A Prediction Model of Preeclampsia in Hyperglycemia Pregnancy

Introduction

The International Federation of Obstetrics and Gynaecology (FIGO) defined “hyperglycemia in pregnancy” (HIP) as hyperglycemia covering all severity levels, which can be divided into DIP (either preexisting diabetes—type 1 or type 2—which antedates pregnancy or is first identified during testing in the index pregnancy) and GDM.¹ According to the latest (10th) edition of the International Diabetes Federation’s Diabetes Map 2021, approximately 16.7% of women of childbearing age worldwide experience varying degrees of elevated blood glucose during pregnancy.² HIP is associated with a significantly increased risk of adverse perinatal outcomes, including preeclampsia, birth injuries, increased caesarean section rates, and an increased incidence of type 2 diabetes mellitus in offspring.^3–5

The International Society for the Study of Hypertension in Pregnancy (ISSHP) defines PE as new onset hypertension and proteinuria after 20 weeks of gestation or with other end-organ damage, that affects 2–8% of pregnancies worldwide, and can lead to preterm delivery, fetal growth restriction, maternal convulsions, damage to vital organs such as the heart, liver and kidneys and in severe cases, maternal and fetal death.^6–8

It has been shown that HIP is associated with an increased risk of PE.⁹ Some studies have shown that mechanisms related to glucose metabolism are involved in the pathogenesis of hypertensive disorders in pregnancy; in addition, PE and HIP share many common risk factors and similarities in pathophysiology, including immune system disorder, vascular endothelial dysfunction, and insulin resistance.^10–15 In this study, we focused on pregnant women with HIP, with the hope of screening those at risk of preeclampsia through simple routine pregnancy care information so that we can take more effective interventions.

Method

Participants and Ethics

Information on 951 pregnant women with hyperglycemia who delivered as an inpatient at the Maternal and Child Health Hospital Affiliated at Anhui Medical University from January 2017 to December 2019 was retrieved through the medical electronic medical record system. The inclusion criteria were as follows: (1) singleton live birth; (2) gestational age at delivery ≥34 weeks; (3) having undergone liver and kidney function at 24–29⁺⁶ weeks of gestation; (4) without chronic hypertension or chronic hypertension combine with preeclampsia; (5) without gestational hypertension; (6) those with age, number of deliveries and basal blood pressure recorded. HIP includes diabetes during pregnancy and gestational diabetes mellitus. The diagnosis of diabetes in pregnancy was made according to the World Health Organization (WHO) criteria; that is, routine testing at any time during pregnancy, if any of the following cut‐off points was met: fasting plasma glucose (FPG) ≥7.0mmol/L; 2-hour plasma glucose (PG) ≥11.1mmol/L following a 75-g oral glucose load; random PG ≥11.1mmol/L in the presence of diabetes symptom.¹⁶ The diagnosis of GDM was made according to the International Association of Diabetes and Pregnancy Study Group (IADPSG) criteria; that is, all patients underwent an oral 75 g glucose tolerance test at 24–28 weeks with overnight fasting for 10 h. Plasma glucose test measurement was performed at 1 and 2 h after oral 75 g glucose tolerance. A GDM diagnosis was made when the plasma glucose values exceeded any of the defined thresholds (fasting: 92 mg/dL (5.1 mmol/L); 1 h: 180 mg/dL (10.0 mmol/L); and 2 h: 153 mg/dL (8.5 mmol/L)).¹⁷

This was a retrospective study of routinely collected clinical data, and exemption from informed consent was approved by the Medical Ethics Committee of Maternal and Child Health Hospital Affiliated at Anhui Medical University (approval number YYLL2022-2108085MH260-03-01). All experimental protocols were approved by the Medical Ethics Committee of Anhui Maternal and Child Health Hospital and all methods were carried out in accordance with the Declarations of Helsinki.

Candidate Predictor Variables

General information of the pregnant women were obtained from the medical records, including maternal age, parity, basal systolic blood pressure (SBP), basal diastolic blood pressure (DBP); liver and kidney function indexes obtained at routine pregnancy care at 24–29⁺⁶ week of gestation, include albumin (ALB), creatine kinase (CK), alkaline phosphatase (ALP), cystatin C (CYSCL), uric acid (UA), creatinine (CREA), alpha hydroxybutyrate dehydrogenase (HBDH), white globule ratio (A/G), aspartate transaminase (AST), total bile acid (TBA), alanine aminotransferase (ALT), total bilirubin (TBIL), glutamyl aminotransferase (GGT), glucose (GLU), total protein (TP), urea nitrogen (BUN), indirect bilirubin (IBIL), direct bilirubin (DBIL) and globulin (GLO).

Statistical Analyses

R language (R.4.2.2) was applied for analysis and the measurement data were expressed as M (IQR); the Wilcoxon rank sum test was used for comparison between groups, while the count data were expressed as n (%) and the Fisher test was used for comparison between groups. P-value <0.05 in two-tailed tests was considered statistically significant.

Model Development and Internal Validation

Interactions and covariance between the variables were first tested using the occurrence of preeclampsia as the dependent variable. Logistic regression models were developed using stepwise regression. Restricted cubic spline (RCS) was used to examine the association between measurement variables and the risk of PE. Interactions and covariance between the measured variables were also tested. The Bootstrap re-sampling (1000 times) method was used for internal validation. Finally, we used a nomogram to visualise the model.

Model Performance Evaluation

The performance of the model was evaluated around three aspects: discrimination, calibration and clinical utility. In this study, AUC was used to evaluate the model’s discrimination ability. Hosmer–Lemeshow tests and calibration curves were used to determine the degree of agreement between predicted probabilities and observed outcomes. Decision curve analysis (DCA) was used to assess clinical validity. Overall model performance was evaluated by using the Brier score. All analyses in this study were performed using R software (version 4.1.3). All tests were two-tailed, and the difference was statistically significant at P <0.05.

Results

Basic Characteristics of Patient Population

The general information on pregnant women with hyperglycemia was divided into PE and non-PE group according to the occurrence of PE. When comparing the general information of pregnant women in the two groups, the difference in age and parity at conception was not statistically significant (P>0.05); the basal systolic blood pressure and basal diastolic blood pressure in the PE group were both higher than that in the non-preeclampsia group, and the difference was statistically significant (P<0.05). Comparison of liver and kidney function l indicators between the two groups: alkaline phosphatase (ALP), cystatin C (CYSCL), uric acid (UA) and glutamyl aminotransferase (GGT) were all higher in the PE group than in the non-PE group, with statistically significant differences (P<0.05), while the differences in other indicators between the two groups were not statistically significant (see Table 1).

Table 1 Basic Characteristics of the Patient Population

Predictive Model Development

The final mathematical equation determined by the logistic regression model to predict PE was as follows: . The predictive model was presented using a nomogram, which can be used to quantitatively predict the risk of preeclampsia in pregnant women of HIP (see Table 2 and Figure 1). Furthermore, a webpage calculation tool (model URL: https://preclampsia-predict.shinyapps.io/dynnomapp/) was created and the values of the variables were entered in the left column of the webpage to derive the probability value of the case being preeclamptic.

Table 2 Multivariate Logistic Regression Analysis of Preeclampsia

Figure 1 A nomogram predicting the risk of preeclampsia in hyperglycemia in pregnancy.

The variables were evaluated using the variance inflation factor (VIF) and Cook’s distance. The VIF of the models, 1.0352–1.0946, was much lower than 10, suggesting that there was no covariance between the variables of the models. The strong point-of-influence analysis showed that Cook’s distance of models was much lower than 1 (see Figure 2). And Figure 3 displays the linear relationship between variables and the risk of PE (all P for non-linearity >0.05). As shown in Figure 3, the risk of PE in pregnant women with HIP decreased with the increase of BUN. On contrary, there was a positive correlation between the levels of UA, GGT, SBP, and CYSCL and the risk of preeclampsia.

Figure 2 Cook’s distance plot of the model.

Figure 3 The relationship between variables and predicted outcomes. (a) Cystatin C and the risk of preeclampsia. (b) Uric acid and the risk of preeclampsia. (c) Glutamyl aminotransferase and the risk of preeclampsia. (d) Blood urea nitrogen and the risk of preeclampsia. (e) basic basal systolic blood pressure and the risk of preeclampsia. Hazard ratios are indicated by solid lines and 95% CIs by shaded areas.

Evaluation of the Prediction Model

AUC values were calculated to assess the discrimination of the predictive model by examining the occurrence of preeclampsia with HIP. As shown in Figure 4, the ROC curve showed an AUC=0.8031 (95% CI: 0.7383–0.8679), with an optimal threshold of 0.0805, at which point the sensitivity was 0.8307 and specificity of 0.6604. When the threshold was set at 0.0270, the sensitivity was 0.9057 and the specificity was 0.4644. The result suggested that the line graphs have good discriminatory and predictive value and can correctly identify HIP pregnant with and without PE.

Figure 4 Receiver operating characteristic (ROC) curve of the model. The green area represents the 95% confidence interval of area under the ROC curve.

The Hosmer–Lemeshow test result showed that the model had a very good fit for the model (χ² = 8.6393, df=8, p = 0.3736). The calibration plot (Figure 5) for the nomogram revealed good predictive accuracy between the actual probability and predicted probability of preeclampsia. Precision was measured by Brier score, which had a value of 0.0461.

Figure 5 Calibration curve plot of the model.

The clinical utility of the model was evaluated using DCA curves. The results are shown in Figure 6. The decision curves show that the net benefit of the prediction model is significantly higher than the two extreme cases, indicating that the nomogram has a high net benefit and predictive accuracy.

Figure 6 Decision curve analysis of the model.

Internal Validation

The resulting model was internally validated by 1000 bootstrap re-samplings. A similar result was observed between the validation cohort and the training cohort. In the validation cohort, the AUC was 0.7886 (95% CI: 0.7750–0.8043) (Figure 7). Meanwhile, the calibration curve showed a good fit during internal validation (Figure 8), while the Brier score was 0.0478 after bootstrap correction. As a result, the prediction ability of the model is acceptable.

Figure 7 Receiver operating characteristic (ROC) curve of the model after 1000 bootstrap re-sampling. The red area represents the 95% confidence interval of area under the ROC curve.

Figure 8 Calibration curve plot of the model after internal validation.

Discussion

Many studies have reported that the risk of preeclampsia was higher in pregnant women with HIP than in those without HIP. For example, the findings of Yogev et al supported the suggestion that HIP promotes the development of PE.¹⁸ It has also been suggested that a history of HIP in the first pregnancy is also a risk factor for PE in the second.^19–22 HIP and PE are inextricably linked, maybe because they have similar risk factors and pathophysiological changes. For example, High-risk factors for HIP and PE both include obesity and advanced age.²³ And study has shown that inflammation and endothelial dysfunction are both involved in the pathogenesis of PE and HIP.²⁴

To date, the early prediction of preeclampsia is challenging because little is known about the aetiology and pathogenesis of PE. Currently, research on the indicators of preeclampsia prediction models has focused on maternal characteristics, medical history, imaging indicators, PLGF, PAPP-A, etc.^25–27 This is an indicator not provided by routine antenatal care, making the applicability and external validation of these models very difficult. Our research proposes a prediction method to address this issue. We developed a preeclampsia prediction model for the HIP population at 24–29⁺⁶ weeks of gestation, using clinical and laboratory data that are readily available during routine antenatal visits. The AUC of the model was 0.8031 (95% CI: 0.7383–0.8679), which is higher than the PE prediction model proposed by Bijl et al.²⁸

Uric acid is an end product of purine metabolism and studies have shown that impaired uric acid metabolism can affect the development of preeclampsia.^29,30 Also, the results of Bellos et al³¹ demonstrate that elevated serum uric acid levels in patients with preeclampsia can be used to predict disease severity and pregnancy complications. This is in line with our study. The results of this study showed that serum uric acid levels were higher in the PE group than in the normal group and were an independent risk factor for PE; also, the risk of PE in pregnant women with HIP increased with increasing uric acid levels. In our study, BUN was a protective factor for patients with PE, and an increase in their levels can to some extent reduce the risk of PE. It is worth mentioning that there are also studies indicating that BUN is a risk factor for preeclampsia, which may be a direction for future research.³² In addition, changes in cystatin C were found to be significant in this study. Each 1 mg/dL increase in cystatin C was associated with a 2.04-fold increase in the risk of preeclampsia in pregnant women with HIP (OR = 2.04; 95% CI:1.51–2.79, P<0.001), suggesting that early changes in renal function are of value in predicting the development of PE in pregnant women with HIP. The reason for this may be that cystatin C correlates with the degree of glomerular endotheliosis, the typical glomerular lesion of PE.³³ And, PE patients may develop systemic small vessel spasms, including those affecting the glomerulus, which worsen as the condition progresses, ultimately leading to further renal function impairment due to glomerular endothelial lesions and renal vascular injuries.³⁴ Serum GGT has certain clinical value for various liver and gallbladder diseases. Previous studies have shown a positive correlation between GGT and the risk of hypertension in adults.³⁵ This study found that GGT levels also affect blood pressure in pregnant women. In our study, as serum GGT levels increase, pregnant women with HIP also have an increased risk of developing PE. This is consistent with the research of Chen et al.³⁶ And they proposed a hypothesis about the pathogenesis, where serum GGT affects the production of free radicals by interacting with iron and other transport proteins.³⁷ Basal blood pressure has been used as a maternal risk factor to construct a preliminary screening model. Moreover, the literature suggests that systolic blood pressure is more important in the development of preeclampsia.³⁸

The significance of our model is to use appropriate statistical methods to analyze readily available data from maternity examinations of pregnant women with HIP, explore high-risk factors for PE in HIP women, and then establish a predictive model that can identify high-risk patients with a low-cost routine. The high-risk patients would then be followed up to develop more targeted screening and treatment plans. For example, when we set the threshold of the model to 0.0270, we can screen 90.57% of PE patients. Although the specificity of the model at this time is 46.44%, we can use a more accurate but cost-effective screening method based on this model, such as the “triple test” method using MAP, PLGF and sFlt-1 measurements, which has a PE detection rate of 79% (specificity of 90%) and an AUC of 0.923.³⁹

In this study, we first compared the differences between the indicators in the two groups through univariate analysis, then excluded variables with high covariance, and finally combined stepwise regression to screen variables to construct the model. The model was further visualised through R language and a website calculator. The prediction models we developed were able to achieve predictive performance comparable to, and in some cases better than, previous prospective cohorts. This study used Bootstrap resampling to internally validate the built model by sampling 1000 times and the results showed that the model had an AUC of 0.7886 at internal validation, suggesting that the model still had good discrimination and calibration correction and also showed that the model was still well calibrated. DCA is a useful method for model evaluation based on a continuous potential risk threshold (x-axis) and the net benefit of using the model to risk-stratify patients (y-axis) demonstrates the clinical utility of the model, with this study showing a high net benefit in the clinical use of the model.

The nomogram is a common predictive model used in many studies in clinical areas. In our study, a nomogram for assessing the risk of developing preeclampsia was created based on the predictor variables screened by a multifactorial logistic regression analysis: the score for each variable corresponds to the score on the topmost scoring axis and then the scores for each variable were summed to give a total score, with the total score corresponding to the value on the bottom axis being the probability of the risk of a severe occurrence of preeclampsia.

There are some limitations in this study. First, a major limitation of our study is a single-center study, which may impact the generalizability and validity of the medical records used. Secondly, external validation of the model is needed to evaluate its value in clinical practice. Thirdly, more machine learning methods are being applied to the construction of predictive models, and random forests and gradient enhancement also have good performance in building models. In the future, we will attempt more modelling methods to optimize the model.

Conclusion

In this study, it was found that UA, GGT, CYSCL and basic SBP were risk factors and BUN was a protective factor for PE in HIP at 24–29⁺⁶ weeks of gestation. According to the findings, we constructed a prediction model based on the information of routine pregnancy care that could accurately predict the risk of PE in HIP patients, which may provide a reference for clinicians to identify high-risk groups early.