A prediction model for cesarean delivery based on the glycemia in the second trimester: a nested case control study from two centers

Abstract Objective Maternal glycemia is associated with the risk of cesarean delivery (CD); therefore, our study aims to developed a prediction model based on glucose indicators in the second trimester to earlier identify the risk of CD. Methods This was a nested case-control study, and data were collected from the 5th Central Hospital of Tianjin (training set) and Changzhou Second People’s Hospital (testing set) from 2020 to 2021. Variables with significant difference in training set were incorporated to develop the random forest model. Model performance was assessed by calculating the area under the curve (AUC) and Komogorov-Smirnoff (KS), as well as accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). Results A total of 504 eligible women were enrolled; of these, 169 underwent CD. Pre-pregnancy body mass index (BMI), first pregnancy, history of full-term birth, history of livebirth, 1 h plasma glucose (1hPG), glycosylated hemoglobin (HbA1c), fasting plasma glucose (FPG), and 2 h plasma glucose (2hPG) were used to develop the model. The model showed a good performance, with an AUC of 0.852 [95% confidence interval (CI): 0.809–0.895]. The pre-pregnancy BMI, 1hPG, 2hPG, HbA1c, and FPG were identifies as the more significant predictors. External validation confirmed the good performance of our model, with an AUC of 0.734 (95%CI: 0.664–0.804). Conclusions Our model based on glucose indicators in the second trimester performed well to predict the risk of CD, which may reach the earlier identification of CD risk and may be beneficial to make interventions in time to decrease the risk of CD.


Introduction
Cesarean delivery (CD) is one of the most common surgical interventions during the childbirth in many countries [1].The global incidence of CD is 21.1%, and it has been estimated to reach 28.5% in 2030 [2].In China, the incidence is more than 40%, far exceeding the global incidence [3,4].Women undergoing CD may suffer from some complications, such as pain, urinary bladder injury, surgical site infection, and deep venous thrombosis, causing a delay for hospital discharge and bringing a burden to healthcare system [5][6][7].Earlier identifying pregnant women with risk of CD may decrease the occurrence of complications for mothers and infants [8].Considering this, it is of great importance to explore tools that can predict the risk of CD earlier for pregnant women in clinical practice.Some researchers have attempted to derive the tools that could provide clinicians with an easily available estimate of CD risk [9][10][11].Burke et al. have developed a model predicting the risk of CD by the maternal anthropometric data and fetal head and abdominal circumference before the onset of labor [12].Levine et al. focused on the demographic characteristics at delivery, including body mass index (BMI), height, Bishop score, and gestational age 40 weeks, to creat a prediction model estimating the risk of CD [13].Although their models show a good performance, they mainly used indicators before the onset of labor or during delivery.To decrease the incidence of CD, earlier prediction is important, which may be beneficial for clinicians and patients to take intervention measures earlier.The second trimester of pregnancy is a critical period for fetal growth, and also an important period for prenatal examination [14,15].The glucose level in the second trimester is a big concern for pregnant women [16].Maternal glucose level may mildly elevate at 24 to 28 weeks of gestation, which may increase the risk of adverse outcomes if not controlled [16].In the multinational Hyperglycemia and Adverse Pregnancy Outcomes (HAPO) study, we found that glucose level was associated with pregnancy outcomes [17].In a cohort study, Djomhou et al. found the close association between high glucose level and the risk of CD [18].So far, models predicting CD risk based on glucose indicators have not been reported.
In this study, we aim to develop a model to predict the CD risk for pregnant women based on the glucose indicators in the second trimester.We also perform an external validation to assess the prediction ability.

Data source and study population
This was a nested case-control study, and participants were collected from the 5th Central Hospital of Tianjin from April 2020 to June 2021 and Changzhou Second People's Hospital from November 2020 to May 2021.This study has been approved by the Ethics Committee of the 5th Central Hospital of Tianjin (WZX-EC-KY2021039) and Changzhou Second People's Hospital (YLJSC006).Written informed consent has obtained from participants or their family members.
Participants met al.l of the following criteria were included: (1) age 18 years; (2) with single pregnancy; (3) receiving prenatal examination and giving birth in the hospitals mentioned above; and (4) with complete clinical data.Participants met one of the following criteria were excluded: (1) with pre-gestational diabetes; (2) with pregnancy through ovulation induction, artificial insemination, or other assisted reproductive technology; (3) with organic or systemic diseases such as cardiovascular diseases, liver and kidney diseases, and blood system diseases before pregnancy; (4) suffering from thyroid dysfunction and other diseases that affect glycometabolism; (5) taking steroids, b-adrenergic receptor agonists, antipsychotic drugs and other drugs that may affect glycometabolism during the pregnancy; (6) with a history of CD; and (7) CD on maternal request.

Data collection
Data were collected as the following: (1) general data: gestational diabetes mellitus (GDM), nationality, pregnant age, marital status, pre-pregnancy body mass index (BMI), first pregnancy, employment situation, history of diabetes in a first-degree relative, history of full-term birth, history of preterm birth, previous abortion, history of livebirth, previous macrosomia, and previous GDM, and comorbidity; (2) laboratory data: fasting plasma glucose (FPG), 1 h plasma glucose (1hPG), 2 h plasma glucose (2hPG), glycosylated hemoglobin (HbA1c).Data on outcome (CD) were also collected.
Pre-pregnancy BMI was evaluated as weight (kg)/height 2 (m 2 ), and the weight and height of pregnant women were measured at the first prenatal examination [19].According to Chinese Adult BMI criteria, these participants were divided into underweight group (BMI < 18.5 kg/m 2 ), normal weight group (18.5 kg/m 2 BMI < 24 kg/m 2 ), and overweight group (BMI 24 kg/m 2 ) [19].
The glucose tolerance was tested through 75-g oral glucose tolerance test (OGTT) between 24-28 weeks' gestation.Samples were collected at fasting state and 1 h and 2 h following the glucose load, and glucose level was detected using glucose oxidase method [15,17].GDM was diagnosed if one or more of the following glucose levels were elevated: FPG 5.1 mmol/L, 1hPG 10.0 mmol/L, and 2hPG 8.5 mmol/L [20].

Development of the prediction model
The prediction model was developed using random forest in this study.Samples from the 5th Central Hospital of Tianjin (n ¼ 291) were selected as the training set, and samples from Changzhou Second People's Hospital (n ¼ 213) were selected as the testing set.Variables with significant difference between CD group and vaginal delivery (VD) group in training set were incorporated to develop the model.The model performance was evaluated using the area under the receiver operating characteristic (ROC) curve (AUC) and Komogorov-Smirnoff (KS), as well as accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV).The important variables were evaluated through feature importance computed by Gini importance.From the Random Forest structure, Gini importance was computed and the measure of the feature importance was determined according to the average over all trees in the forest.

Statistical analysis
The normally distributed measurement data were described as mean ± standard deviation (Mean ± SD) and compared using independent-samples t test between groups.The non-normally distributed measurement data were expressed as median [M (Q1, Q3)] and compared using Mann-Whitney U rank sum test between groups.The count data were expressed as number (n) and percentage (%), and intergroup comparison was performed using chi-squared or Fisher's exact test.The missing data on HbA1c in the testing set were addressed by multiple imputation, and sensitivity analysis was performed before and after imputation.The performance of our model in patients with or without GDM, with different ages (age 35 years and <35 years), and with different pre-pregnancy BMIs was also assessed.All statistical analyses were conducted using Python 3.7.3(Python Software Foundation, Delaware, USA) software, and p < .05 was regarded as statistical significance.

Baseline information of patients
Table 1 displays that a total of 504 women were enrolled according to the inclusion and exclusion criteria.In the training set, 291 women were included, with the mean FPG of 4.75 ± 0.49 mmol/L, the mean 1hPG of 8.29 ± 1.86 mmol/L, the mean 2hPG of 7.24 ± 1.59 mmol/L, and the mean HbA1c of 5.03 ± 0.43%.Of these women, incidence rate of CD was 36.43% (n ¼ 106).In the testing set, 213 women were included, and 4.23% of them (n ¼ 9) missed data on HbA1c, which was then addressed by multiple imputation.Sensitivity analyses showed that the results were consistent before and after imputation (p ¼ .968)(Supplementary Table S1).Significant difference in pregnant age, pre-pregnancy BMI, employment situation, history of diabetes in a first-degree relative, comorbidity, FPG, and 1hPG was observed between training set and testing set.

Difference analysis of training set
Table 2 shows the comparison of patients' characteristics between CD group and VD group in the training set.We found the significant difference in prepregnancy BMI, first pregnancy, history of full-term birth, history of livebirth, 1hPG, and HbA1c between the two groups (all p < .05).

Development and validation of the model predicting the risk of CD
Variables with significant difference between CD group and VD group in the training set were used to develop the model.Considering that women with abnormal glucose tolerance had a high risk of CD [21], FPG and 2hPG were also incorporated into the model.The AUC and KS value of this model was 0.852 [95% confidence interval (CI): 0.809-0.895]and 0.58, respectively, indicating the high discriminative ability of the model (Figure 1(A,B)).At a cutoff value of 0.36, accuracy, specificity, sensitivity, PPV, and NPV were 0.756 (95%CI: 0.707-0.805),0.670 (95%CI: 0.603-0.738),0.906 (95%CI: 0.850-0.961),0.611 (95%CI: 0.535-0.688),and 0.925 (95%CI: 0.881-0.970),respectively.Results were shown in Table 3.According to feature importance of this model, pre-pregnancy BMI, 1hPG, 2hPG, HbA1c, and FPG were identified as the more significant predictors for the risk of CD (Figure 2).

Discussion
CD brings some serious complications and causes a long-term effect on women's health [22,23], thereby earlier identifying the risk of CD and making interventions in time are important for the purpose of decreasing the incidence of CD.In this study, we mainly focused on the glucose indicators in the second trimester to develop a random forest model for the prediction of CD risk.Our results found that the model showed a good performance in predicting the risk of CD.According to the feature importance, prepregnancy BMI, 1hPG, 2hPG, HbA1c, and FPG were identified as the more significant predictors for CD.The model also performed well in women without GDM, with age < 35 years, and with overweight.
In this study, our model showed an AUC of 0.852, indicating that it performed well in the prediction of CD risk.Previously, some prediction models were developed to predict the risk of CD.The model developed by Burke et al. mainly focused on indicators before the onset of labor and the AUC was 0.69; compared to their model, our model was better in the discriminative ability [12].The model of Levine et al. was mainly used to predict the CD risk after an induction of labor, which may not be beneficial to adopt some interventions to improve the outcome [13].Our model payed attention to the important role of glucose indicators in the second trimester to reach the prediction of CD risk earlier.This can help clinicians make interventive measurements earlier to decrease the incidence of CD.Although there was significant difference in some variables in the baseline between training set and testing set, external validation showed the good performance of our model (AUC ¼ 0.734), indicating that our model had a wide applicability.In addition, our model performed well in patients without GDM, age < 35 years, and overweight.Also, our study identified 1hPG, 2hPG, HbA1c, FPG, and pre-pregnancy BMI as the more significant predictors for CD based on the feature importance.Maternal glycemia during pregnancy is a focus of attention because pregnant women are easier to occur gestational glucose intolerance, which is closely associated with the high risk of adverse pregnancy outcomes [24].Former studies have linked glucose level to adverse perinatal outcomes, and hyperglycemia increased the risk of CD [25,26].HbA1c was used to assess glycemic control in the past 2-3 months, and has been found to independently predict the risk of CD [27].Lemaitre et al. have also reported that high HbA1c level was associated with the higher CD risk [28].Consistently, our study identified HbA1c as one of the significant predictors for CD.In addition, real time OGTT blood glucose values appear to better reflect the dynamic glycemic status during the pregnancy compared to HbA1c [27].Studies have displayed that FPG, 1hPG, and 2hPG were associated with the risk of CD [25,29,30].In agreement with previous studies, our model showed 1hPG, 2hPG, and FPG were important predictors for CD.The pre-pregnancy BMI have been reported as important influence factors for CD [31].Similarly, our study found that prepregnancy BMI was an important predictor for CD.
There are some advantages in our study.First, we focus on the glucose indicators in the second trimester to develop a prediction model for CD, and the model displays a good performance to predict the risk of CD.Second, multiple imputation is used to handle the missing data to keep the statistical power and unbiased results.Third, the model performance has been validated by an external validation, and results are concordant, suggesting the reliability and wide applicability of the model.Also, some limitations in our study should be concerned.First, we only include women with single birth; thus, the predictive ability of this model for those with multiple birth is unclear.Second, the sample size in our study is relatively small, and larger sample size should be collected for further study in the future.Third, our study excludes women with pre-gestational diabetes.The performance of our model for women with pre-gestational diabetes is unknown.Fourth, women with history of CD are excluded, and further studies should be conducted in women with previous CD.Fifth, factors in the third trimester, such as glycemic control in the third trimester and overall fetal growth, are not considered in our study.In the future, studies considering factors in the third trimester are needed to further predict the risk of CD.

Conclusion
Our study developed and validated a prediction model for the risk of CD based on the glucose indicators in the second trimester, and the model displayed a good predictive performance.Applying this model, obstetricians may earlier predict the risk of CD, which is beneficial to carry out reasonable prenatal management earlier for women at high risk; thereby decreasing the risk of CD.Our model should be cautiously used in women with pre-gestational diabetes or previous CD, and further studies should be conducted in these populations.

Figure 1 .
Figure 1.Receiver operator characteristic (ROC) curve (A) and KS curve (B) in the training set.

Figure 2 .
Figure 2. The importance of the included variables.

Figure 3 .
Figure 3. Receiver operator characteristic (ROC) curve (A) and KS curve (B) in the testing set.

Table 1 .
Characteristics of included participants.

Table 2 .
Comparison of patients' characteristics between CD and VD groups.

Table 4 .
Performance of the prediction model based on GDM, age, and pre-pregnancy BMI.