The incidence of pre-eclampsia (PE) is 2%–8% and results in 7%–8% of neonatal deaths or defects, (Ananth et al., 2013) accounting for 6% of maternal deaths during the perinatal period. Fetal growth restriction (FGR) is a serious complication of hypertensive disorder complicating pregnancy (HDCP), accounting for 42.3% of all perinatal deaths in China. (Vigil-De et al., 2003) Pre-eclampsia complicated by FGR may result in increased perinatal mortality. Barker’s (Barker et al., 1989) hypothesis suggests that FGR affects fetal, physical, and intellectual development in childhood and adolescence and significantly increases the incidence of adult metabolic syndrome. However, an early diagnosis of FGR is not easy, resulting in a relative delay in intervention and a lack of effective treatments. If a feasible prediction method can be found, it will be highly valuable in improving the outcomes of PE complicated by FGR.

Artificial neural network (ANN) models can predict complex relationships between variables that other models cannot compare, such as logistic regression models. Research in the biomedical field shows that the ANN is more accurate in classifying dependent variables than the logistic regression model. (Renganathan, 2019) Pre-eclampsia is affected by multiple factors, and common prediction method models cannot capture the complex cross-correlation of these factors. Artificial neural networks have been studied in PE diagnoses and delivery mode predictions. (Xu et al., 2019) Nair(Nair, 2018) constructed an ANN-based machine learning model using the identified biomarker microarray dataset and successfully classified PE and normotensive populations. The advantages of ANNs in dealing with these complex and heterogeneous features are confirmed again in the present study, which adopts the ANN to investigate the influence of maternal clinic characteristics and relevant blood parameters on pregnancy outcomes in PE complicated by FGR and to assess their value in predicting the occurrence of adverse perinatal outcomes.

Table 1 presents the demographic and obstetric characteristics of the study population. The patients in the concurrent group had a higher pre-pregnancy BMI and were older than those in the PE and FGR groups. Women with more PE in prior pregnancy in the concurrent group complicated the FGR group. The differences between the three groups in adverse pregnancy and delivery gravity and parity were not statistically significant. The selection of research objects is shown in Figure 1.

Analysis of maternal pregnancy outcomes showed that the incidences of fetal distress, placental adhesion, and postpartum hemorrhage were higher in the concurrent group than in the PE group. The incidences of HELLP, ICP, postpartum hemorrhage, premature rupture of membranes, placental abruption, placental adhesion, and fecal contamination of amniotic fluid were higher in the concurrent group than in the FGR group.

Analysis of neonatal–perinatal outcomes revealed that the incidences of prematurity, mild neonatal asphyxia, severe asphyxia, and admission to the newborn intensive care unit were higher in the concurrent group than in the PE and FGR groups. These differences were statistically significant (Table 2).

The following hematological parameters with statistically significant differences were screened by intergroup difference analysis (p < 0.05): inflammatory parameters (neutrophil-to-lymphocyte ratio [NLR] and platelet-to-lymphocyte ratio [PLR]), coagulation parameters (prothrombin time [PT] and thrombin time [TT]), lipid parameters (high-density lipoprotein [HDL], low-density lipoprotein [LDL], and triglyceride [TG]), platelet parameters (platelet count [PLT], mean platelet volume [MPV], and plateletcrit [PCT]), uric acid (UA), lactate dehydrogenase (LDH), and total bile acids (TBA). All these parameters could be used as input nodes for the ANN as the basis for constructing the ANN models, as illustrated in Table 3.

Cases were randomly divided into training and testing sets with a ratio of 2:1. A total of 15 parameters, including the serological parameters that were statistically different between groups and the maternal ages and pre-pregnancy BMIs, were used as input nodes to construct the ANN model. The Batch_Size was selected as 10. and the learning rate was set to 0.3. After experiment, the optimal number of hidden layer nodes was selected. The number of hidden layer nodes in each model was 10 for preeclampsia complicated with fetal growth restriction model, eight for postpartum hemorrhage model, eight for fetal distress model, eight for premature rupture of membranes model, eight for preterm delivery model, and 10 for FGR severity model. The trained ANN was tested for accuracy using the data in the testing set. The accuracy of the ANN for predicting the occurrence of the following diseases and adverse outcomes was as follows: 84.3% for PE complicated by FGR, 81.9% for predicting the severity of FGR, 80.3% for premature rupture of membranes, 80.1% for postpartum hemorrhage, 77.6% for fetal distress, and 76.3% for provider-initiated preterm delivery. The details are shown in Table 4.

Pre-eclampsia is a disease peculiar to pregnancy. The perinatal mortality rate is increased, and it is closely related to the short-term and long-term complications of the mother and child. Because an early diagnosis of FGR is not easy, it is very important to study and predict the risk factors of PE complicated by FGR. This study retrospectively analyzed the clinical data and hematological parameters, such as inflammatory, coagulation, and blood lipid indexes, and the platelet parameters of pregnant women in PE complicated by FGR. There were 15 risk factors of PE complicated by FGR screened out, and a prediction model for predicting disease occurrence and adverse perinatal outcomes was established by using an ANN. This study analyzed the clinical characteristics and perinatal outcomes of PE complicated by FGR. The results showed that the age and pre-pregnancy body mass index of women with PE complicated by FGR were larger than those of women with PE or FGR alone. The incidences of fetal distress, placental adhesion, postpartum hemorrhage, preterm delivery, and mild and severe neonatal asphyxia were significantly higher in patients with PE complicated by FGR than in those with only PE or FGR. These results confirm that women with a high body mass index before pregnancy and older women were more likely to have PE complicated by FGR. The perinatal outcomes in patients with PE complicated by FGR were worse than in those with only PE or only FGR. Savitri et al. (Savitri et al., 2016) also found that the pre-pregnancy BMI, rather than the weight gain during pregnancy, determined the level of blood pressure during pregnancy and was related to the higher probability of PE and its complications. However, the close relationship between older age and PE is well known. Another study found that the older age of pregnant women is related to uterine and placental dysfunction; (Lean et al., 2017) this should be the key to the adverse consequences of FGR and its effects on the fetus caused by the aging process. The most common cause of FGR is placental insufficiency due to impaired placental blood flow and vascular abnormalities, (Alfaidy et al., 2014; Alfaidy et al., 2020), (Alfaidy et al., 2014; Alfaidy et al., 2020) which is similar to the pathogenesis of PE. Therefore, patients with PE are prone to FGR, which aggravates the original pathological state and causes the perinatal outcome of the mother and child to be worse than with PE or FGR alone.

Maternal serological parameters changed markedly with the onset and progression of PE. This study found that 13 parameters, including blood lipid indexes (HDL, LDL, TG), coagulation indexes (PT, TT), platelet parameters (PLT, MPV, PCT), inflammatory indexes (NLR, PLR), uric acid, LDH, and total bile acid were related to PE complicated by FGR. Furthermore, the levels of TG and LDL were higher, and the HDL level was lower in patients with PE complicated by FGR than in those with only PE, and the differences were statistically significant. The mechanism for this might be because increased serum TG concentrations and increased LDL production during pregnancy cause increased calcium levels in vascular smooth muscle cells, raise the degree of oxidative stress, and injure the vascular endothelial cells, thus triggering PE. (Goulopoulou and Davidge, 2015) Furthermore, dyslipidemia disrupts endothelial cell function and stimulates endothelial cells to synthesize more thromboxane, which in turn causes blood vessels to spasm. This would lead to shallow chorionic villus implantation, which can reduce blood flow to the placenta and uterus and impede nutrient supply and fetal development, which can further induce the occurrence of FGR in patients with PE.

The results of the oral glucose tolerance test during mid-pregnancy and the fasting plasma glucose during late pregnancy were compared among the three groups in the present study. The results showed that the differences were not statistically significant. The reason for these outcomes might be correlated with the fact that patients with gestational diabetes mellitus or a pre-pregnancy history of diabetes mellitus were excluded from the present study. Whether glucose metabolism influences the occurrence of PE complicated by FGR, and the mechanisms involved, requires further investigation.

Platelet activation may occur during the early stage of PE. (Kohli et al., 2016; Armalyz et al., 2018) The results of the present study showed that PLT in patients with PE complicated by FGR was lower than in those with only PE and that MPV and platelet distribution width (PDW) were higher, and the differences were statistically significant (p < 0.05). This may be correlated with platelet aggregation and depletion due to injury to the vascular endothelium in patients with PE complicated by FGR. This aggregation and depletion would cause the average PDW to increase due to the reduced platelet count, resulting in increased blood viscosity and the potential for microthrombosis. Microcirculatory disorders might also lead to ischemia and hypoxia, which could further affect the growth and development of all organs and the fetus. (Kurtoglu et al., 2015) The present study also found that PT and TT in patients with PE complicated by FGR were lower than in those with only PE or only FGR, while D-dimer levels were higher in patients with PE complicated by FGR than in those with only FGR, and the differences were statistically significant (p < 0.05). Shortened activated partial thromboplastin time and PT are closely correlated with a hypercoagulable state and represent a higher likelihood of thrombosis, leading to multiple thromboses, reduced fibrinolytic activity, and impaired ability to repair endothelial cells.

The occurrence of PE is mainly correlated with elevated levels of the inflammatory response. (Zhao et al., 2017) The platelet-to-lymphocyte ratio is closely correlated with immune response and can be considered a class of immune response factors. (Jeon et al., 2017) The neutrophil-to-lymphocyte ratio is strongly correlated with cardiovascular diseases (such as coronary heart disease), inflammatory diseases, and obstetric and gynecological malignancies. (Myatt and Cui, 2004) The results of the present study provide further evidence that inflammatory immune mechanisms might play a role in disease progression in patients with PE.

Most current prediction studies are based on logistic regression models to predict only FGR or PE. (Yu et al., 2017; Saleh et al., 2021; Feng et al., 2022) Most of these studies were single-index predictive analyses, such as the glucose and lipid metabolism index or single coagulation index, with unsatisfactory sensitivity and specificity. For example, Ren Danyu et al. (Ren and Wang, 2019) analyzed the coagulation function index and its application value in judging the severity of PE, and the area under the receiver operating characteristic curve was 0.67 (p > 0.05). Few studies have been conducted on creating a predictive model for FGR in patients with PE. Artificial neural networks are distributed parallel information-processing algorithmic mathematical models that imitate the behavioral characteristics of animal neural networks. This type of network relies on the complexity of the system to process information by adjusting the interconnectivity between a large number of internal nodes. Wang et al. (Wang et al., 2016) used logistic regression and ANN models to evaluate the high-risk factors of PE before and during pregnancy and compared the efficiency of the two in diagnosing PE. They found that the fitting effect of the ANN in grading PE in the third trimester of pregnancy is more suitable for multivariate analysis of complex diseases due to logistic regression, and the efficiency of ANN in disease diagnosis and prediction has also been confirmed in many studies. (Ibrahim et al., 2022; Miao et al., 2022)

At present, relevant data from ultrasounds are used in predicting PE and have proven to be valuable. (Lai et al., 2022) Ultrasound measurement is considered a bone marker, neglecting soft tissues, such as fetal fat and muscle, and the cross-section selection and fetal position’s influence will lead to estimation errors. The pathophysiological changes of PE are systemic multi-system and multi-organ damage, which will cause harm to the mother and child. Ultrasound generally monitors the changes in the fetus and uterus, which may occur later than the hematological parameters and other indicators reflecting the state of systemic diseases. Therefore, selecting clinical characteristics and hematological parameters for prediction is more meaningful than ultrasounds. The present study combined ANNs, interdisciplinary cooperation, and an artificial intelligence system to build an effective and simple predictive model based on hematological parameters and clinical characteristics of pregnant women. The internal verification showed that the ANN accuracy rate of predicting PE complicated by FGR is 84.3%, the accuracy rate of predicting FGR severity is 81.9%, and the accuracy rate of predicting preterm birth is 76.3%. Moreover, these models all have a good F1 score (this value balances the recall and accuracy of the model and is a good evaluation index of the ANN model). In addition, the ANN model for predicting postpartum hemorrhages, premature rupture of membranes, and fetal distress has a good accuracy rate. However, due to the influence of the basic incidence rate of disease, the data groups are in an unbalanced state, and the accuracy and F1 scores of the model are low, though this cannot deny the value of the model.

In pregnant patients with PE, monitoring the changes in relevant hematological and metabolic parameters during pregnancy offers a useful guide to predicting the development of FGR and improving perinatal outcomes. The present predictive model, combining coagulation parameters, lipid metabolism, and inflammatory hematological parameters, has a good predictive value in predicting the occurrence of PE complicated by FGR and pregnancy complications, such as postpartum hemorrhage and fetal distress.