional layer, and w i j l − 1 is the kernel weight from i th neuron in l − 1 th layer to j th neuron in l th layer. N l − 1 specifies the total number of filters in l − 1 th layer and conv 1 D represents the 1D-CNN operation. The convolutional layer employed the ReLU activation function to convert the intermediate output z j l into its final output a j l using equation (3).

The pooling layer’s objective is to minimize the size of feature maps. During max-pooling, the largest value was selected from the area of the feature space. Equation (4) represents the output y j l due to the max-pooling layer.

The fully linked MLP layer took its input from the flattened output of the pooling layer. The error rate for backpropagation is computed using binary cross-entropy, as shown in equation (5).

where m represents the total number of training examples, y i is the actual label, and y _ pred i is the predicted output of the i ^th training example. The weights and biases were changed during backpropagation using equations (6) and (7). In this process, weight and bias derivatives were applied.

where α is the learning rate in the proposed network. The backpropagation seeks to optimize the learning process by minimizing the value of the binary cross-entropy error specified in equation (5).

5.1 Experimental setup

The proposed method was employed using Python 3.9.7, TensorFlow 2.8.0, and Keras 2.8.0. All experiments were conducted with Intel R i5, RAM: 8 GB, x64-based processor.

5.2 Experiment one: parameter tuning of 1D-CNN layer

The proposed deep learning model is a layered architecture in which the core layer is 1D-CNN that decides classification performance. The most critical parameters that affect the convolution operation in the 1D-CNN model are the kernel_s and n_filters required for the optimal performance of the model. The following experiments were conducted to decide the optimal settings for these two parameters. The base condition of the CTU-UHB dataset, without any class balancing method, was used for performance tuning experiments. The layer-wise setting and hyperparameter settings were as per Tables 2 and 3. Considering initial experimentation, the model was trained using 50 epochs and 64 training examples per batch.

5.2.1 Experiment for n_filters

The dataset consists of 506 normal delivery records and 46 cesarean section records, as explained in Section 3.1. This unbalanced condition is considered for deciding on two crucial parameters, kernel_s and n_filters. During the initial preliminary experimentation, the performance was observed to be highly dependent on the n_filters. Therefore, initial experimentation for parameter tuning is conducted for the appropriate value of n_filters. The layer-wise parameters were in accordance with Table 2, except for n_filters. The training parameters were as mentioned in Table 3. The initial setting for kernel_s was kept as kernel_s = 15, and the n_filters = [1,2,3,4,5,6] were considered for tuning. The values are selected as per initial preliminary experiments. The model was trained using the cross-fold technique with 10 folds, and the result was averaged across 10 folds. Thirty-one simulations were carried out for each choice of n_filters, and statistical analysis of Acc, Se, Sp, QI, and AUC parameters was done by plotting the results using boxplots, as illustrated in Figure 2. The topmost plots denote the Acc and Pre plots; the middle row represents the Se and Sp boxplots, whereas the QI and AUC boxplots are represented in the bottommost row. The Acc and Pre boxplots suggested that n_filters = 1 is the optimal choice, whereas the Se parameter suggested that n_filters = 5 is an excellent choice.

Figure 2

The boxplots for averaged performance parameters using different values of n _ f i l t e r = [ 1 , 2 , 3 , 4 , 5 , 6 ] across 10 folds. From top-left to top-right: Acc, Pre; from middle-left to middle-right: Se, Sp; from bottom-left to bottom-right: QI, and AUC. (a) Acc boxplot, (b) Pre boxplot, (c) Se boxplot, (d) Sp boxplot, (e) QI boxplot, and (f) AUC boxplot.

The Se parameter is helpful to avoid unnecessary cesarean sections, while the Sp parameter is significant to avoid fetal compromise. The Sp parameter boxplot suggested that n_filters = 1 will be the best option. Boxplots of QI and AUC parameters also confirmed the same value. The n_filters = 1 value also yielded promising results in the acceptable range regarding the Se parameter and, therefore, was selected for further tuning the kernel_s parameter, as explained in Section 5.2.2.

5.2.2 Experiment for kernel_s

The following experiment was conducted to determine the value of the kernel_s parameter used in the convolutional layer. The n _ f i l t e r s = 1 was considered for choosing the appropriate value of kernel_s. The various values of k e r n e l _ s = [ 5 , 10 , 15 , 20 , 25 , 30 ] were considered for experimentation. The cross-fold technique was used to train the model with 10 folds, averaging the results. A total of 31 simulations were carried out for each choice of kernel_s. The resulting boxplots of Acc, Pre, Se, Sp, QI, and AUC parameters are shown in Figure 3. The model’s performance in terms of the Se parameter was excellent for k e r n e l _ s = 10 ; similar performance was be observed for k e r n e l _ s = [ 15 , 20 , 25 ] . Although the Sp parameter’s value was low due to very few cesarean section records in the unbalanced dataset, it was better for k e r n e l _ s = 10 with the highest median value indicated by the orange line in the boxplot. The Sp parameter value was good for k e r n e l _ s = 5 , but there was a high variance from the median value to Q3, so this choice of kernel_s was not considered further. The choice of k e r n e l _ s = 10 can be further verified using boxplots of other parameters such as Acc, Pre, QI, and AUC, as illustrated in Figure 3. The optimal values selected for n _ f i l t e r s = 1 and k e r n e l _ s = 10 were used further for various dataset imbalance handling scenarios, as discussed subsequently.

Figure 3

The boxplots for averaged performance parameters using different values of k e r n e l _ s = [ 5 , 10 , 15 , 20 , 25 , 30 ] across 10 folds. From top-left to top-right: Acc, Pre; from middle-left to middle-right: Se, Sp; from bottom-left to bottom-right: QI, and AUC. (a) Acc boxplot, (b) Pre boxplot, (c) Se boxplot, (d) Sp boxplot, (e) QI boxplot, and (f) AUC boxplot.

5.3 Data balancing methods

The data imbalance in CTU-UHB regarding 506 normal delivery records (controls) and 46 cesarean section records (cases) was addressed by adopting the strategies outlined in subsections from Sections 5.3.1 to 5.3.4. The 1D-CNN model’s layer-wise settings and training hyperparameters were as per Tables 2 and 3. with n _ f i l t e r s = 1 and considering k e r n e l _ s = 10 for the 1D-CNN layer in the proposed model. A total of 31 experiments were conducted for each strategy. Statistical parameters were used to assess the model’s performance following Section 3.

5.3.1 Unbalanced dataset

The first method can be viewed as a base condition in which the dataset was assessed without class balancing. The model was trained using 50 epochs and 64 training examples per batch. The training loss function and training Acc function are depicted in Figure 4 in parts (a) and (b), respectively, indicating training stability and convergence. The model was trained using the cross-fold technique with 10 folds, as explained in Section 5.2.

Figure 4

Training functions for fold=1 (Unbalanced dataset). (a) Training logloss plot and (b) training Acc plot.

5.3.3 The weighted binary cross-entropy method

The imbalance in the dataset was addressed using the weighted binary cross-entropy approach. The weights were assigned to each class considering the proportion of 8.33% minority class vs 91.66% majority class. The weights resulted in one misclassification from the case category, contributing to as many as 11 misclassifications in the control category (considering 46 cases vs 506 controls in the dataset); the cross-fold method with 10 folds, a batch size of 64 examples. The 1D-CNN model was converged with 200 epochs during the training phase. The training logloss and Acc functions can be observed in Figure 5.

Figure 5

Training functions for fold = 1 (weighted binary cross-entropy). (a) Training logloss plot and (b) training Acc plot.

5.3.4 Oversampling using SMOTE

In the current approach, using SMOTE for oversampling, the minority class was made proportionate during the training phase. The cross-fold validation method was used with 10 folds, of which ninefolds were sent for SMOTE and then utilized for training, and onefold was preserved for testing. The model was empirically examined with different epoch values, and its performance was most remarkable for epoch 30, which is an outstanding accomplishment for the suggested model. The training logloss and Acc functions are illustrated in Figure 6.

Figure 6

Training functions for fold = 1 (SMOTE). (a) Training logloss plot and (b) training Acc plot.

6.1 Unbalanced dataset

The unbalanced dataset was used as input for the 1D-CNN model, and Table 4 shows the various statistical parameters’ values during the training and testing phases. The median Se, Sp, and QI values were 0.9980, 0.1900, and 0.4355, whereas the AUC was 0.7140. The high value of Se shows that the model accurately detected cases of normal delivery, whereas low Sp suggests that cesarean section delivery cases were inadequately classified. The unbalanced data may be a contributing factor to this bias. The model demonstrated exceptional Acc, Pre, and F1-score performance during the test phase. Figure 7 shows the individual fold and overall mean AUC values.

Figure 7

Fold-wise AUC values (unbalanced dataset).

6.2 Undersampling

The undersampling scenario considered 46 records from both classes. Table 4 lists the results for the undersampling example during the training, validation, and testing phases. This experiment produced low Se, Sp, QI, and AUC levels. The low results may be attributed to the small number of training instances used to build the model.

6.3 The weighted binary cross-entropy method

The weighted binary cross-entropy method resulted in boxplots for Se, Sp, and QI, as shown in the top row of Figure 8. The Se parameter ranged from 0.8757 (min) to 1.00 (max). The range of the Sp parameter was found to be 0.48 (min) to 0.9 (max). The low variance in the Se parameter suggests the model’s stability. The median Se, Sp, and QI are reported in Table 4 as 0.9625, 0.7650, and 0.8581, respectively. The Se value is adequate for identifying normal delivery records. The results obtained for the Sp and QI parameters are much better than those of the unbalanced dataset. Other metrics, such as Acc, Pre, and F1-Score, are also improved in the current scenario. Figure 9 provides the fold-wise AUC and mean AUC values.

Figure 8

The boxplots for averaged performance parameters using n _ f i l t e r = 1 and k e r n e l _ s = 10 across 10 folds under the weighted binary cross-entropy scenario. From top-left to top-right: Se, Sp; from bottom-left to bottom-right: QI, and AUC. (a) Se boxplot, (b) Sp boxplot, (c) QI boxplot, and (d) AUC boxplot.

Figure 9

Fold-wise AUC values (weighted binary cross-entropy).

6.4 Oversampling using SMOTE

The SMOTE scenario used in the training phase resulted in Se, Sp, and QI boxplots, as represented by the bottom row in Figure 10. The range of Se parameter was 0.9881 (min) to 1.00 (max). The low variance from median to Q3 and zero variance from Q3 to max in the Se boxplot indicate a stable and consistent result. The range of the Sp parameter was found to be 0.81 (min)–0.92 (max). Once again, a low variance from the median to the Q3 value assures the model’s stability. The median Se value of 0.9981 and the Sp value of 0.88 are well balanced enough to discriminate between normal and cesarean sections, as supported by the outstanding QI of 0.9372 and AUC of 0.9510. The median values are represented in Table 4. Figure 11 depicts the fold-based and mean AUC values.

Figure 10

The boxplots for averaged performance parameters using n _ f i l t e r = 1 and k e r n e l _ s = 10 across 10 folds under SMOTE scenario. From top-left to top-right: Se, Sp; from bottom-left to bottom-right: QI, and AUC. (a) Se boxplot, (b) Sp boxplot, (c) QI boxplot, and (d) AUC boxplot.

Figure 11

Fold-wise AUC values (SMOTE).

6.5 Comparison with existing literature

The findings of the suggested methodology are compared with the cutting-edge methods, with the CTU-UHB dataset serving as the fundamental criterion for the comparison. Table 5 summarizes the comparison’s results. The comparison uses the pertinent metrics Se, Sp, QI, and AUC. Case and control values are also mentioned since they influence the interpretation of the Se and Sp parameters. In the SMOTE oversampling situation, the Se and QI of the proposed technique are greater than those described in [17], while Sp is slightly lesser. In addition, complex engineering of features was required in [17]. Considering the scenario with weighted binary cross-entropy, the proposed approach has shown superior results than AUC [19] and [23]. Although Sp and AUC parameters in [21] are superior to the proposed method in the SMOTE scenario, Se is much better in the suggested method, and QI is almost equal to that in [21]. In addition, Saleem et al. [21] required complex feature engineering. Given the condition of an unbalanced dataset in [22] and the current work, the Se is still higher in this situation. Although [24] yielded better results regarding Sp, QI, and AUC, a few records from the CTU-UHB dataset were chosen for this study. In contrast, the present research used all available records, and Se is still performing better. The current analysis used a one-dimensional FHR signal, while a two-dimensional image generated from a one-dimensional FHR signal was used as input for the 2D-CNN method in [24]. The generation of two-dimensional images is an overhead and will require more computational power. In addition, one notable limitation when employing 2D-CNN is high computational requirements, which demand specialized hardware, especially for training. In this regard, 2D-CNN is not appropriate for applications that operate in real-time on mobile and low-power/low-memory devices [13]. All the parameters are higher in the proposed method than those in the study by Alsaggaf et al. [25] with the requirement of significant feature engineering. The segmentation approach was employed for class balancing in [27], and the current system’s performance utilizing the SMOTE scenario is superior for all assessment parameters. GAN addressed the imbalance in the dataset [28], but the suggested scheme excelled in all parameters.

6.6 Discussion

The CTG is an important fetal monitoring instrument during the intrapartum and antepartum phases. The CTG tracings are evaluated according to the FIGO criteria, yet inter-observer and intra-observer variability remains and may lead to unnecessary surgical births and cesarean sections. Recently, machine learning techniques have improved classification performance by reducing variance. An optimal set of input features, which remains challenging and time-consuming, influences the classification performance. The proposed model employed a 1D-CNN, suitable for a one-dimensional input, and has the notable trait of combining the feature mining and categorization methods in a single unit. The class balance is rarely present in medical datasets and is the second aspect influencing classification performance. The current study addresses the class imbalance problem using various scenarios: (i) considering the original unbalanced dataset; (ii) undersampling the dataset; (iii) employing the weighted binary cross-entropy method; and (iv) oversampling the dataset with SMOTE. The parameter tuning for the 1D-CNN model is done for two essential parameters: n_filters and kernel_s, as it highly impacts the overall performance of convolutional operation. The Se for the imbalanced dataset is 99.80%, which is an excellent classification rate for true positives (normal delivery cases) and suggests that unnecessary operational deliveries can be avoided. The bias introduced by an unbalanced dataset may account for a lower Sp value in this scenario. Due to the insufficient training examples, the outcome of the undersampling scenario needed improvement. During the weighted binary cross-entropy scenario, the 1D-CNN model performs significantly better than the unbalanced scenario, as Sp increases by 57.5%, QI by 42.26%, and AUC by 20.4%. However, a slight decrease in the Se by 3.55% can be observed here. The overall outcome due to the weighted cross-entropy method is significantly better than the unbalanced scenario. Compared to the weighted binary cross-entropy technique, SMOTE significantly increased around 11.5% in Sp, 7.91% in QI, 3.3% in AUC, and 3.56% in Se. Figure 12 shows a comparison graph of the four crucial parameters, Se, Sp, QI, and AUC, generated using four class balancing scenarios: unbalanced dataset, undersampling, weighted binary cross-entropy, and SMOTE. The graph demonstrates that the 1D-CNN model outperforms in terms of Sp, QI, and AUC in the SMOTE scenario compared to the rest.

Figure 12

1D-CNN model performance under different class balance scenarios.

The boxplots for Se, Sp, QI, and AUC also mark stability and consistent performance due to the very low variance in the acceptable range during the SMOTE scenario, as shown in Figure 10.

The Se plays a crucial role in preventing unnecessary interventions during labor. Apart from undersampling, the suggested model is robust enough to demonstrate consistent Se in all circumstances. The Sp is a crucial metric as it helps avoid fetal compromise; hence, a rise in Sp in the SMOTE scenario is paramount. Although the Sp value in the training phase is relatively high in the SMOTE scenario due to the availability of oversampled case records, it dropped during the testing phase. The reason might be a smaller number of case records available during the testing phase. The learning rate decay and batch normalization method lead to early network convergence, ultimately lowering the number of training epochs in unbalanced and SMOTE conditions. This feature makes the proposed network worth using with portable devices, where less memory and power requirements must be satisfied. Second, the low computational requirement of the 1D-CNN model, without any additional overhead for input preparation during the testing phase, makes it suitable for real-time applications. The system’s stable and consistent performance may be due to correctly handling zero values in the FHR signal during the input normalization phase. The most crucial factor significantly impacting the overall performance is the proper selection of n_filters and kernel_s for the convolutional layer. After conducting many trials in each case, the authors carefully selected these values during fine-tuning experimentation. The fine-tuning of these values leads to satisfactory results regarding Se, Sp, QI, and AUC. The experiments for hyperparameter setting and layer-wise parameters also contributed to the overall results. However, the authors strongly feel that other deep learning methods combined with CNN might help increase overall predictive capability in SP, QI, and AUC values. The overall outcomes are promising for using this SMOTE method to avoid unnecessary cesarean sections and to detect fetal compromise. The comparison with the state-of-the-art methods is also encouraging toward the model’s usefulness in clinical practices.