Early Prediction Model for Preterm Birth Combining Demographic Characteristics and Clinical Characteristics

Objective. To create early prediction models for preterm birth (PTB) based on the Chinese population, combining demographic characteristics and clinical characteristics. Methods. A retrospective study on 15197 pregnant women who were recruited in Obstetrics and Gynecology Hospital of Zhejiang University from January1, 2017 to December 31, 2017. Demographic characteristics and clinical characteristics were collected and were randomly divided into the observation group (80%) and the validation group (20%). Multivariable Logistics regression analysis was performed to develop a risk prediction model in the observation group and the validation group. It was evaluated by the value of area under the curve (AUC) of receiver operating characteristic (ROC). Finally, we got a simple scoring system to present the preterm birth risk. Results. There were 1082 pregnant women (8.9%) developed PTB in the observation group and 316 pregnant women (10.3%) in the validation group. Gravidity, educational level, residence, previous history of PTB, twin pregnancy, pre-gestational diabetes mellitus (type I or II), chronic hypertension, placenta previa, gestational hypertension were signicant predictors of future PTB. These factors were all included in the model, the AUC was 0.746 with sensitivity of 61.4% (95%CI: 61.4-66.7%) and specicity of 86.6% (95%CI: 85.2-87.9%) at the threshold score of 8. Conclusion. PTB be predicted by demographic characteristics and clinical characteristics pre-pregnancy predicting as early as

Background WHO de nes preterm birth (PTB) as all births before 37 weeks of gestation, or fewer than 259 days from the rst date of a woman's last menstrual period 1 ,but in China, PTB means delivery after 28 gestational weeks but before 37 gestational weeks. PTB leads to perinatal mortality and morbidity, short-term or longterm complications like respiratory distress syndrome, bronchopulmonary dysplasia, necrotising enterocolitis, intraventricular hemorrhage, infections, feeding di culties, visual or hearing problems and so on 2 ., However prevention of PTB, improvement of the intrauterine condition before delivery, heightening the survival rate of newborns has become a challenging public issue.
A study published by Lacent found the incidence of PTB was 7.3 per 100 births [95% con dence interval (CI): 7.0-7.6%] or 6.7 per 100 live births (95%CI: 6.4-7.0%) during 2015 to 2016 in China 3 . The rising PTB rate has been reported by recent researches 4 , therefore how to predict and prevent PTB is particularly important. Most studies of PTB prediction involved ultrasound cervical length examination or biomarkers like fetal bronectin respectively or jointly 5 . Previous studies suggested that PTB was related with maternal demographic characteristics like age, social-economic factors, and clinical characteristics like hypertensive disordersin during or before pregnancy, sexually transmitted disease, history of miscarriage, stillbirth or PTB, multiple pregnancy, placenta abruption, placenta previa, ect 367 . An American study once established a prediction model of PTB in 2012 6 , but they served for one or multiple population as most studies did, and few studies provided risk assessment for an individual.
For the rst time, this paper combines demographic and clinical characteristics to establish a PTB prediction model and validate it based on a large population of Chinese. We established a simple PTB risk scoring system, a personalized PTB risk assessment, and a theoretical basis to prevent PTB.

Participants and data
This large retrospective study was conducted among 18000 pregnant women who did prenatal care and nally delivered at Women's Hospital School of Medicine Zhejiang University during January 1, 2017 to December 31, 2017. The exclusion criteria were the following: (1) Therapeutic induced labor because of stillbirth or birth defects (n = 246); (2) Not Han nationality (n = 34); (3) Women with incomplete information (n = 2523). Finally, a total of 15197 women remained in our study. QMmedical electronic medical records system and Union electronic medical records query system were utilized to obtain demographic and clinical characteristics information and telephone follow-up survey was conducted for patients with uncompleted obstetrics history. All women were randomly divided into the observation group and the validation group in our study, the observation group contained about 80% of all (n = 12131), whereas the validation group contained about 20% (n = 3066).

Clinical risk factors
According to previous studies of PTB, we gathered data including age,parity, gravidity, height, weight (before pregnancy), social-economic factors (career, educational level, residence), pre-pregnancy BMI, family history of diabetes mellitus (de ned as a rst-degree relative with diabetes mellitus type I or II), previous obstetrics history (PTB, spontaneous abortion, macrosomia, therapeutic induced labor because of stillbirth or birth defects), the complications of pregnancy (twin pregnancy, chronic hypertension, hypertensive disorders of pregnancy, pre-gestational diabetes mellitus (type I or II), gestational diabetes mellitus (GDM), placenta previa, scar uterus, uterine myoma) and other clinical data. Pre-pregnancy BMI was calculated as weight (kg) /(height (m))2. The pre-pregnancy BMI consisted of four groups by Chinese BMI standard: lean group (< 18.5 kg/m2), normal group (18.5-23.9 kg/m2), overweight group (24-27.9 kg/m2) and obese group (> 28 kg/m2). According to Chinese physical labor strength classi cation criteria (GB3869-1997), career was divided into "light labor" (such as staffs, civil servants etc.), "medium labor" (such as teachers, students, individual operators), "heavy labor" (such as workers, farmers, medical workers, police and soldiers etc.) and "others". Educational level was divided into low (lower than college, middle (college), and high (higher than college). Residence was divided into rural and urban areas. The clinical features were classi ed as positive and negative according to the presence or absence of the corresponding history.

Statistical analysis
For statistical analysis of the data SPSS statistical package for social sciences. Release 21.0 was used.
Univariate logistic regression analysis determined signi cant risk factors for PTB and reported an odds ratio (OR) for each risk factor (unadjusted); multivariate logistic regression identi ed the adjusted odds ratio (AOR) for all related factors in the scoring system. Bilateral P < 0.05 was considered statistically signi cant.

Clinical scoring system
Data from the observation group were used to develop a simple prediction scoring system. If women's clinical characteristics were signi cantly associated with the prevalence of PTB (P < 0.05) in univariate analysis, they would enter multivariate logistic regression. Clinical score for each risk factor was derived from rounded log adjusted odds ratio score following multivariate logistic regression, which ultimately create a clinical scoring system for independent PTB predictors.
Scoring system performance was examined at different score thresholds with intervals of 0.5 points in the validation group to estimate the optimal scoring threshold. Each score equal to or greater than the threshold score was assumed to be PTB, and then numbers of the assumed and real PTB conditions under each threshold score were separately counted, respectively. Sensitivity, speci city, positive and negative predictive values were calculated for each score, and the Wilson method estimated the 95% CI.
Performance of the clinical scoring system would be assessed by AUC.

Demographic and clinical risk factors
Of 12131 women in the observation group, 1082 women (8.9%) developed PTB. The correlation between each risk factor included in logistic regression and PTB at univariate and multivariate levels were shown in Table 1. And we can observe that gravidity, age, pre-pregnancy BMI, educational levels, residence, family history of diabetes mellitus, previous obstetrical history like PTB, spontaneous abortion and complications in pregnancy like twin pregnancy, chronic hypertension, pre-gestational diabetes mellitus (type I or II), placenta previa, scar uterus, gestational hypertension associated with PTB independently (P < 0.05). However, parity (P = 0.329), previous history of macrosomia (P = 0.181), fetal abnormality (P = 0.111), GDM in this pregnancy (P = 0.067), uterine myoma (P = 0.549), career (P > 0.05, except "light labor" or "others" groups) were not signi cantly correlated with PTB. CI: 2.6-4.1, P < 0.001), respectively. Interestingly, women who had a higher educational level (university or above) or lived in cities had less incidence of PTB (P < 0.001).

Clinical scoring system
After multiple logistic regression, the score for each clinical risk factor in the observation group was derived from the natural logarithm of AOR, multiplied by 10 for convenience ( Table 2). For example, more than two deliveries would receive 1.8 points on "gravidity", while a twin pregnancy would receive 27.7 points on "twin pregnancy", and so on. A pregnant woman, would receive a total score after counting 9 items like "gravidity","educational level", "residence", "previous history of PTB", "twin pregnancy", "pregestational diabetes mellitus (type I or II)", "chronic hypertension", "placenta previa", "gestational hypertension" which was a simple clinical scoring system for PTB.  Fig. 1 illustrates the trade-offs between sensitivity and speci city for the scoring system. The area under the curve is shown as 0.749 (95%CI: 0.732-0.767) (Fig. 1), indicating a 74.9% probability that a randomly selected patient with PTB will receive a higher risk score than a randomly selected patient without PTB. Our study found that twin pregnancy was the strongest predictor of PTB, and twin pregnancy was 16.1 times more likely to have PTB than singleton pregnancy (95% CI: 13.1-19.8, P < 0.001). Women with previous history of PTB had an increase risk to induce PTB than those without (OR = 2.9, 95%CI: 2.1-4.0, P < 0.001). Meta-analysis of 6 cohort studies showed that women with a history of PTB in twin pregnancy had a signi cantly higher risk of PTB in subsequent singleton pregnancy (OR = 3.34; 95% CI: 1.50-7.45).
The earlier delivered twin pregnancy the higher were chances of PTB in subsequent singleton pregnancy 9 .
In a cohort study involved 7633 pregnant women, Shen et al. 12 observed that both gestational hypertension (AOR = 1.8) and preeclampsia (AOR = 7.1) were signi cantly associated with PTB (P < 0.01), and they also found that folate had a protective effect on early-onset preeclampsia (P < 0.01). Premkumar et al. 13 discovered 8.8% women in 23425 singleton pregnancies developed PTB and 3.8% diagnosed with chronic hypertension. There were increased odds of PTB among women with chronic hypertension (AOR = 2.74, 95%CI:2.28-3.29, P < 0.001). In addition to this, a recent study published in The Lancet showed low-dose aspirin administered between 6 full gestational weeks and less than 14 gestational weeks among women from low-income and middle-income countries, would reduce the incidence of PTB before 34 weeks in pregnancy complicated with hypertensive disorders of (OR = 0.38, 95%CI:0.17-0.85, P = 0.015) 14] . For women with chronic hypertension or gestational hypertension, if they can identify the related high-risk factors, supply folate earlier, keep close prenatal care, control blood pressure during pregnancy and take low-dose aspirin, etc. may reduce the incidence of PTB 14 .
In addition, a retrospective study found that 46.8%(29/62) placenta previa cases were presented by PTB, whether it was spontaneous or iatrogenic 15 . Previous studies had discovered that the risk factors of placenta previa include multiparas, advanced maternal age (> 35 years old), prior cesarean delivery etc 16 .
Our study found that increasing maternal age (P = 0.008) and gravidity (P < 0.001), history of cesarean delivery (P = 0.041) heightened the PTB rate at univariate level respectively, but they were excluded by multivariate logistic regression nally. A recent Chinese study has also shown the incidence of PTB Interestingly, we also discovered that women who lived in rural area or had lower educational level (below college degree) were more likely to have PTB, the AOR were 1.59 (95%CI: 1.37-1.86, P < 0.001) and 1.45 (95%CI: 1.22-1.72, P < 0.005), respectively. A study from Norway pointed out that the risk of PTB was 1.5fold increase in single mothers or mothers with lower educational level 20 . In Michigan state, USA, every percent increase of unemployment during the rst trimester of pregnancy was related to a 3% increase of PTB 21 . A meta-analysis of 12 European cohort studies also revealed that women with lower educational level had a higher PTB rate, the most Signi cant of which were in Netherlands (P = 0.001) and Norway (P = 0.009). The incidence of PTB was 7.0% among women with lower educational level in Netherlands and 4.9% among those well-educated. Corresponding PTB rate was 9.7% and 5.9% in Norway women respectively 22 . A systematic review believed that this might be related to the limited data collection and the laggard facilities in economic backward areas 23 . We presume that well-educated women and urban women may pay more attention to their perinatal health,may be more likely to attend routine or extra prenatal care persistently. They have access to medical resources when needed, and most of them can get better nutrition during pregnancy. For example, our hospital supplies several free classes to pregnant women per week, which covers some general knowledge of pregnancy complications, home care for women or newborns and so on. We also provide remote fetal heart monitoring services to women who demand, this may decrease the occurrence of PTB to a certain degree.
All information our prediction model may need, can be easily obtained from women's history and demographic characteristics without blood tests or other invasive inspections (except for examinations that are necessary to assess the severity of pregnancy complications). Therefore, this makes it convenient, practical, e cient and non-invasive. This prediction model helps to identify high PTB risk group or individual a during early pregnancy even before pregnancy. The risk of PTB increases according to the rising (or high), scores especially for those with scores greater than 15 points, the risk of PTB is as high as 50%. Close pregnancy monitoring or targeted intervention for women with high PTB risk ,may reduce the incidence of PTB. However, there are many limitations in our research. Firstly, it's prone to memory bias being a retrospective study, information bias due to the exclusion of some samples with incomplete information. In addition, our population was concentrated in a large specialist women hospital rather than a community hospital, which may increase the number of women with higher risks.
Most of the population sample lives in Hangzhou or surrounding cities in Zhejiang Province, so we cannot represent all regions or ethnicities. Although previous study thought that it might be more e cient when combining demographic and clinical characteristics with cervical length data 8 , we failed to obtain it because our hospital didn't take cervical length measurement as a routine perinatal monitoring content.
Thus, future prediction model can consider to include some laboratory indicators and cervical length data before or during pregnancy, which may improve the accuracy and e cacy of the prediction model, besides that, they can verify whether the existing prevention for PTB is effective or not.

Conclusion
In summary, our prediction model predicts the possibility of PTB based on the demographic and clinical characteristics like gravidity, educational level, residence, previous history of PTB, twin pregnancy, pregestational diabetes mellitus (type I or II), chronic hypertension, placenta previa, gestational hypertension. We can identify high PTB risk group or individual during early trimester even before pregnancy.
Reinforcing intensive perinatal care and preventive measures speci c to each risk factor may help reduce PTB and improve pregnancy outcomes.