Prediction of premature rupture of fetal membranes using deep learning in East China

Abstract

If premature rupture of fetal membranes (PROM) can be forecasted, doctors can formulate individualized medical treatment plans and optimize the utilization efficiency of unevenly distributed medical resources in China to lower the odds of PROM, preterm birth, and neonatal mortality. We collected the medical records of 20,392 dyads of mothers and term-birth neonates who had received prenatal care services from January 1, 2014 to December 31, 2019 in Hangzhou, Zhejiang province, East China. According to participants’ home and working addresses, maternal exposure to air pollution and meteorological conditions was estimated. Deep learning was used to predict the odds of PROM occurrence. The efficiency of Large Language Model—DeepSeek was tested in healthcare settings. Of 32 clinical covariates have been identified to be statistically significantly associated with PROM, 25 variables—7 positively and 18 negatively linked to PROM—can be detected at least one week before PROM or delivery. Using the Bonferroni correction as a stricter classification tool, 10 out of 32 clinical covariates were statistically associated with PROM. Air pollution exposure and meteorological conditions that were associated with PROM were identified. Based on these findings, approximately 86.1% of PROM cases can be forecasted using deep learning. Thus, individualized treatment can be crafted and vital medical resources can be allocated in advance. DeepSeek can facilitate healthcare processes and optimization of medical resources, showing its uniformity, thoroughness, and robustness. However, the improvement of prediction accuracy for PROM was accompanied by increasing false-positive cases, which is a paradox that needs to be solved.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-48769-9.

Introduction

Premature rupture of membranes (PROM) refers to the fetal membrane rupture before the onset of delivery¹. Among different nations, the incidence of PROM ranged from 5.0%² to 18.7%³. PROM causes approximately one-third of preterm birth (PTB)⁴, which has been the leading factor of infant morbidity and mortality worldwide^5,6. Approximately 50% of neonatal deaths and 75% of all perinatal mortality were ascribed to PTB in recent decade⁷, in part due to the lung hypoplasia⁸, implying PROM accounted for approximately 17% of neonatal mortality and 25% of perinatal deaths worldwide. Those neonates who have survived PROM often suffer from multiple diseases, including cerebral palsy, delays in development, impaired immune function, reduced respiratory function, and sensory disturbance⁹.

Inflammation reaction, cell apoptosis, collagen metabolism, and oxidative stress may lead to PROM¹. Amniotic-fluid infection is also an important cause of PROM¹⁰. Microbes (e.g., Ureaplasma) can reach the cervix and cause membrane rupture^11,12 Physiological and physical status of pregnant women, such as short cervical length, premature birth that occurred during previous pregnancy, positive fetal fibronectin screening¹³, pre-pregnancy weight¹⁴, cigarette smoking, vaginal bleeding during pregnancy, and genitourinary infection¹⁵, also plays an important role in PROM incidence. Maternal serum amino acid (e.g., lysine, glycine, glutamic acid) levels are associated with occurrence of PROM¹⁶. Administrations of prophylactic antibiotics^17–19, probiotics²⁰, and corticosteroids²¹ are common avenues to lower the risks of PROM.

Environmental and meteorological factors may also trigger PROM. An increasing number of studies pointed out that PROM was associated with maternal exposure to ambient air pollutants, including SO₂²², NO₂²³, O₃^24,25, CO²⁶, PM_2.5^27–32, and PM₁₀³³. However, in another report, there was no significant association found between maternal PM_2.5 exposure and PROM³⁴. Meteorological conditions³⁵, such as heatwaves^36,37, cold spell³⁸, diurnal temperature variation³⁹, wind speed, humidity⁴⁰, and barometric pressure^41,42, were linked to PROM. However, current investigations have mainly focused on only one variable or limited risk factors that may cause PROM. There is limited evaluation of the combined impacts of multiple contributing factors (e.g., maternal physiological characteristics, air pollutants, and meteorological conditions) on PROM in China so far.

In recent decade, machine learning (ML) has made breakthrough progress in predicting diseases⁴³. In the field of public health, it is difficult to conduct large-scale randomized double-blind experiments on humans. Causal ML (CML) can deal with multi-dimensional and unstructured datasets and offer flexible, data-driven approaches for projecting the outcome of treatment, thus making personalized treatment strategies possible⁴⁴. For instance, genetic polymorphisms are related to inter-individual differences in in efficacy of medications⁴⁵, and CML can learn these differences, assess the heterogeneity in treatment among patients, and determine the optimal treatment plan⁴⁴.

In this study, we investigate the association of PROM with clinical covariates, air pollution exposure and meteorological conditions, establish a CML-based pathway to predict the incidence of PROM, test the reliability and robustness of the proposed approach, and proffer a method to personalize the treatment for reducing the risks of PROM, improving point-of-care practice, and enhancing the utilization efficiency of medical resources. Due to data limitation on preterm PROM, this study focuses on term PROM.

Methods

Study population

We collected the electronic health records of 21,646 dyads of mothers and neonates who were admitted for medical care and maternity services from January 1, 2014 to December 31, 2019 in southern Zhejiang. All participants’ home and working addresses were in southern Zhejiang province. Each clinical entry record contained a full-cycle procedure of child-bearing, laboratory infection test result, diagnosis of diseases, and prescription of medical treatments. The overall PTB rate was 5.8%.

The datasets used for the analysis encompassed ten categories of data: (i) maternal personal covariates; (ii) clinical conditions; (iii) habits during pregnancy; (iv) diseases; (v) maternal medication usage; (vi) home and working addresses; (vii) neonatal information; (vii) air pollution levels; and (ix) meteorological conditions. This study was approved by the ethics committee of School of Medicine, Zhejiang University (Trial No. 2024 − 1181). All study procedures were conducted in accordance with the guidelines of the Declaration of Helsinki⁴⁶.

Maternal and neonatal covariates

The maternal physiological and physical covariates encompassed nine variables, including age, height, weight, body mass index (BMI), gravidity, parity, gestational age (GA), number of abortions, and cervical cerclage surgery. Maternal personal conditions included eight variables, such as intraoperative bleeding, pregnancy mode including natural, rtificial insemination and In Vitro Fertilization (IVF), delivery mode, body temperature on admission, leukocyte count, neutrophil granulocyte count, C reactive protein (CRP) level, and erythrocyte count. Maternal habits had two variables, including tobacco smoking and alcohol consumption, which were self-reported. Maternal comorbidities and complications had fourteen variables, including polyhydramnios, maternal diabetes, gestational diabetes mellitus (GDM), preeclampsia, intrahepatic cholestasis syndrome, chronic hypertension, hyperlipidemia, polycystic ovary syndrome (PCOS), uterine fibroids, PROM, gonococcal infection, Mycoplasma hominis infection, Ureaplasma urealyticum infection, and Group B streptococcus (GBS) infection. Maternal medication treatment had ten variables, including the usage of antibiotics, dexamethasone (DXM), magnesium sulphate, ritodrine, nifedipine, m-phenyltriphenylene, indomethacin suppository, dydrogesterone, and progesterone. Home address included two variables: urban and rural. Neonatal covariates had seven variables, including intrauterine infection (e.g., umbilical cord blood infection, placental infection, and amniotic fluid infection), fetal intrauterine distress, neonatal asphyxia, and APGAR (activity, pulse, grimace, appearance and respiration) scores (1 and 5 min after birth), sex, and birthweight. Small for GA (SGA), large for GA (LGA), and appropriate for GA (AGA) were defined as birthweight less than 10th, over 90th, and between the 10th and 90th percentiles, respectively.

Air pollution exposures and meteorological conditions

Hourly meteorological data at the city center of Hangzhou for the period between January 1, 2013, and December 31, 2019, including temperature, relative humidity (RH), wind speed, wind direction, rainfall, and atmospheric pressure, were obtained from the Hangzhou Meteorological Bureau. We also acquired hourly ambient air pollution data, including the levels of SO₂, NO₂, O₃, CO, PM_2.5, and PM₁₀, from six environmental monitoring stations in Hangzhou between January 1, 2013 and December 31, 2019⁴⁷. Synchronized hybrid ambient real-time particulate analyzers (SHARP 5030i) from Thermo Scientific™ were used to detect the atmospheric concentrations of PM_2.5 and PM₁₀. The atmospheric concentrations of SO₂, NO₂, O₃, and CO were measured using Model 450i, Model 42i, Model 49i-PS, and Model 48i-TLE from Thermo Scientific™, respectively. We chose air pollutant data from the atmospheric monitoring station closest to the participant’s home address to calculate the exposure level. The pregnancy was divided into three trimesters. First trimester was defined as the period from the LMP to the end of the 12th week of pregnancy; the second trimester was the period between the 13th week + 1 day and 27th week + 6 days; and the third trimester included the period from the 28th week and beyond.

Statistical analysis and machine learning

Statistical significance was defined as the two-tailed p-value less than 0.05. The association between the two datasets was perceived as statistically significant when the p-value was smaller than 0.05. Statistical analyses were performed in Statistical Product and Service Solutions (SPSS) software package (https://spssau.com/) and R (the R Foundation for Statistical Computing; version 4.5.1). A Logistic Regression Model (LRM) was used to fit the association of PROM with clinical covariates, air pollution exposures, and meteorological conditions. The regression coefficient (β-value) and its 95% confidence interval (CI) were generated via SPSS. β-value is statistically meaningful only under the condition of p-value less than 0.05. Restricted Cubic Spline (RCS) curve was used to elucidate odds ratio (OR) and relevant 95% CI. Multiple comparisons using Bonferroni correction (0.05/32 clinical indicators = 0.0015625) were applied to minimize the risk of obtaining false positive results. According to the Bonferroni correction, p-values less than 0.0015625 were considered statistically significant, while those between 0.0015625 and 0.05 were deemed suggestive significant. We utilized packages rms (6.2-0.2) to fit the LRM, dplyr (1.1.4) to manipulate data, and ggplot2 (3.5.1) to generate graphs.

Three machine learning models including Random Forest (RF), Multi-layer Perceptron (MLP), and XGBoost were used to predict PROM. For the MLP regressor, we used four hidden layers with ReLU activation, each containing 64 units. The optimizer was applied with a constant learning rate of 0.001, and the number of iterations was set to be 5000. For the XGBoost regressor, the maximum depth of trees was set to be 6. We set the learning rate at 0.3, and the number of iterations was 100. For the RF regressor, we used 10 trees without maximum depth. The minimum samples required to split an internal node was set to be 2, and the leaf node was set to be 1; the criterion was Mean Squared Error (MSE). The taxonomy of maternal and neonatal covariates was incorporated as dichotomous or tripartite variables, shown in Table S1. For example, we converted DXM usage as a binary variable (used = 1; not used = 0). The datasets were randomly classified into training (90%) and testing (10%).

A Model Card was provided in the Supplementary Materials. There were no potential medical dataset biases in this study, because all medical data of pregnant women and neonates were carefully recorded. But there were potential dataset biases in maternal exposure to air pollutants and meteorological conditions, because we have only known their home and working addresses, but did not know the exact location where pregnant women were at every moment of their daily lives. We did not conduct data cleaning and preprocessing steps. We did not combine data from multiple sources. The method of data splitting was random. The data splitting could mimic anticipated real-world applications. The data splitting procedure was not chosen to avoid data leakage.

Model evaluation and validation

To avoid data leakage, all predictive modeling procedures were conducted using leakage-safe data splitting. Model training, hyperparameter tuning, and preprocessing were performed exclusively on the training set, while model performance was evaluated on an independent held-out test set.

Model discrimination was assessed using the area under the receiver operating characteristic curve (AUROC) and the area under the precision–recall curve (AUPRC), which is particularly informative for imbalanced outcomes such as PROM. Model calibration was evaluated using the Brier score, calibration intercept and slope, and calibration plots.

Classification performance was further evaluated by reporting confusion matrices, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) at a predefined probability threshold of 0.5. Additional threshold-free evaluation was conducted using ROC and precision–recall curves.

The DeepSeek

The name of the Large Language Model used in this study is the DeepSeek, which can be obtained online (https://www.deepseek.com/). The version we used was R1. The manufacturer of the DeepSeek was Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd. The date of use of the DeepSeek was September 5, 2025. The authors of this study are fully responsibility for the contents from the DeepSeek.

Results

Descriptive clinical statistics

Table 1 shows the basic clinical statistics of 20,392 dyads of mothers and neonates who were born with term birth (TB). The mean maternal age, gravidity, parity, height, weight, and BMI were 30.01 ± 3.87 years, 1.92 ± 1.08 times, 1.38 ± 0.50 times, 160.79 ± 4.64 cm, 67.97 ± 8.21 kg, and 26.28 ± 2.93, respectively. Defined as less than 28 completed weeks of GA⁴⁸, extremely PTB (EPTB) incidence was 1.6%. The numbers of neonates who had gestational age (GA) between 37 and 42 weeks, between 28 and 37 weeks, and less than 28 weeks were 20,392 (94.2%), 902 (4.2%), and 352 (1.6%), with PROM rates of 13.8%, 35.8%, and 88.9%, respectively. 12,860 mothers were nulliparous before this delivery. Average GA for all TBs was 39.08 ± 1.07 weeks. During delivery, pregnant women who did not experience from PROM had 37.7% more intraoperative bleeding compared to those who suffered from PROM on average. Of 13 women who got pregnant using artificial insemination, 12 suffered from PROM and only 1 did not. There was only one woman who smoked tobacco during pregnancy; she got PROM. Of 143 women who consumed alcohol during pregnancy, 138 suffered from PROM and 5 did not. Prevalences of maternal diabetes and GDM among PROM women—1.9% and 1.2%—were much lower than non-PROM women—13.3% and 9.7%, respectively. According to erythrocyte count, anemia rates among PROM and non-PROM women were 18.9% and 19.0%, respectively. Of 26 women who had DXM treatment, 23 suffered from PROM and 3 did not. Of 2814 women who had PROM, 1.0% got infections and 4.8% received medical treatment (excluding antibiotics). Of 17,578 non-PROM women, 0.9% got infections and 4.2% received medical treatment (excluding antibiotics). Medical treatments of antibiotics, DXM, magnesium sulphate, ritodrine, nifedipine, indomethacin, and progesterone were mainly used for pregnant women who had a shortened cervix and optimization of neonatal health.

Table 1.

Basic clinical statistics of participants with TB.

Variables	PROM (n = 2814)	Non-PROM (n = 17578)
Maternal age (years)	29.55 ± 3.73	30.08 ± 3.89
Maternal height (cm)	161.05 ± 4.56	160.75 ± 4.65
Maternal weight (kg)	67.81 ± 8.03	67.99 ± 8.23
Maternal BMI (kg/m2)	26.13 ± 2.90	26.30 ± 2.93
Gravidity	1.55 ± 0.76	1.98 ± 1.11
Parity	1.30 ± 0.49	1.39 ± 0.51
Number of previous miscarriages	0.24 ± 0.61	0.59 ± 0.87
GA (weeks)	38.92 ± 1.04	39.10 ± 1.08
Cervical cerclage surgery	N = 0 (0%)	N = 9 (0.05%)
Intraoperative bleeding (ml)	219.4 ± 152.4	302.01 ± 143.96
Delivery mode	Natural: 76.8%; Cesarean: 21.3%; Vagina assisted: 1.9%	Natural: 52.8%; Cesarean: 46.9%; Vagina assisted: 0.3%
Body temperature on admission	Lower than 37.5 °C: 99.5%; Higher than 37.5 °C: 0.5%	Lower than 37.5 °C: 95.8%; Higher than 37.5 °C: n = 4.2%
Leukocyte count (×109/L)	9.73 ± 2.85	10.19 ± 2.40
Neutrophil granulocytes (×109/L)	7.47 ± 2.77	8.40 ± 2.79
CRP (mg/L)	7.06 ± 15.16	7.71 ± 12.88
Erythrocyte count (g/L)	Larger than 110: N = 2281 (81.1%) Between 60 and 110: N = 533 (18.9%)	Larger than 110: N = 14,233 (81.0%) Between 60 and 110: N = 3326 (18.9%) Less than 60: N = 19 (0.1%)
Smoking tobacco during pregnancy	N = 1	N = 0
Alcohol consumption during pregnancy	N = 138 (4.9%)	N = 5 (0.03%)
Antibiotic	N = 2804 (99.6%)	N = 17,318 (98.5%)
DXM	N = 23 (0.8%)	N = 3 (0.02%)
Magnesium sulphate	N = 84 (3.0%)	N = 572 (3.3%)
APGAR-1 scores (one minute after delivery)	8.02 ± 3.24	9.84 ± 0.92
APGAR-5 scores (five minutes after delivery)	9.29 ± 1.71	9.93 ± 0.66
Infant sex	Male: N = 1501 (53.3%) Female: N = 1313 (46.7%)	Male: N = 9064 (51.6%) Female: N = 8514 (48.4%)
Male neonatal birthweight (g)	3379.93 (N = 1501)	3444.76 (N = 9064)
Female neonatal birthweight (g)	3286.45 (N = 1313)	3344.07 (N = 8514)
From admission to delivery	18.88 ± 17.92	38.49 ± 29.24
From delivery to hospital discharge	92.01 ± 28.97	70.99 ± 19.44
From admission to discharge	110.90 ± 27.12	109.47 ± 35.29
Home address	Urban: N = 1672 (59.4%) Rural: N = 1142 (40.6%)	Urban: N = 10,382 (59.1%) Rural: N = 7196 (40.9%)
Working address	Urban: N = 1828 (65.0%) Rural: N = 986 (35.0%)	Urban: N = 11,456 (65.2%) Rural: N = 6122 (34.8%)

Numbers of male and female TB neonates accounted for 51.8% and 48.2%, respectively. Average male and female TB neonatal birthweight with mothers who experienced PROM was 3379.93 and 3286.45 g, respectively. Average male and female neonatal birthweight who had non-PROM mothers was 3444.76 and 3344.07 g, respectively. The numbers of neonates who were born in 37th, 38th, 39th, 40th, 41 st, and 42nd weeks of GA were 1367 (6.7%), 4931 (24.2%), 6937 (34.0%), 5100 (25.0%), 2014 (9.9%), and 43 (0.2%), respectively. The thresholds of SGA and LGA male and female neonates for each GA are shown in Table S2. As GA increased from 37 to 42 weeks, the thresholds of SGA and LGA male neonates increased from 2650 to 3700 to 3250 and 4250 g, respectively; the thresholds of SGA and LGA female neonates increased from 2600 to 3650 to 3200 and 4000 g, respectively. The average time interval between admission and delivery, delivery and hospital discharge, and admission and hospital discharge were 35.98 ± 29.07, 73.68 ± 22.05, and 109.66 ± 34.36 h, respectively.

Association between PROM and clinical indicators

The associations between PROM and medical indicators are shown in Table 2. The Euler’s number e to the power of the β-value is equal to odds ratio (OR). According to p-value, we identify 32 clinical covariates—8 positively and 24 negatively—that were statistically significantly associated with PROM. Seven clinical covariates that were significantly positively linked to PROM, including height (Fig. 1), pregnancy mode, tobacco smoking, alcohol consumption, antibiotics use, DXM use, and GBS, can be known prior to the onset of delivery. 18 out of 24 clinical covariates that were significantly negatively correlated with PROM, which were maternal age (Fig. 2), BMI (Fig. 3), gravidity (Fig. 4), parity, previous abortions (Fig. 5), GA (Fig. 6), body temperature, leukocyte count (Fig. 7), neutrophil granulocytes (Fig. 8), CRP (Fig. 9), hypertension, chronic maternal diabetes, GDM, preeclampsia, intrahepatic cholestasis syndrome, hyperlipidemia, uterine fibroids, and fetal intrauterine distress, can be detected before delivery. Neonatal birthweight (Fig. 10) and intraoperative bleeding (Fig. 11) were also associated with PROM, but they cannot be detected prior to delivery. As Fig. 1 indicates, if maternal height was between 145 and 165 cm, the odds of PROM increased as maternal height increased; if maternal height was higher than 165 cm, the odds of PROM decreased as maternal height increased. Figure 2 shows that when maternal age was under 27-year-old, the odds of PROM increased as maternal age increased; when maternal age was between 27 and 34, the odds of PROM decreased as maternal age increased; when maternal age was greater than 34, the odds of PROM remained roughly stable. Figure 3 illustrates that a linearity between odds of PROM and BMI; when BMI increased, the odds of PROM declined steadily. Figures 4 and 5 show that the odds of PROM moved firmly downwards as gravidity and previous abortions increased, respectively. Figure 6 exhibits that the odds of PROM decreased as GA increased when GA was less than 38.5 weeks and more than 39.6 weeks; however, between 38.5 and 39.6 weeks, the odds of PROM increased as GA increased. Figures 7 and 8, and Fig. 9 indicate that the associations of the odds of PROM with leukocyte count, neutrophil granulocytes, and CRP, respectively, all show a U-shape.

Table 2.

Association between PROM and clinical covariates.

Category	Variables	p-value	β-value	95% CI of β-value
Maternal personal covariates	Maternal age (years)	p < 0.001	− 0.004	− 0.005 to − 0.003
	Maternal height (cm)	p = 0.001	0.002	0.001 to 0.003
	Maternal weight (kg)	p = 0.261
	Maternal BMI (kg/m2)	p = 0.003	− 0.002	− 0.004 to − 0.001
	Gravidity	p < 0.001	− 0.044	− 0.048 to − 0.040
	Parity	p < 0.001	− 0.041	− 0.050 to − 0.032
	Previous abortions	p < 0.001	− 0.057	− 0.062 to − 0.051
	GA (weeks)	p < 0.001	− 0.019	− 0.023 to − 0.015
	Cervical cerclage	p = 0.230
Maternal indicators	Intraoperative bleeding (ml)	p < 0.001	− 0.000	− 0.000 to − 0.000
	Pregnancy mode	p < 0.001	0.173	0.104 to 0.243
	Delivery mode	p < 0.001	− 0.105	− 0.114 to − 0.095
	Body temperature	p < 0.001	− 0.124	− 0.149 to − 0.099
	Leukocyte count (×109/L)	p < 0.001	− 0.009	− 0.011 to − 0.007
	Neutrophil granulocyte (×109/L)	p < 0.001	− 0.014	− 0.016 to − 0.012
	CRP (mg/L)	p = 0.017	− 0.000	− 0.001 to − 0.000
	Erythrocyte count	p = 0.807
Maternal lifestyle during pregnancy	Tobacco smoking	p = 0.012	0.862	0.186 to 1.538
	Alcohol consumption	p < 0.001	0.833	0.777 to 0.888
Maternal comorbidities and complications	Polyhydramnios	p = 0.689
	Maternal hypertension	p < 0.001	− 0.130	− 0.194 to − 0.066
	Maternal diabetes	p < 0.001	− 0.131	− 0.146 to − 0.117
	GDM	p < 0.001	− 0.134	− 0.150 to − 0.119
	Preeclampsia	p < 0.001	− 0.139	− 0.205 to − 0.072
	Intrahepatic cholestasis syndrome	p < 0.001	− 0.123	− 0.162 to − 0.084
	Hyperlipidemia	p < 0.001	− 0.115	− 0.162 to − 0.068
	PCOS	p = 0.738
	Uterine fibroids	p < 0.001	− 0.133	− 0.187 to − 0.078
Microbial infection	GBS	p = 0.008	0.307	0.081 ~ 0.532
	Mycoplasma hominis infection	p = 0.537
	Ureaplasma urealyticum infection	p = 0.689
	Gonococcal infection	p = 0.276
Intrauterine infection	Amniotic fluid (checked microbial positive)	p = 0.167
	Umbilical cord blood infection (checked microbial positive)	p = 0.106
	Placental infection (checked microbial positive)	p = 0.130
Clinical and medical treatments	Antibiotic	p < 0.001	0.102	0.061 to 0.144
	DXM	p < 0.001	0.748	0.615 to 0.880
	Magnesium sulphate	p = 0.453
	Ritodrine	p = 0.741
	Nifedipine	p = 0.144
	m-Phenyltriphenylene	p = 0.076
	Indomethacin suppository	p = 0.233
	Dydrogesterone	p = 0.572
	Progesterone	p = 0.388
Neonatal information	Fetal intrauterine distress	p < 0.001	− 0.128	− 0.197 to − 0.059
	Neonatal asphyxia	p = 0.135
	APGAR-1 score (one minute after delivery)	p < 0.001	− 0.084	− 0.087 to − 0.081
	APGAR-5 score (five minutes after delivery)	p < 0.001	− 0.090	− 0.095 to − 0.085
	Infant sex	p = 0.080
	Birthweight	p < 0.001	− 0.000	− 0.000 to − 0.000
Time interval (hours)	From admission to delivery	p < 0.001	− 0.003	− 0.003 to − 0.002
	From delivery to leave the hospital	p < 0.001	0.005	0.005 to 0.005
	From admission to leave the hospital	p = 0.065
Address	Home (urban or rural)	p = 0.827
	Work (urban or rural)	p = 0.722

Fig. 1.

Association between maternal height and PROM.

Fig. 2.

Association between maternal age and PROM.

Fig. 3.

Association between maternal BMI and PROM.

Fig. 4.

Association between gravidity and PROM.

Fig. 5.

Association between previous abortion times and PROM.

Fig. 6.

Association between GA and PROM.

Fig. 7.

Association between leukocyte count and PROM.

Fig. 8.

Association between neutrophil granulocytes and PROM.

Fig. 9.

Association between CRP and PROM.

Fig. 10.

Association between Neonatal birthweight and PROM.

Fig. 11.

Association between intraoperative bleeding and PROM.

During regression modeling study, 12 (age, gravidity, height, weight, BMI, miscarriage, GA, intraoperative bleeding, neonatal birthweight, erythrocyte count, neutrophil granulocyte count, and CRP) out of the 32 clinical indicators fitted logistic regression model. Among them, 10 (age, gravidity, height, miscarriage, GA, intraoperative bleeding, neonatal birthweight, erythrocyte count, neutrophil granulocyte count, and CRP) showed significant (p < 0.0015625) association with PROM, whereas the indicator BMI showed suggestive significant (p < 0.05). Most of the regression curves of PROM and the clinical indicators were nonlinear (p-nonlinear < 0.05), except the neonatal birthweight (p-nonlinear = 0.48) and the BMI (p-nonlinear = 0.72).

China has strict laws on maternal and child healthcare: for every two weeks during the third trimester of pregnancy, doctors must conduct detailed physical examinations on each pregnant woman and fetus. Therefore, we perceive that these 25 clinical covariates that were statistically significantly associated with PROM can be measured at least two weeks before delivery. The statistically significant association of PROM with maternal age, BMI, gravidity, parity, GA, abortion times, and previous mode of delivery was also found in Egypt⁴⁹. We acknowledge that risk factors such as preeclampsia and fetal growth restriction may increase the risk of iatrogenic PTB rather than PROM. We also acknowledge that PCOS, fibroids, ICP, GDM (unless with polyhydramnios), hypertension are not typically associated with PROM. The association does not necessarily reflect causality. Here we only show the link between PROM and potential risk factors, but we have not demonstrated their casual links.

Association between PROM and air pollution exposure

Table S3 shows the association between PROM and long-term maternal exposure to air pollution. There was no significant association between PROM and long-term maternal exposure to SO₂ and CO. PROM was significantly associated with maternal exposure to O₃ and PM₁₀ during the first trimester, maternal exposure to NO₂, PM_2.5, and PM₁₀ during the second trimester, maternal exposure to NO₂, O₃, PM_2.5, and PM₁₀ during the whole pregnancy. For every 10 µg/m³ increase in maternal NO₂ exposure during the second trimester and the entire pregnancy, PROM incidence increased by 0.4% (95% CI: 0–0.8%) and 1.0% (95% CI: 0.3–1.7%), respectively.

For every 10 µg/m³ increase in maternal O₃ exposure during the entire pregnancy, PROM incidence increased by 0.7% (95% CI: 0.2–1.1%). For comparison, every 10 µg/m³ increase in the 3-day average concentration (lag02) of O₃–8 h was linked to an increment in PROM of 5.42% in Chinese province of Henan²⁴. For every 10 µg/m³ increase in maternal PM_2.5 exposure during the second trimester and the entire pregnancy, PROM incidence increased by 0.4% (95% CI: 0.1–0.7%) and 1.0% (95% CI: 0.4–1.5%), respectively; for every 10 µg/m³ increase in maternal PM₁₀ exposure during the second trimester and the entire pregnancy, PROM incidence increased by 0.3% (95% CI: 0–0.5%) and 0.6% (95% CI: 0.2–1.0%), respectively. For comparison, an increase of 10 µg/m³ in SO₂, NO₂ and O₃ was linked to increased odds of PROM by 8.74%, 3.09%, and 1.68%, respectively; no significant association between PM_2.5 and PROM was detected²². Maternal PM_2.5 exposure may raise the odds of PROM by as much as 15%¹. Maternal exposure to PM₁₀ during the third trimester was linked to increased risks of PROM³³.

Then we calculated the association between maternal exposure to air pollution weeks before delivery and PROM, shown in Table S4. Average maternal SO₂ exposure levels one week and two, three, and four weeks before delivery were all positively significantly linked to PROM. Maternal NO₂ exposure from three and four weeks before delivery to delivery was positively significantly linked to PROM. Maternal PM_2.5 and PM₁₀ exposure four weeks before delivery was positively significantly associated with PROM. SO₂ level was not significantly correlated with PROM in terms of trimester exposure, indicating the impact of short-term exposure to SO₂ before delivery may be more profound than long-term exposure.

Association between PROM and meteorological conditions

Table S5 sketches the association between PROM and meteorological conditions. PROM was not statistically significantly linked to temperature, rain, air pressure, and wind speed. PROM was significantly negatively associated with RH during the first and the third trimesters and the entire pregnancy. For every 10 units decrease in RH during the first trimester, the third trimesters, and the entire pregnancy, the odds of PROM increased by 1.4% (95% CI: 0.2–2.7%), 2.1% (95% CI: 0.8–3.5%), and 3.7% (95% CI: 1.6–5.8%), respectively. Table 3 shows the meteorological impacts on PROM estimated by our study and other investigations. Then we calculated the association between meteorological conditions weeks before delivery and PROM, shown in Table S6. Average RH from three and four weeks before delivery were significantly negatively linked to PROM. Average wind speed from three and four weeks before delivery until delivery were significantly positively correlated with PROM.

Table 3.

Meteorological impacts on PROM.

Variable	Meteorological condition	PROM odds	Period	Cohort	Reference
Temperature	1 °C increase during the week before delivery	4% increase in term PROM risk	2002–2008	15,381 singleton pregnancies with PROM in the United States	35
Temperature	Non-intense heatwaves	9–14% increase in PROM risks	2008–2018	190,767 subjects with 16,490 (8.6%) spontaneous PROMs in the United States	36
Temperature	Exposure to extreme cold (below − 2 °C) and extreme heat (above 32 °C)	0.5% and 2.2% increase in PROM risks, respectively	2015–2017	3,255 pregnancies with PROM in China	37
Temperature	Extremely high temperatures (32 °C)	6-day lagging relative risk of 1.08 for term PROM	2015–2019	77,941 pregnant women; 11,873 pregnant women were diagnosed with PROM	39
Strong winds, humidity	–	–	1999	–	40
Air pressure	–	Mixed results: positively, negatively, or no independent relationship	–	–	41,42
RH	10 units increase during the entire pregnancy	3.7% increase in PROM risks	2014–2019	20,392 dyads of mothers and neonates	This study

Machine learning for predicting PROM incidence

First, we use RF and XGBoost to generate the Feature Importance (FI) of medical covariates which can be known before delivery. According to the FI generated by RF, the medical variable that was the most influential factor in shaping the prediction result was neutrophil granulocyte count (8.96%), followed by leukocyte count (8.91%), GA (6.86%), CRP (6.56%), gravidity (6.27%), maternal age (5.80%), parity (5.40%), maternal height (5.39%), BMI (5.34%), weight (5.31%), and erythrocyte count (5.17%); these 11 factors combined accounted for 69.97%. According to the FI generated by XGBoost, the highest score was alcohol consumption (17.40%), followed by chronic maternal diabetes (16.04%), gravidity (7.31%), previous abortive times (7.16%), body temperature on admission (5.64%), leukocyte count (4.94%), GA (4.38%), neutrophil granulocyte count (3.39%), antibiotics use (3.34%), and DXM use (3.18%); these 10 factors combined accounted for 72.77%. The disparity between the FI generated by RF and XGBoost reflects the differences in weight estimates among different machine learning algorithms.

Second, we used all the clinical covariates as the inputs of machine learning model to predict the odds of PROM. We divided all datasets into equal parts. We used 90% for training and 10% for testing, and repeated this procedure for 10 times to cover all the PROM datasets.The outputs of RF and XGBoost were dichotomous; the outputs of MLP were numeric. The threshold of MLP prediction was set as 0.5. If the predicted value by MLP was greater than 0.5, PROM tended to occur. The accuracy of RF, XGBoost, and MLP for predicting PROM was 37.9% (1067/2814), 45.1% (1269/2814), and 40.2 (1132/2814), respectively. The accuracy of RF, XGBoost, and MLP for predicting non-PROM was 99.0%, 97.7%, and 98.2%, respectively. For all the cases RF, XGBoost, and MLP predicted as PROM, 86.0%, 75.9%, and 78.6% suffered from PROM, respectively. For all the cases RF, XGBoost, and MLP predicted as non-PROM, 90.9%, 91.7%, and 91.1% did not suffer from PROM, respectively.

Third, we use all the clinical covariates that can be obtained before delivery as the inputs of machine learning model to predict the odds of PROM. The accuracy of RF, XGBoost, and MLP for predicting PROM was 15.8%, 26.3%, and 17.2%, respectively. The accuracy of RF, XGBoost, and MLP for predicting non-PROM was 98.3%, 96.6%, and 98.7%, respectively. For all the cases RF, XGBoost, and MLP predicted as PROM, 59.6%, 55.6%, and 68.1% suffered from PROM, respectively. For all the cases RF, XGBoost, and MLP predicted as non-PROM, 87.9%, 89.1%, and 88.2% did not suffer from PROM, respectively. Due to the much higher proportion of non-PROM compared to PROM, it was more reasonable to set the threshold of whether PROM occurs to be below 0.5. For further validation, we set the threshold of MLP at lower than 0.5, as shown in Table S7. When threshold was set at 0.4 and 0.3, approximately 26.6% and 40.5% of all PROM cases can be forecasted, respectively. When the threshold was set at 0.2, we can accurately forecast 60.2% (1695/2814) and 81.6% (14349/17578) of PROM and non-PROM occurrence, respectively; for all cases MLP predicted as PROM and non-PROM, 34.4% and 92.8% suffered and did suffer from PROM, respectively. When the threshold was set at 0.1, we can accurately forecast 85.5% (2406/2814) and 59.9% (10526/17578) of PROM and non-PROM occurrence, respectively; for all cases MLP predicted as PROM and non-PROM, 25.4% and 96.3% suffered and did suffer from PROM, respectively.

Fourth, we use all the clinical covariates that can be obtained before delivery combined with air pollution exposure and meteorological conditions during the first trimester (O₃, PM₁₀, and RH) and second trimester (NO₂, PM_2.5, and PM₁₀) that were statistically significantly correlated with PROM as inputs of machine learning model to predict the odds of PROM. The results turned out to be better. The accuracy of RF, XGBoost, and MLP for predicting PROM was 14.9%, 26.5%, and 18.2%, respectively. The accuracy of RF, XGBoost, and MLP for predicting non-PROM was 98.9%, 96.6%, and 97.5%, respectively. For all the cases RF, XGBoost, and MLP predicted as PROM, 65.6%, 55.6%, and 54.0% suffered from PROM, respectively. For all the cases RF, XGBoost, and MLP predicted as non-PROM, 87.7%, 89.1%, and 87.5% did not suffer from PROM, respectively. Despite the fact that, if we set the threshold of MLP to be 0.1, approximately 86.1% (2422/2814) of PROM cases can be forecasted with an overall accuracy of 23.2%, this high level of prediction accuracy may be not quite helpful in practice, because we set the threshold at 0.1, which means there would be many false positive cases that reduce the usefulness of predictions in clinical settings.

Finally, the model performance was evaluated. In the independent test set, the PROM prediction model demonstrated good discriminative ability, with an area under the receiver operating characteristic curve (AUROC) of 0.78 (Fig. 12). Given the imbalanced distribution of PROM, precision–recall analysis was additionally performed, yielding an area under the precision–recall curve (AUPRC) of 0.39, substantially exceeding the baseline event prevalence and indicating meaningful predictive utility (Fig. 13). Calibration analysis showed a Brier score of 0.10, and logistic recalibration yielded calibration intercept and slope estimates that did not suggest substantial systematic over- or underestimation of risk. The decile-based calibration plot demonstrated reasonable agreement between predicted probabilities and observed PROM rates across risk strata (Fig. 14). At this threshold, the model achieved a sensitivity of 0.18, specificity of 0.98, positive predictive value of 0.54, and negative predictive value of 0.88. These findings indicate that the model prioritizes high specificity and moderate precision at the default threshold, while sensitivity remains limited. Importantly, the previously reported statement that the model “forecasted 86.1% of PROM cases” reflected recall obtained at a low probability threshold without accounting for precision; when evaluated using balanced threshold conditions and comprehensive reporting of discrimination, calibration, and predictive values, the model demonstrates a more moderate and clinically interpretable performance profile.

Fig. 12.

Discrimination and calibration performance of the PROM prediction model in the independent test set.

Fig. 13.

Precision–recall curve for PROM prediction.

Fig. 14.

Calibration plot for the PROM prediction model.

Discussion

Why PROM prediction is important

Over the past 40 years, China’s overall healthcare and medical levels have achieved significant progress: the average national life expectancy increased from 67.77 years in 1981 to 77.93 years in 2020; nationwide infant mortality rate decreased from 34.7‰ in 1981 to 5.0‰ in 2020⁵⁰. However, uneven distribution of medical resources (e.g., clinical materials, human resources, and high-level medical services) has posed a serious challenge to healthcare improvement in China, leading to regional disparities in medical utilization efficiency and negative health outcomes. To address this challenge, China has carried out a medical reform since 2015 to introduce a hierarchical medical system (HMS) for effectively allocating medical resources according to the actual needs of patients. This HMS reform has improved accessibility of public medical resources in large cities, but did not reduce the inequality of healthcare resource distribution between urban and rural areas. For instance, physician-to-bed ratio was lower in less developed regions in China. Since the challenges imposed on the health system have shown a character of regional heterogeneity in China, a practical, cost-effective, and region-specific medical approach is imperative to decrease the non-communicable disease burden. Here, we argue that if diseases can be forecasted prior to the onset, policy-makers and clinicians can early schedule the medical resource and craft treatment plans for potential patients in advance. In doing so, potential patients can be prioritized for special healthcare, the odds of diseases can be lowered, and the utilization of medical resources can be optimized. In this study, we take PROM for an example to illustrate how to improve the utilization efficiency of medical resources via disease prediction.

The first need for PROM prediction is to gain time. Once vaginal liquid discharge is observed, regardless of the amount, medical attention must be sought immediately; as PROM occurrence is confirmed, the pregnant women with should be hospitalized to receive medical treatments as fast as possible to reduce the risk of infection. As Table 1 exhibits, the average time interval between hospital admission and delivery of PROM women was 51% (19.6 h) shorter than that of non-PROM ones, indicating women suffering from PROM need to give birth as soon as possible to ensure the health of both mother and fetus; the average time interval between delivery and hospital discharge of PROM women was 30% (21.0 h) longer than that of non-PROM ones, implying women suffering from PROM require much more medical care after childbirth.

Given the lopsided distribution of medical resources in China, forecast of PROM is more urgent in the rural areas than urban. Therefore, if we can help the doctors identify women with higher odds of PROM before it happens, preemptive medical actions can be implemented to forestall PROM; if PROM still occurs with the aid of preemptive clinical treatment, early prediction can help doctors to better mobilize medical resources for postpartum care.

Main findings

Our main findings indicated that PROM was associated with 25 clinical variables—7 positively and 18 negatively—that can be detected before delivery. In terms of trimester exposure or entire pregnancy exposure, PROM was significantly positively associated with NO₂, O₃, PM_2.5, and PM₁₀. In terms of weeks exposure prior to delivery, PROM was significantly positively associated with SO₂, NO₂, PM_2.5, and PM₁₀. Thus, long-term maternal exposure to NO₂, O₃, PM_2.5, and PM₁₀ and short-term maternal exposure to SO₂, NO₂, PM_2.5, and PM₁₀ were identified as triggers for PROM. In terms of meteorological conditions, both long-term and short-term RH was significantly negatively linked to PROM, indicating the wetter the air, the less likely PROM would occur. Based on these findings, we propose an avenue to predict the odds of PROM weeks before delivery. Via PROM prediction, we can provide advance forecasts and early warnings for the occurrence of PROM.

Underlying mechanism

Air pollution may cause oxidative stress, inflammation reactions, cytotoxicity, and barrier structure disturbance, leading to PROM^1,51. Equally important, exposure to air pollution may result in harmful effects on collagen. Multiple types of collagens and fibrins are present in the fetal membrane, providing tensile strength and elasticity of tissues to prevent membrane from rupture that may be caused by stretching forces as fetus grows⁵². Inflammation reactions may degrade the extracellular matrix and disrupt membrane function⁵³. Placental mitochondrial change may be induced by intensified oxidative stress due to prenatal PM₁₀ exposure⁵⁴. Maternal exposure to NO₂, O₃, and PM_2.5 may trigger the formation of reactive oxygen species, destabilizing membrane function¹.

The highest rate of collagen synthesis occurs in the first trimester¹. In this study, we have found that there was significant association of PROM with O₃ and PM₁₀ exposure during the first trimester. Moreover, exposure to air pollution may result in an increased placental mutation rate⁵⁵. Air pollution has been linked to methylation loss in DNA and mitochondrial dysfunction⁵⁶. Reportedly, DNA methylation and mitochondrial DNA alteration may all contribute to adverse birth outcome in pregnant women and fetuses⁵⁷. However, there is still a lack of research on how air pollutants affect the probability of PROM occurrence by altering genes and epigenetics.

Paradox in early warning of PROM

As we have crafted a pathway to forecast PROM and tested its efficacy, the next step is to pass this information on to pregnant women susceptible to PROM as an early warning. As mentioned above, if the threshold is set to be 0.1, approximately 86.1% of PROM cases can be forecasted by deep learning using medical covariates that were statistically significantly linked to PROM and can be detected prior to delivery as input parameters of the model. However, this prediction also included non-PROM cases that were nearly three times more than PROM cases. Therefore, we face an irreconcilable paradox: the more accurate the PROM prediction, the lower the number of PROM occurrence that can be predicted; the more cases of PROM occurrence that can be predicted, the less accurate the prediction of PROM will be. Even though we can forecast PROM weeks before its occurrence, to decrease the odds of PROM, a lot of medical resources will be inevitably wasted, such as communication costs. However, we deem that these costs are worth spending, especially for those pregnant women who live in rural areas. Due to the uneven scattering of medical resources, when PROM occurs upon pregnant women who live in rural areas, they can find a Tertiary-A hospital easily; however, when pregnant women living in rural areas experience PROM, especially for those who live in remote regions or mountain areas where high-speed railway or other modes of public transportation do not cover, they often cannot obtain matching medical resources as soon as those living in urban areas.

Identifying susceptible PROM window

The parameters that were statistically significantly linked to PROM often showed non-linear relationship and sometimes even a U-shaped. For instance, as Fig. 2 shows, even though maternal age was significantly negatively linked to PROM in overall assessment, this phenomenon mainly occurred in pregnant women between the ages of 26 and 34. The probability of PROM increased with age in pregnant women under 26 years old and over 34 years old. However, as most pregnant women were between 26 and 34 years old, there was an overall significant negative correlation between PROM and age. As Fig. 1 shows, when the mother’s height was between 150 and 164 cm, the incidence of PROM increased with height; when the maternal height was greater than 164 cm, the incidence of PROM decreased as height increased. As Fig. 6 shows, the overall odds of PROM decreased as GA increased; however, this trend was reversed when GA was around 39 weeks. Moreover, since gravidity and parity often increased with maternal age, we need to establish causal relationship and distinguish causality from association, which was the hardest part in the field of public health.

As Figs. 7, 8 and 9 illustrate, the association of PROM with leukocyte count, neutrophil granulocyte, and CRP showed a U-shaped link. PROMs with high CRP levels and leukocytosis may increase the risk of other adverse birth outcomes⁵⁸. When leukocyte count was around 10.1 × 10⁹/L and neutrophil granulocyte count was around 11.0 × 10⁹/L, the odds of PROM were the lowest. There was no significant association between PROM and erythrocyte count. Based on these clinic findings, prophylactic antibiotics, probiotics, and corticosteroids can be used as preemptive actions to decrease the odds of PROM.

Personalized treatment

Using deep learning, we can craft a personalized treatment plan and evaluate its effectiveness⁴⁴. As the risk factors that were associated with higher odds of PROM have been identified, the probability of PROM occurrence can be reduced by influencing those risk factors through manual intervention. Using deep learning for prediction, it was found that their average probability of suffering from PROM was as high as 97.2%; if they have stopped alcohol consumption during pregnancy, their average probability of suffering from PROM was only 23.8%.

Strengths and limitations

The primary strength of our study is the quality and quantity of the data. All the clinical data were recorded in detail. The data of air pollution levels and meteorological conditions were provided by local government. We have fully known the home and working addresses of each participant, which allowed us to find the nearest monitoring station to obtain exposure data.

Several limitations should also be noted. First, maternal exposure to heavy metals, including chromium⁵⁹, vanadium⁶⁰, and lead⁶¹, was found to be associated with PROM. Maternal exposure to rare earth elements was also linked to PROM⁶². In this study, we did not record maternal exposure to heavy metals and rare earth elements. Second, due to the mobility of people, participants may go places far away from home or even move outside the study area. We only know their home and working addresses, but did not know if they travel to other cities, other provinces, or even overseas, which resulted in inaccurate estimates of maternal exposure levels. Third, ambient air pollution levels and meteorological conditions do not necessarily reflect the indoor settings. Air conditioning, heating, and air purifier, which have been popular in household in East China, can lead to the significant change in exposure levels. We perceive the only circumstance in which the assessment on maternal exposure can be flawless was during the COVID-19 lockdown. However, even in the COCID-19 quarantine and lockdown, during which people cannot leave their living places, there is an obvious discrepancy between indoor and outdoor air quality, temperature, and RH, leading to a certain degree of inaccuracy. Fourth, we find the statistical link between RH and PROM, but have not established the causality, whose importance is greater than association. Fifth, the association of alcohol consumption with PROM was striking, but we must point out that the habit of routinely drinking alcohol during pregnancy was self-reported by mothers, and we were not fully aware of exact amount of alcohol consumptions. Also, we defined “routinely” as more than once a week during entire pregnancy, we did not get the information of exact times a mother consumed alcohols every week, which may lead to undeniable bias. Sixth, since the toxicity of different PM components may vary significantly across China⁶³, we must monitor the proportion of different components in PM in the future work to more accurately assess the maternal exposure. Seventh, this study only explores the numerical correlation of PROM with meteorological conditions and air pollutants, which cannot indicate causality. For instance, even though we find the significant association between PROM and RH, there might be no causal relationship between them; current analysis can only show linkage, but not causality. Eighth, Area Under the Receiver Operating Characteristic Curve (AUROC), Area Under the Precision-Recall Curve (AUPRC), sensitivity, specificity, and confusion matrices are indispensable parameters for evaluating binary classification models. However, we consider that p-Value and FI were enough to let us find the important inputs for the machine learning models. We plan to discuss these important parameters in another full-length article. If we overly focus on these parameters, readers may be misled to overlook the thrust of this article, which is to predict the incidence of PROM months before its actual occurrence.

Moreover, distinguishing between pre-labor PROM (pre-delivery but at term) and preterm pre-labor PROM is of importance, because their perinatal consequences are quite different for mothers and neonates. In the future work, we should focus more on preterm PROM. It is also worth noting that in preterm PROM, approximately half of mothers may labor within 7 days and their morbidity was highly likely due to PTB; a certain proportion of mothers may develop chorioamnionitis which increased maternal and neonatal adverse events. We need to investigate the impacts of chorioamnionitis on PROM.

Previous PTB (medical history) or cervical surgery are risk factors for a repeat PTB, but not necessarily PROM. A short cervix (in someone with a previous history of PTB) and raised fetal fibronectin in a symptomatic patient are significant risk factors or a useful clinical tool to predict PROM. Other risk factors such as fatty acid concentrations in deficient mothers and dental care may impact the odds of PROM. What’s more, intra-amniotic inflammation with preterm PROM accounted for a substantial proportion of PTB cases⁶⁴. Kynurenine, tryptophan, and neopterin may act as indicators to show the disorders of pregnancy⁶⁵. To more accurately predict the odds of PROM, we need to include more risk factors in the machine learning models.

The use of antibiotics and corticosteroids may reduce neonatal morbidity after PROM but whether they can be used as a preventative treatment for reducing the odds of PROM is not clear yet and needs more demonstrations. Interleukin-6 level in amniotic fluid is a better indicator than white blood cell counts to show microbial invasion of amniotic cavity⁶⁶. In the future work, we need to measure maternal interleukin-6 concentration and compare its impact on PROM with that of white blood cell count. The ORACLE trials may be a guideline⁶⁷ for our future work. Through PROM prediction, testing and evaluation of health care applications of Large Language Models⁶⁸ can be used in this case.

Conclusion

In summary, clinical covariates that were significantly associated with PROM has been identified. Air pollution exposure and meteorological conditions may ratchet the early warning of PROM another notch. Based on these findings, to our knowledge, we first propose an approach that has practical application value to forecast the probability of PROM occurrence for crafting personalized treatment plans and optimizing the utilization efficiency of medical resources.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

R. F. and C.B. ran the SPSS and wrote the manuscript. J. X ran the machine learning models. X. W ran the R and prepared all the figures. C. Y prepared all the tables. Z. L. provided the meteorological data.

Funding

This study was supported by the “Pioneer” and “Leading Goose” R&D Program of Zhejiang province (No. 2025C02123), Zhejiang Provincial Education Project (No. Y202351306), Fundamental Research Funds for the Central Universities (No. 226-2024-00040), and Zhejiang Provincial Natural Science Foundation (No. LQN25D050006).

Data availability

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Code availability

All the code used in study to run machine learning models is accessible online (https://github.com/superxjm/PROM_prdict/tree/main). The input datasets are available from the corresponding author and can be given on reasonable request via e-mail.

Declarations

Ethics approval and consent to participate

This study was approved by the ethics committees of the Sir Run Run Shaw Hospital, School of Medicine, Zhejiang University (Trial No. 2024 − 1181). Written Informed consent was obtained from all participants and their legal guardians.

Competing interests

The authors declare no competing interests.