Introduction

More than 254 million pregnancies were recorded each year between 2015 and 20191, of which > 73 million ended in abortion2. Consequently, ˃140 million babies are born annually around the world3. In Asia, there are 16 live births per 1000 people; the corresponding figure in the United Arab Emirates (UAE) is 10 live births per 1000 people4 Some infants are born preterm (PTB) and/or with low birthweight (LBW). Newborns weighing < 2500 g are considered to be LBW infants, which can be caused by several factors, such as high parity, maternal diet, shorter interpregnancy intervals, premature delivery, and socioeconomic issues. PTB and very PTB (vPTB) refer to the delivery of an infant before 37 and 32 weeks of gestation, respectively. PTB and LBW are strongly associated with each other5,6,7,8,9,10,11,12. For example, a PTB infant is 18 times more at risk of being an LBW infant and vice versa12. Furthermore, LBW and PTB have substantial health impacts11 and are important determinants of infant health and survival. Compared to infants with normal birth weight, LBW infants are at a higher risk of perinatal death and have a greater chance of developing serious developmental problems, including mental retardation, low IQ, visual and auditory impairments, long-term disabilities, and premature death13,14,15. In contrast, PTB can place the child at a higher risk of serious health issues, such as gastrointestinal, respiratory, hearing, vision, cognitive, and growth problems as well as correlating a greater chance of permanent disability and death16. PTB is a major global concern that affects > 15 million infants every year, of which almost 1 million do not survive17. To increase awareness and improve treatment, World Prematurity Day and Prematurity Awareness Month are observed on November 17 and every November, respectively18,19. With a prevalence of 6.3%, PTB is a concern in the UAE12. Efforts have been made to prevent PTB births; however, their prevalence is still high20. Early detection and management can improve outcomes for both mothers and infants. In recent years, machine learning (ML) models have shown promising performance in various domains, such as in the field of obstetrics for LBW and PTB prediction.

Although many studies for LBW and PTB prediction have been performed (Table 1) these efforts have two major limitations. First, many pregnancies occur each year; however, the medical datasets are not publicly available due to privacy reasons, and most of the datasets often suffer from class imbalance, making it difficult to accurately predict minority classes. Data balancing techniques can be employed; however, they do not always effectively capture the distribution of the minority class, resulting in subpar performance. Anomaly detection algorithms can help address this issue. Second, ML models overlook inter-entity relationships and rely only on grid-based data18,19. As a result, patients in a dataset are treated as independent and uncorrelated21. This is not always accurate because patients can be correlated based on shared diseases or comorbidities12,20. These complex relational structures pose a challenge to the ability of ML models to extract meaningful information from data. Consequently, the effectiveness of ML models in uncovering valuable insights is hindered.

Table 1 Previous studies on the outcomes of adverse pregnancies.
Full size table

To overcome these limitations, we propose a novel method to predict adverse pregnancy outcomes (LBW, PTB, and vPTB) using a node embedding-based graph outlier detection algorithm22. The knowledge graph is developed from a well-curated dataset representative of the Emirati population, and a GAE is employed for outlier detection. Our solution considerably improves the performance of the LBW and PTB prediction models in both parous and nulliparous women, demonstrating the effectiveness of node embedding and graph outlier detection.

Related works

In this section, we provide a brief overview of studies conducted on adverse pregnancy outcomes using ML-based models. For example, Feng et al.23 utilized ultrasound features to predict fetal weight using data from 7875 women with 190 LBW samples. Trujillo et al.24 used support vector regression with a radial basis function kernel to estimate BW from a dataset of 250 women and 23 features such as maternal height, weight, and body mass index (BMI).

Similarly, various ML models (such as RF, SVM, NB, LR, DT, KNN, neural network, MLP, and ensemble models) have been used in various studies to predict LBW in infants6. The features used in these studies include average clinical attachment loss, clinical measures, education, gender, gestation age, height, hypertension condition, income, last weight recorded before conceiving, maternal age, medical history, mother’s age, newborn weight, number of children, parents’ education, periodontal parameters, place of residence, prepregnancy BMI, weight gain during pregnancy, and smoking. These studies used datasets ranging from 189 to 215,568 patients and achieved accuracy ranging from 72 to 97% with varying specificity and sensitivity. Important risk factors associated with LBW included the last weight recorded before conceiving, mother’s age, prepregnancy BMI, and weight gain during pregnancy. Further details can be found in25,26.

There is a comprehensive literature on PTB analysis from the medical perspective; however, there are limited studies predicting PTB using ML algorithms. For instance, Mercer et al.27 performed logistic regression-based analysis for PTB infants using data from 2929 women (multiparous: 1711, nulliparous: 1218). Lee et al.28 proposed ML methods to predict PTB and its important risk factors on a dataset of 596 patients. Further, they revealed that BMI was the most important risk factor followed by hypertension, diabetes mellitus, prior cone biopsy, and prior placenta previa for PTB prediction. Tran et al.29 proposed stabilized sparse logistic regression for PTB prediction and estimated important risk factors on a dataset of 15,814 women for PTB prediction, achieving an AUC of 0.85 at 34 weeks of gestation.

Taha et al.12 conducted a study based on statistical analysis to identify the factors associated with PTB and LBW in the UAE. They used a dataset of 1610 (Emiratis and expatriates) mothers from Abu Dhabi, UAE. Furthermore, they showed that PTB and LBW are highly associated with each other. They also revealed that Arab national women are twice at a higher risk of having PTB compared to non-Arab mothers. Sun et al.30 evaluated multiple ML algorithms and found that RF performed well (AUC of 0.89), with age, waist size, height, mean platelet volume, globulins, and serum inorganic phosphorus being key risk factors associated with PTB.

Koivu et al.31 used the CDC and NYC datasets (for external validation). An AUC of 0.67 was achieved for the CDC dataset, while a maximum AUC of 0.64 was achieved for the NYC dataset using the ANN and LightGBM models. Raja et al.32 also employed an ML model for PTB prediction. Belaghi et al.33 created a dataset of 112,963 nulliparous women with various risk factors selected based on the literature34,35 for PTB prediction. Belaghi et al.36 performed multivariate statistical analysis on a dataset of 267,226 births for PTB and spontaneous PTB in multiparous and nulliparous women. They found that the AUC using logistic regression in the first trimester was 0.68 and 0.73 for nulliparous and multiparous women, respectively, while it was 0.72 and 0.78, respectively, in the second trimester.

Diaz et al.37 proposed a ML model-based methodology for PTB prediction in chronodisrupted mothers. They used a dataset of 380 births (preterm: 157, term: 223) and showed that features, such as sleeping habits, were important for PTB prediction. Lee et al.38 used multiple ML models (LR, ANN, RF) for PTB using 90 different features, including particulate matter and depression. They used a dataset of 405,586 participants for the classification of PTB into four different categories. However, the dataset was highly imbalanced (2.22% in the minority class). Experiments were performed to show that an AUC of 0.52–0.58 was achieved.

Methods

An overview of the proposed system, which contains several modules, is shown in Fig. 1. In the first module, tabular data are transformed into a knowledge graph, and node embedding features are extracted. Node embedding vectors and tabular data are fed as inputs into the graph auto encoder (GAE). The yellow dots in Fig. 1 at the input of the GAE signify outliers, which are challenging to reconstruct at the output of the GAE because of their deviation from the norm. This deviation results in an increased reconstruction error; therefore, these points are referred to as outliers. We evaluate the results based on different metrics such as AUC and AUC-PR. Each module is explained in detail below.

Figure 1
figure 1

Proposed methodology for a node embedding-based graph outlier detector.

Full size image

Ethics statement

Our analysis used data from a prospective cohort study conducted in Al Ain, Abu Dhabi, UAE. The study included pregnant women from the Emirati population who were recruited and followed up via medical records in hospitals. As described in a previous study39, the participants completed a baseline questionnaire at recruitment. The study received approval from the Abu Dhabi Health Research and Technology Ethics Committee (DOH/CVDC/2022/72) and was in complete agreement with the Declaration of Helsinki. All participants provided written informed consent prior to data collection. Furthermore, all experiments were performed in accordance with relevant guidelines and regulations.

Dataset

The dataset used in this analysis was previously described and was obtained from an ongoing prospective maternal and child cohort, the Mutaba`ah Study, in Al Ain, UAE39. The risk factors used in this analysis were selected based on the literature and medical justification. Parous is defined as having given birth to a child, and nulliparous denotes a woman who has not given birth to a child. Data for 3508 patients were used, of which 2708 were parous (2411 with normal births and 297 with PTB) and 801 were nulliparous (of which 97 had experienced PTB). Only 35 of the 2708 parous women had experienced vPTB, whereas 22 of the 801 nulliparous women had vPTB delivery. Full details are provided in Table 2.

Table 2 Distribution of the dataset.
Full size table

Problem formulation and graph construction

An outlier, or anomaly, refers to instances in a dataset that considerably deviate from the majority of observations40. Despite their rarity, outliers can provide valuable insights and information. In our particular dataset, for example, vPTB samples constitute only a small fraction (1.3%) of the total observations, but their detection and prediction are important. Our method builds upon a dataset \(D\) that comprises patients with varying risk factors. Based on GAE, reconstructing patients who are rare (PTB, vPTB) is challenging, leading to higher reconstruction errors; as such, they are considered to be outliers. Our approach was inspired by Du et al.22; however, we introduced a new graph construction method. Moreover, our approach utilizes node embedding rather than a cosine similarity adjacency matrix, leading to improved GAE predictive capabilities. To construct a graph, various entities, such as patients, demographics, pregnancy conditions, medical history before pregnancy, and fetal characteristics, were identified as nodes. Each patient was assigned a unique identifier to create a distinct node, and relationships were established between the nodes to show edges. The topology of the graph was determined based on the similarity of its node properties such that patients with similar complications were positioned closer together25,41. To implement this in Neo4j, we used primary key and foreign key concepts to create nodes and relationships between them. In particular, we created a node for each entity identified in the text, such as a patient node, demographics node, and pregnancy condition node. Each node exhibits its own set of properties such as patient age, medical history, and fetal characteristics. The relationships between nodes were established based on the relationships identified in the text. For example, we created a "HAS_DEMOGRAPHICS" relationship between a patient node and a demographic node to indicate that the patient has demographic information.

Node2Vec

The Node2Vec node embedding algorithm was employed to capture the structural and semantic relationships between nodes and learn their low-dimensional representations. Node2vec is a scalable algorithm that employs a flexible objective function balancing the preservation of local and global graph structures. Based on a graph G = (V, E), Node2Vec learns an embedding for each node in the graph such that the embeddings capture the structural information of the graph. It generates random walks from the graph, which form sequences of nodes obtained by randomly traversing the graph, with a biased transition probability that balances between breadth-first and depth-first search strategies. For each node in the graph, the algorithm samples a set of random walks by performing a fixed number of steps, where each step follows a transition probability based on the node's proximity to the current node, as controlled by two hyperparameters, \(p\) and \(q\). The transition probability is then defined as follows:

$${P}_{xy}=\left\{\begin{array}{c}\frac{1}{p} \quad if\, {d}_{xy}=0\\ 1 \quad if\, {d}_{xy}=1 \\ \frac{1}{q} \quad otherwise\end{array}\right.$$

where \({d}_{xy}\) is the shortest path distance between nodes \(x\) and \(y\); \(p\) and \(q\) control the possibility of returning to the previous node and exploring new nodes, respectively. After generating random walks, the Node2Vec algorithm optimizes a skip-gram model to learn node embeddings that capture the local and global structural information of the graph. In particular, the skip-gram model aims to maximize the possibility of predicting context nodes given the target node in the random walks. Details regarding the Node2Vec algorithm can be found in25 and42.

FastRP

FastRP is a rapid and scalable algorithm for learning low-dimensional representations of nodes in large-scale graphs43. FastRP is based on randomized projections that map high-dimensional vectors to lower-dimensional space while preserving pairwise distances. This approach allows us to compute low-dimensional vector representations for each node in a graph, which can then be used for various downstream tasks. The FastRP algorithm factorizes the adjacency matrix of a graph into two low-rank matrices, representing row and column embeddings. Row embeddings are computed using a randomized projection technique that involves multiplying the adjacency matrix with a random Gaussian matrix. This results in a low-rank approximation of the adjacency matrix that preserves the pairwise distances between nodes. Column embeddings are obtained by transposing the row embeddings. The FastRP algorithm is defined as \(A \approx XWXT\), where \(A\) is the adjacency matrix of the graph, \(X\) is the row embeddings, \(W\) is the projection matrix, and \(XT\) is the column embeddings. The projection matrix \(W\) is computed by multiplying \(A\) by a random Gaussian matrix \(R\), then orthogonalizing the resulting matrix using \(QR\) decomposition. Row embeddings \(X\) are obtained by multiplying \(A\) by projection matrix \(W\), whereas column embeddings \(XT\) are obtained by transposing \(X\).

Graph auto encoder

GAE provides a powerful unsupervised learning framework for obtaining meaningful representations of graph-structured data. It comprises an encoder, a decoder, and a loss function that work together to learn the features between nodes and their neighbors. The encoder maps input data into a lower-dimensional representation, while the decoder maps the representation back to the data. For dataset \(X\), its associated node embedding from graph \(G\), weight matrices \({W}^{\left(i\right)}\), and bias matrices \({b}^{\left(i\right)}\), the GAE can be represented as follows:

$$Z=f\left(X,A\right)=LeakyReLU((LeakyReLU(X\cdot G\cdot {W}^{\left(0\right)}-{b}^{\left(0\right)})G\cdot {W}^{\left(1\right)}-{b}^{\left(1\right)})$$

The loss function, which measures the difference between the reconstructed output of GAE \(Z\) and the input \({X}^{\prime}\), is as follows:

$$f\left(X,A\right)=\sum \left(L\left({X}^{\prime},Z\right)\right)=\sum {\Vert {X}^{\prime}-Z\Vert }^{2}$$

The aim of GAE training is to uncover the underlying relationships between any patient \({p}_{i}\) and its neighboring patients. At the output of the GAE, \({p}_{i}\) and its neighbors are reconstructed such that the reconstruction error is minimized. Because the majority of the patients experience normal circumstances, they will be easier to reconstruct through GAE. Outliers present a challenge in the reconstruction because they significantly diverge from the norm.

Evaluation metrics

We used two commonly used performance metrics, namely, the area under the receiver operating characteristic (ROC) curve and the area under the precision–recall (PR) curve (AUC-PR), to evaluate the performance of the proposed method44. AUC-ROC and AUC-PR are important performance metrics for evaluating ML models. AUC-ROC measures the ability of the classifier to distinguish between positive and negative classes, whereas AUC-PR measures the precision–recall trade-off of the classifier. Details regarding the calculation and choice of these metrics can be found in25 and44.

Experiments and results

Graph construction was performed using Neo4j, and GAE was performed using MATLAB 2018B. All experiments were repeated ten times and the average results are presented. After performing knowledge graph creation, we were able to identify 2737 and 830 entities, known as nodes, by establishing 19,987 and 6061 relationships for parous and nulliparous patients, respectively. These include 2698 and 815 nodes with 19,203 and 5812 relationships for the LBW datasets in parous and nulliparous women, respectively. The graph produced from the initial risk factors using Neo4j is shown in Fig. 2.

Figure 2
figure 2

(a) Knowledge graph sample. Blue, yellow, and red nodes represent patients with their respective PIDs, risk factors, and target class information, respectively. (b) Sample from the congested graph for clear demonstration.

Full size image

We adopted the same experimental setup as discussed in22 and initialized key hyperparameters, such as the learning rate and the architecture of hidden layers. We tried different learning rates and hidden layer structures to determine the best setup. Finally, we found that a learning rate of \({1e}^{-9}\) and a three-layer architecture worked best in our experiments. Traditional GAE is sensitive to the value of \(k\); hence, we performed multiple experiments to determine the optimal value of k. The best performance was achieved for k = 90 (Fig. 3). Therefore, for GAE, the value of k was set to 90 for all experiments.

Figure 3
figure 3

AUC for different values of k for GAE.

Full size image

The performance results in Table 3 compare different methods for LBW prediction in parous and nulliparous women. The methods evaluated include original GAE; FastRP with 16, 32, and 64 features; Node2Vec with 16, 32, and 64 features; and a combination of FastRP and Node2Vec with features. The results show that the LBW prediction performance varies across different methods and feature sizes. The combination of FastRP and Node2Vec with 32 features performed relatively better than the other methods for LBW prediction in parous women with AUC-ROC and AUC-PR values of 0.6077 and 0.1545, whereas FastRP with 64 features achieved the best performance for nulliparous women with AUC-ROC and AUC-PR values of 0.5796 and 0.2305, respectively. The original GAE performed worst for parous and nulliparous women with AUC values of 0.4982 and 0.4891, respectively. The overall improvements in AUC-ROC for parous and nulliparous women using node embedding GAE were ~ 11% and 9%, respectively.

Table 3 Experimental results for LBW prediction using anomaly detection algorithms.
Full size table

The results of the PTB prediction in parous and nulliparous women are shown in Table 4. The AUC values for all methods are ~ 0.5. Among the methods for parous women, FastRP with 16 features performs the best. The AUC-ROC of the combined 64 features was higher; however, FastRP with 16 features achieved better performance with AUC-ROC and AUC-PR values of 0.5026 and 0.1491. For nulliparous women, the best AUC-ROC of 0.5647 was achieved using FastRP with 64 features that exhibits more than a 5% improvement relative to GAE-based predictions.

Table 4 Experimental results for PTB prediction using anomaly detection algorithms.
Full size table

Table 5 shows the results for vPTB prediction. The best results for parous women were achieved using Node2Vec with 16 features, showing an AUC-ROC of 0.5756 with more than 7% improvement compared to the original GAE-based implementation. Similarly, an improvement of > 11% was observed for nulliparous women, which achieved an AUC-ROC of 0.6696 compared to 0.5555 achieved by the original GAE-based approach.

Table 5 Experimental results for vPTB prediction using anomaly detection algorithms.
Full size table

Patient explanations

Knowledge graphs offer a more effective approach to explain the reasons behind an outcome. Figure 4 shows three patients, PID 1, PID 3, and PID 12, who were predicted as outliers for PTB using the proposed method. Note that PID 12 did not experience PTB, although it was classified as an outlier using the GAE-based method. All three patients exhibited common risk factors, including anxiety about their upcoming birth and a prior history of PTB. Moreover, PID 12 shares additional risk factors with either PID 1 or PID 3, including previous pregnancy loss, infection of the amniotic sac, planned pregnancy, education level, and delivery of a female baby. In addition, the GAE-based method displayed the ability to identify specific risk factors for individual patients.

Figure 4
figure 4

Explanation of GAE detection for PTB prediction.

Full size image

Figure 5 shows an example of patients with vPTB complication predicted as an outlier using the proposed method. The green patient nodes with IDs 43 and 30 are correctly detected as outliers, whereas patients 27 and 38, who did not experience vPTB, are detected as outliers by the GAE-based method. Figure 6 shows a use case for LBW, where two patient nodes (556 and 2410) are accurately recognized as outliers. However, the third patient node (PID 2440), who did not experience LBW, was wrongly classified as an outlier.

Figure 5
figure 5

Predicted very PTB outlier patients using the GAE-based method.

Full size image
Figure 6
figure 6

Predicted LBW outlier patients using the GAE-based method.

Full size image

Discussion

Herein, we used node embedding-based GAE outlier detection to address the limitations of traditional ML models in handling class imbalance and inter-entity relationships in medical datasets. This method detects outliers in graph-structured data, where each node and edge represent an individual patient and the relationships or interactions between patients, respectively. Notably, the prediction model performance was improved for PTB, vPTB, and LBW when we incorporated node embedding and GAE. For nulliparous women, an AUC-ROC of 0.6696 demonstrated an improvement of > 11% compared to an AUC-ROC of 0.5555 achieved using the original GAE-based approach. For parous women, an AUC-ROC of 0.5756 showed more than 7% improvement compared to the original GAE-based implementation. For LBW prediction, the AUC-ROC and AUC-PR values for parous women using node embedding GAE were 0.6077 and 0.1545, respectively, demonstrating an 11% improvement. Furthermore, the AUC-ROC and AUC-PR values for nulliparous women were 0.5796 and 0.2305, respectively, demonstrating a 7% improvement.

In addition to the improved predictions, we performed individual patient analysis for outlier (PTB and LBW) prediction. Based on our findings, we identified three patients (PID 1, PID 3, and PID 12) as outliers for PTB, indicating that all three patients are at higher risk of PTB. Notably, PID 12 was known to have delivered a full-term baby, despite sharing several risk factors with PID 1 and PID 3, such as infection of the amniotic sac, exposure to passive smoking, premature rupture of membranes, and concern for birth. These risk factors have been previously reported to be highly associated with PTB delivery8,45,46,47,48. There are several potential explanations for this anomaly. First, it should be noted that no predictive model, including autoencoder models, is 100% accurate. The performance of these models heavily relies on the quality and quantity of data used for training, and our model was trained on a relatively small dataset. Second, PTB can have multiple underlying causes and risk factors, some of which may not have been included in our list of risk factors used to train our model. Factors such as ultrasound parameters, biomarkers, and fetal fibronectin are known to contribute to PTB risk and may not have been captured in our dataset. Third, there may be human errors associated with the data used in this study during the data collection process or in the patient’s medical history. For example, the patient’s gestational age or delivery date may have been recorded incorrectly, or there may have been missing data that could have influenced the model’s predictions. These protective factors are related to the doctor–patient relationship. For example, receiving regular prenatal care from a healthcare provider can help identify and address any potential health issues that may increase the risk of PTB. Early intervention and optimal treatment of potential health issues such as hypertension and gestational diabetes, maintaining healthy diets and exercise, and regular cervical length assessments are some ways to help reduce the risk of PTB and promote the health and well-being of both the mother and baby. Similar explanations can be used to describe LBW anomalies.

Our model’s prediction for PTB and LBW showed improved performance compared with the traditional GAE approach; however, this result should always be considered in the context of the patient’s individual circumstances, medical history, and healthcare interventions provided. Moreover, the model can assist in the early detection of patients at risk of PTB and LBW, thus allowing for timely and personalized interventions to prevent adverse outcomes. These results can serve as a starting point for further research and pave the way for the development of more effective and accurate prediction models for adverse pregnancy outcomes.

PTB and LBW are the major causes of infant morbidity and mortality, and early prediction can help clinicians manage and prevent PTB and LBW. Our model was trained on a prenatal and clinical dataset to detect patterns associated with PTB/LBW. Consequently, our model can be used to predict the likelihood of PTB/LBW in new patients. Thus, clinicians can use this model to screen and risk stratify new patients as well as identify patterns in PTB/LBW data that are unique to individual patients, ultimately personalizing treatment plans that are tailored to the specific needs of each patient.

Some limitations to our model’s predictions must also be acknowledged. First, the dataset size and population diversity may not be representative of other populations. As such, larger and more diverse datasets may lead to different results, necessitating further investigation. However, the results of this study may be generalizable to other populations with appropriate modifications. The data used contain a majority of binary variables. Therefore, it is challenging to accurately capture and model inter-entity relationships. The proposed solutions may not effectively capture all complex relationships in the data, leading to less accurate performance. Furthermore, formal hyperparameter optimization procedures were not conducted. Instead, various settings were explored to assess their influence on the model’s performance, leaving room for potential improvements using advanced transformer-based generative models. Moreover, using a GAE and node embedding can be computationally expensive and may require high computational resources.

While our model’s prediction for PTB and LBW showed an improved performance compared with the traditional GAE approach, it is important to emphasize that the purpose of this research is to assist and augment the capabilities of clinicians rather than replace them. In addition, the practical implementation of this setup has not yet been adopted, presenting a potential avenue for future exploration. Furthermore, to ensure responsible and beneficial deployment of these technologies, it is important to consider the ethical considerations associated with the application of ML in healthcare, addressing concerns related to privacy, informed consent, and bias. Future studies could explore these issues in more detail and guide how to address them in practice.

Conclusion

The results of this study demonstrate the potential for incorporating node embedding and graph outlier detection as a means of improving the accuracy of prediction models for adverse pregnancy outcomes (e.g., PTB and LBW), offering a unique advantage over traditional ML models. This prediction model can be a useful tool for identifying new patients at higher risk of PTB/LBW. By flagging these patients for further evaluation and interventions, this technique can potentially improve PTB/LBW outcomes and reduce associated health risks. Using the benefits of our predictive model, healthcare professionals can be proactive in PTB/LBW risk prediction and provide personalized and targeted patient-centered medical treatment, thus improving the morbidity and mortality outcomes in pregnancy.