Risk Factors Associated With Postpartum Hemorrhage In Full-Term Singleton Delivery In A General Hospital In Liaoning Province: Construction Of A Risk Assessment And Prediction Model

Background:Postpartum hemorrhage(PPH)is one of the leading causes of maternal mortality,but the majority of these deaths are preventable.In routine clinical practice,doctors roughly predict the probability of PPH by assessing clinical history,perfecting physical examination and laboratory tests.However,due to the limited sensitivity and specificity of various tests and the low incidence of PPH,the previous methods of using bleeding assessment tools to assess the risk of bleeding,such as structured history collection and systematic assessment scale,have low efficacy in assessing the incidence of PPH.In recent years,with the opening of the two-child policy,the incidence of PPH is rising,and the studies on constructing PPH risk prediction model become hot spots again.However,firstly,most of the current studies do not distinguish PPH between cesarean section and vaginal delivery;Secondly,the selection of model predictors was limited by the lack of medical history collection and sample size.Therefore,most clinical prediction models lack relatively comprehensive evaluation.This study aims to explore the common risk factors in vaginal delivery patients with PPH,and the PPH subgroups classificaed by etiology(PPH of uterine atomy,PPH of placental factors,PPH of canal Trauma,and PPH of coagulation abnormal)the potential risk factors and incidence,and to construct the risk prediction model of PPH and each subgroup,verify and evaluate the model,so as to provide a multi-angle theoretical basis for PPH prevention and control.Objective:1.In this study,we aim to analyze the association between various factors,such as maternal characteristics,pregnancy history,complications during pregnancy,the process of labor,and neonatal conditions,with the risk of postpartum hemorrhage(PPH)in a population of women undergoing vaginal delivery in a large public tertiary maternity hospital.We will explore potential risk factors and their predictive efficacy for PPH caused by uterine atony,placental factors,genital tract lacerations,and coagulation abnormalities,based on the main etiologies of PPH.2.Logistic regression analysis will be used to determine the impact weights of each factor on the probability of PPH occurrence.Subsequently,separate models will be developed and graphically represented for each of the four etiological subgroups to enhance the accuracy of predicting high-risk individuals for PPH and provide valuable insights for PPH prevention and management.3.In addition,we plan to construct and evaluate predictive models for PPH and its etiological subgroups using machine learning algorithms,aiming to provide reliable reference for personalized assessment and prevention of PPH in high-risk pregnant women.Methods:1.In this cohort study,the obstetric ward in Shengjing Hospital of China Medical University was used as the investigation site to investigate the women who completed vaginal delivery in the hospital from January 1,2016 to December 31,2020,and the fetal birth outcomes of the included women were followed up.The survey contents included the basic characteristics of the puerpera,the history of pregnancy and pregnancy,the complications during pregnancy,the process of delivery and the situation of the newborn.After defining the relevant definitions,the subjects were divided into two subgroups according to the occurrence of PPH\PPH of uterine atomy\PPH of placental factors\PPH of canal Trauma and PPH of coagulation abnormal.The basic characteristics,pregnancy history,complications during pregnancy,delivery process and neonatal conditions of the two groups were analyzed,and the high-risk factors closely related to the occurrence of postpartum hemorrhage with different causes were identified.2.On the basis of the first part of the survey,the subjects were divided into two subgroups,and the differences between the two groups were compared.Univariate and multivariate analyses were used to screen the independent risk factors for PPH.The study population was divided into training set and test set according to the ratio of 7:3.The conventional Logistic regression analysis was used to establish the risk prediction model of PPH and each subgroup,and the performance of the model was evaluated and verified from three aspects of discrimination,calibration and clinical utility.The PPH clinical prediction models in other studies were compared to comprehensively evaluate the model.The above statistical analysis was accomplished using R version 3.6.3 software.3.On the basis of the second part,SMOTE oversampling method was used to divide the research data into training set and test set with a ratio of 7:3 according to whether PPH occurredsetset.The training set was used to construct the machine learning prediction model,and the test set was used to evaluate the performance of the machine learning prediction model.Four different machine learning algorithms,including random forest,naive Bayes,Adaboost and Gradient Boosting Decision Tree(GBDT),were used to construct the prediction model using the training set.After comparing the discrimination performance of different algorithms,the Gradient Boosting Decision Tree was used to build a machine learning model,and the model was verified and evaluated from two aspects of discrimination and calibration.Finally,optimized Machine-Learning Prediction Models were built into a desktop application software through Python platform.The above statistical analysis was accomplished using Python3.8.5 software.Results:1.According to the inclusion and exclusion criteria,a total of 24833 patients were included in this study,with an average age of 29.66±3.53 years old and an average blood loss of 393.54±92.53 ml.The overall incidence of PPH,UA-PPH and PF-PPH was 6.54%,4.93%and 0.97%,respectively,CT-PPH was 0.56%,and CA-PPH was 0.31%.Age,race,occupation,pre-pregnancy BMI(OR for Overweight=2.16,OR for Obesity=2.43,reference to BMI<18.5Kg/m~2),delivery times(OR for first delivery=2.50,refercence to second delivery),labor gestational age(OR for great than 40 weeks=2.92,reference to less than 38 weeks),diabetes(OR=1.62),hypertension(OR=2.68),anemia(OR=1.95),coagulant function abnormality(OR=3.92),combination of uterine myoma/uterine adenomyosis(OR=3.08),hydramnios(OR=1.68),premature rupture of membranes(OR=2.15),placental abruption(OR=8.19),scar uterus incubation period(OR=3.05),the total labor time(OR=2.04),the first labor time(OR=3.25),the first"active"labor time(OR=3.17),the second labor time(OR=4.76),the third labor time(OR=4.91),combine the placenta residue/placental adhesion/placenta accreta(OR=2.22),analgesic delivery(OR=3.02),instrument-assisted delivery(OR=2.30),cervical/vaginal/vulva laceration(OR=1.48)and neonatal weight(OR for greater than 4000g=9.40,reference to less than 2500g)are closely related to the occurrence of PPH.Among them,common rist factors such as pre-pregnancy BMI,delivery times,labor gestational age,diabetes,hypertension,anemia,combination of uterine myoma/uterine adenomyosis,premature rupture of membranes,labor time,combined with residual placenta/placenta adhesion/placenta accreta,analgesic delivery and neonatal weight were all associated with adverse outcomes of PPH in each subgroup.In the UA-PPH subgroup,age,race,occupation,hydramnios,placental abruption,scar uterus incubation period were six additional independent risk factors for UA-PPH except for common factors.In the PF-PPH subgroup,three additional risk factors such as race,occupation,hydramnios were associated with PF-PPH except for the common factors.In the CT-PPH subgroup,instrument-assisted delivery,cervical/vaginal/vulva laceration and soft birth canal laceration were closely related to the occurrence of PPH except for common factors.In the CA-PPH subgroup,five additional factors such as occupation,smoking,hydramnios,placental abruption,scar uterus incubation period were independent risk factor for CA-PPH except for common factors.2.The clinical prediction model was established after Logistic regression analysis of the overall PPH,and the high discrimination degree was obtained,namely,the AUC of the training set was 0.807(95%CI:0.792-0.821),and the AUC of the test set was 0.803(95%CI:0.781-0.825).The calibration curves were all close to 45 degrees,and the DCA curve showed good clinical utility.In the Nomogram,placental abruption and neonatal weight had a greater weight in the model.In the UA-PPH subgroup,the AUC of the training set was 0.794(95%CI:0.777-0.811),and the AUC of the test set was 0.782(95%CI:0.782).0.756-0.807),and the calibration curves are close to the 45 degree line.The DCA curve reflects the clinical utility well,and the scoring weight of each risk factor in the Nomogram is basically the same as that of PPH.In the PF-PPH subgroup,the AUC in the training set was 0.796(95%CI:0.761-0.830),and the AUC in the test set was only 0.739(95%CI:0.666-0.813).The calibration curves deviated from the 45 degree line,and the DCA curve reflected poor clinical utility.In the CT-PPH and CA-PPH subgroups,the AUC of the training set was 0.935(95%CI:0.901-0.969)and 0.807(95%CI:0.901-0.969),respectively.0.792-0.821),and the AUC of test set was 0.943(95%CI:0.903-0.982)and 0.807(95%CI:0.781-0.825),but all the calibration curves deviated from the 45 degree line.The clinical utility of DCA curve was poor,and the modeling effect was not ideal.3.Using SMOTE oversampling,the study data was expanded to 46420 cases.In PPH-MLPM,the AUC of training set and test data were 0.995(95%CI:0.995-0.996)and0.989(95%CI:0.9888-0.991),respectively.In the test set,the area difference under the Receiver operating characteristic of Random forest,Adaboost and GBDT methods and logistic regression was 0.079,1.116 and 0.186,respectively,which was statistically significant.The area under the Receiver operating characteristic of Random forest,Adaboost and GBDT methods was significantly higher than the area under the Receiver operating characteristic of the traditional logistic regression model.The AUC of UA-PPH-MLPM was 0.996(95%CI:0.996-0.997)in the training set and 0.991(95%CI:0.990-0.993)in the test set.The AUC of PF-PPH-MLPM in training set and test set were0.9996(95%CI:0.9995-0.9997)and 0.9978(95%CI:0.9971-0.9985).The AUC of CT-PPH-MLPM in the training set was 0.9999(95%CI:0.9998-0.9999),and the AUC in the test set was 0.9996(95%CI:0.9994-0.9998).The AUC of CA-PPH-MLPM in the training set was 0.9997(95%CI:0.9995-0.9998),and the AUC in the test set was 0.9988(95%CI:0.9984-0.9993).The discrimination of the above models is high,and the calibration curves are close to the 45 degree line,the AUC of each models is statistically greater than that of the conventional Logistic model,and the prediction accuracy is high.Conclusion:1.The occurrence of PPH and subtype PPH in cesarean deliveries is associated with pre-pregnancy BMI,delivery times,labor gestational age,diabetes,hypertension,parity,anemia,combination of uterine myoma/uterine adenomyosis,premature rupture of membranes,labor time,combined with residual placenta/placenta adhesion/placenta accreta,analgesic delivery and neonatal weight common factors,while there also exists different special risk factors for each subtype.2.Apart from commonly shared factors such as parity,pre-pregnancy BMI,gestational age at delivery,diabetes,hypertension,anemia,comorbid uterine fibroids/endometriosis,premature rupture of membranes,prolonged labor,and placental remnants/adhesions/implantation,we also identified specific risk factors associated with specific etiologies of postpartum hemorrhage.For example,polyhydramnios is associated with PPH caused by uterine atony,while epidural analgesia,instrumental delivery,and cervical/vaginal/perineal lacerations are related to PPH caused by genital tract trauma.3.The PPH-Logistic prediction model uses factors such as ethnicity,occupation,pre-pregnancy BMI,parity,gestational age,diabetes,hypertension,anemia,abnormal coagulation function,uterine fibroids/adenomyosis,polyhydramnios,premature rupture of membranes,placental abruption,scarred uterus,total duration of labor,latent phase duration,active phase duration of the first stage of labor,duration of the second stage of labor,duration of the third stage of labor,retained placenta/adherent placenta/placental implantation,epidural analgesia,instrumental delivery,cervical/vaginal/perineal lacerations,and newborn weight as predictors.The model exhibits high discrimination,high predictive accuracy,and good clinical utility.However,in the etiological subgroup,discrimination,calibration,and clinical utility may vary.4.In the etiological subgroup,the UA-PPH-Logistic prediction model demonstrates superior discrimination,calibration,and clinical utility,with high predictive accuracy and clinical applicability.However,the PF-PPH-Logistic,CT-PPH-Logistic,and CA-PPH-Logistic prediction models show less desirable performance in certain aspects,with lower calibration and clinical utility.5.Through external validation using real-world data,the PPH-Logistic and UA-PPH-Logistic prediction models demonstrate good discrimination,calibration,and clinical utility.On the other hand,the PF-PPH-Logistic,CT-PPH-Logistic,and CA-PPH-Logistic prediction models show good discrimination but exhibit less desirable results in terms of calibration and/or clinical utility validation.6.By constructing machine learning prediction models for postpartum hemorrhage and its etiological subgroups and evaluating the performance of the models,it is shown that these machine learning prediction models have high predictive accuracy and discrimination for various types of postpartum hemorrhage.7.In the testing set,the machine learning prediction models demonstrate good performance,with the machine learning prediction models for postpartum hemorrhage and its etiological subgroups exhibiting better discrimination than the logistic prediction models.However,the evaluation results on a real-world external validation set show that in terms of calibration,the machine learning prediction model for postpartum hemorrhage caused by soft birth canal factors overestimates the clinical predictive risk.Additionally,the machine learning prediction model for postpartum hemorrhage caused by abnormal coagulation function cannot currently be extrapolated for clinical prediction.8.To leverage the advantages of machine learning-based clinical prediction models and enhance their role in clinical decision-making and healthcare resource allocation,the project team has developed and validated a clinical application software for machine learning prediction models.This software can be used by healthcare professionals to better understand and address the complexities of the medical environment,enabling the implementation of more accurate and efficient management strategies and providing a valuable tool for risk assessment of postpartum hemorrhage in clinical practice.9.Future research should focus on obtaining support from large-scale,multi-center datasets and continuously improving these prediction models to enhance their accuracy and reliability.This will provide more effective support for the prevention and treatment of postpartum hemorrhage in full-term cephalic deliveries.