{"paper_id":"96cc4a54-f215-48d9-ae53-91ab19744299","body_text":"1 \n \nTITLE \nDevelopment of risk prediction models for preterm delivery in a rural setting in Ethiopia \n \nAUTHORS \nClara Pons-Duran, PhD\n1*, Bryan Wilder, PhD1,2*, Bezawit Mesfin Hunegnaw, MD MPH3, \nSebastien Haneuse, PhD4, Frederick G. B. Goddard, MS PhD1, Delayehu Bekele, MD MPH1,5, \nGrace J. Chan, MD MPH PhD1,3,6 \n \nAFFILIATIONS \n1 Department of Epidemiology, Harvard T.H. Chan School of Public Health, Boston, MA, USA \n2 Machine Learning Department, Carnegie Mellon University, Pittsburgh, PA, USA \n3 Department of Pediatrics and Child Health, St. Paul’s Hospital Millennium Medical College, \nAddis Ababa, Ethiopia \n4 Department of Biostatistics, Harvard T.H. Chan School of Public Health, Boston, MA, USA \n5 Department of Obstetrics and Gynecology, St. Paul’s Hospital Millennium Medical College, \nAddis Ababa, Ethiopia \n6 Division of Medical Critical Care, Boston Children’s Hospital, Department of Pediatrics, \nHarvard Medical School, Boston, MA, USA. \n* Joint first authors \n \nCORRESPONDING AUTHOR \nClara Pons-Duran, Department of Epidemiology, Harvard T.H. Chan School of Public Health, \n677 Huntington Ave, Kresge 913, Boston, MA 02115 \ncponsduran@hsph.harvard.edu \n \nWORD COUNT \n3072\n \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted December 24, 2022. ; https://doi.org/10.1101/2022.11.04.22281948doi: medRxiv preprint \nNOTE: This preprint reports new research that has not been certified by peer review and should not be used to guide clinical practice.\n\n2 \n \nABSTRACT \nBackground. Preterm birth complications are the leading causes of death among children under \nfive years. A key practical challenge, however, is the inability to accurately identify pregnancies \nthat are at high risk of preterm delivery, especially in resource-limited settings where there is \nlimited availability of biomarkers assessment. \nMethods. We evaluated whether risk of preterm delivery can be predicted using available data \nfrom a pregnancy and birth cohort in Amhara region, Ethiopia. All participants were enrolled in \nthe cohort between December 2018 and March 2020. The study outcome was preterm delivery, \ndefined as any delivery occurring before week 37 of gestation regardless of vital status of the \nfetus or neonate. A range of sociodemographic, clinical, environmental, and pregnancy-related \nfactors were considered as potential inputs. Cox and accelerated failure time models, and \ndecision tree ensembles were used to predict risk of preterm delivery. Model discrimination was \nestimated using the area-under-the-curve (AUC). Additionally, the conditional distributions of \ncervical length (CL) and fetal fibronectin (FFN) were simulated to ascertain whether those \nfactors could improve model performance. \nResults. A total of 2493 pregnancies were included. Of those, 138 women were censored due to \nloss-to-follow-up before delivery. Overall, predictive performance of models was poor. The \nAUC was highest for the tree ensemble classifier (0.60, 95%CI [0.57, 0.63]). When models were \ncalibrated so that 90% of women who experienced a preterm delivery were classified as high \nrisk, at least 75% of those classified as high risk did not experience the outcome. The simulation \nof CL and FFN distributions did not significantly improve models’ performance. \nConclusions. Prediction of preterm delivery remains a major challenge. In resource-limited \nsettings, predicting high-risk deliveries would not only save lives, but also inform resource \nallocation. It may not be possible to accurately predict risk of preterm delivery without investing \nin novel technologies to identify genetic factors, immunological biomarkers or the expression of \nspecific proteins.\n \n \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted December 24, 2022. ; https://doi.org/10.1101/2022.11.04.22281948doi: medRxiv preprint \n\n3 \n \nINTRODUCTION \nGlobally, almost 15 million babies are born preterm before 37 weeks of gestation each year [1]. \nA variety of factors are known to be associated with risk of preterm birth, including obstetrics \nhistory, anthropometric measurements, infections, ultrasound measurements and biological and \ngenetic markers [2,3]. Accurate prediction tools to identify women at an increased risk of \npreterm delivery would allow policy makers, practitioners and researchers to target interventions \ndesigned to reduce preterm deliveries. Some studies conducted in high income settings \ndeveloped predictive models to classify women based on their risk for preterm birth considering \nmultiple maternal characteristics [4,5]. Their discriminative performance was modest (area under \nthe receiving operator characteristic curve (AUC) from 0.62 to 0.70), with generally lower \nperformance when performing external validation [5]. \nTargeting interventions to high-risk pregnancies is a critical challenge because of the lack of \naccurate prediction tools. Some published models were developed for women with a priori \nknown risk factors like preterm labor or multiple pregnancy [6-8]. Other models used predictors \nthat are not readily available in resource-limited settings such as cervical length (CL), bacterial \nvaginosis, fetal fibronectin (FFN), cytokine concentration and other biomarkers [9-11]. \nAdditionally, to our knowledge, no published model handled competing risks of stillbirth, or \nconsidered a combined outcome of preterm delivery regardless of vital status of the fetus or \nneonate. However, preterm and stillbirth share common causes and risk factors, and it is likely \nthat the biological mechanisms that trigger preterm labor or rupture of membranes may lead to \nthe delivery of a preterm stillborn in extreme cases [12,13]. Overall, there is a gap in the \ndevelopment of prediction tools that are accurate and applicable to the general population, with \nand without a priori risk, especially in low-resource countries where data on biomarkers that \ncould contribute to improve model performance are not commonly available. \nMost preterm birth cases occur among women without known risk factors [14,15]. Limited \navailability of promising biomarkers to predict preterm birth in low-resource settings make it \ncritical to develop context-specific predictive tools. We use large datasets from the Birhan \npregnancy cohort [16], to test whether it is possible to predict risk of preterm delivery in rural \nEthiopia. Additionally, we aim to ascertain whether it would be effective to invest in the \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted December 24, 2022. ; https://doi.org/10.1101/2022.11.04.22281948doi: medRxiv preprint \n\n4 \n \ncollection of known key predictors such as CL or FFN to improve accuracy of predictions by \nsimulating the conditional distribution of those two factors. \n \nMETHODS \nStudy design and setting \nWe conducted a cohort study in the Birhan field site, including 16 villages in Amhara region, \nEthiopia, covering a population of 77,766, to estimate morbidity and mortality outcomes among \n17,108 women of reproductive age and 8,554 children under-five with house-to-house \nsurveillance every three months. The site is a platform for community and facility-based research \nand training that was established in 2018, with a focus on maternal and child health [17]. Nested \nin the site is an open pregnancy and birth cohort that enrolls approximately 2,000 pregnant \nwomen and their newborns per year with rigorous longitudinal follow-up over the first two years \nof life and household data linked with health facility information [16]. The catchment area is \nrural and semi-urban, covers both highland and lowland areas, and includes two different \ndistricts, Angolela Tera, and Kewet/Shewa Robit. \nWe used data from the Birhan Health and Demographic Surveillance System (HDSS) and the \nnested pregnancy and birth cohort, Birhan Maternal and Child Health (MCH) cohort to develop a \nseries of risk prediction models for preterm delivery [16,17]. The HDSS provides estimates and \ntrends of health and demographic outcomes including morbidity among women of reproductive \nage and children under two years and births, deaths, marriages, and migration in the entire \npopulation [17]. The pregnancy and birth cohort, generates evidence on pregnancy, birth, and \nchild outcomes using clinical and epidemiological data at both the community and health facility \nlevel [16]. \nStudy participants \nThe sample for this study included women enrolled during pregnancy in the MCH cohort \nbetween December 2018 and March 2020. They were followed-up in home and facility visits \nthrough delivery. Women who were enrolled during pregnancy and were followed up beyond 28 \nweeks of gestation were included in the study. This gestational age cut-off was used because in \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted December 24, 2022. ; https://doi.org/10.1101/2022.11.04.22281948doi: medRxiv preprint \n\n5 \n \nEthiopia stillbirths are considered ≥ 28 weeks. We excluded newborns with implausible \ngestational ages at birth: <28 weeks due to the definition of stillbirth, and ≥ 46 weeks. \nStudy variables and definitions \nThe outcome of this study was preterm delivery, a composite indicator defined as any delivery \noccurring before 37 completed weeks of gestation, regardless of vital status of the fetus or \nneonate. This includes both preterm births (live birth prior to completion of week 37 of \ngestation) [18], as well as stillbirths (any fetal death after 28 completed weeks of gestation) [19] \nwhich occurred before 37 weeks of gestation. \nGestational age was estimated using the best available method from ultrasound measurements, \nreported date of last menstrual period, fundal height or maternal recall of gestational age in \nmonths. Detailed information on these estimations can be found elsewhere [20]. \nThe selection of potential predictors was guided by literature review and expert knowledge from \nstudy obstetricians and pediatricians. Predictors with low prevalence rates in the sample (rare \nevents with \n≤ 5 cases) were dropped. Over 70 socio-demographic, biological, environmental and \npregnancy-related predictors were included in the initial models. The complete list of assessed \npredictors can be found in Table S1 (Online Supplementary Document 1). Dummy variables \nindicating missingness for each predictor were included as additional variables, an approach that \nis justified for predictive models because it reflects the complete state of knowledge available at \nthe time of prediction. \nAnalysis \nDescriptive statistics \nA descriptive analysis of the background characteristics of women who experienced term \ncompared to preterm delivery was performed using t-test for continuous variables, chi-square test \nfor most of the binary variables and Fisher’s exact test for multiple gestations, to test for \nstatistically significant differences between groups. \nPrediction models \nFive models were fit to predict risk of preterm delivery, including linear models and nonlinear \ndecision tree approaches. All five strategies were designed to predict the outcome of preterm \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted December 24, 2022. ; https://doi.org/10.1101/2022.11.04.22281948doi: medRxiv preprint \n\n6 \n \ndelivery using information available at 28 weeks of gestation. The first four models were time-\nto-event methods which modelled the time until delivery from the 28th week gestation mark, \naccounting for left truncation and right censoring of person-time. Left truncation arises when \nwomen are enrolled beyond 28 weeks of gestation while right censoring arises when follow-up \nceases prior to observation of the event of interest (e.g. due to outmigration or loss-to-follow-up). \nThe first model was a Cox proportional hazards model, fit using the R package survival [21]. \nSecond, an accelerated failure time model was fit with a log-logistic distribution using the R \npackage flexsurv [22]. Third, a decision tree was fit using the R package LTRtrees that extends \nprevious uses of a decision tree in survival analysis to account for left truncation and right \ncensoring (LTRCART, left truncation right censoring classification and regression trees) [23]. \nFourth, a decision tree ensemble was implemented using the eXtreme Gradient Boosting (XGB) \nR package which uses a Poisson likelihood function proposed by Fu and Simonoff (2017) to \naccount for right censoring and left truncation [23,24]. Finally, a fifth analysis based on a XGB \nclassification model was fit using a binary outcome (i.e. whether delivery was preterm or not) \ninstead of the time-to-event. During fitting this last model, data which was either right censored \nor left truncated was excluded. \nModels were fit and evaluated using 5-fold cross validation due to the need to evaluate models \non out-of-sample data while reserving as much data as possible for fitting [25]. All models were \nevaluated on the same held-out dataset within each fold, regardless of which data or methods \nwere used while fitting the model. Model performance was assessed using the AUC to assess \naccuracy at binary classification, and the c-index to assess the fraction of pairs for which \npredicted risk was concordant with delivery time. For both metrics, a value of 0.5 represents a \nrandom prediction which is uncorrelated with the true outcome. Larger values indicate more \naccurate predictions, and a value of 1 represents predictions which are perfectly concordant with \nthe true outcome. \nSimulation of cervical length and fetal fibronectin \nA final analysis was performed to simulate the potential impact of including CL and FFN as \npredictors. These two variables were found in past work to be significantly associated with \npreterm birth [26-28]. Since they are not regularly collected in the study region, we used \nsimulation to assess the potential gain from collecting them. The simulation used data from the \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted December 24, 2022. ; https://doi.org/10.1101/2022.11.04.22281948doi: medRxiv preprint \n\n7 \n \nMFMU PREDS study [29], a study which screened 2929 women for risk factors for preterm \nbirth in the United States. PREDS study identified CL and FFN as key predictors for preterm \nbirth [30,31]. Details on the simulation model and comparison between the simulated and real \nmeasurements can be found in the Supplemental Methods and Results (Online Supplementary \nDocument 1). \n \nRESULTS \nThe sample composed 2834 pregnancies. Among those we excluded 75 (2.6%) records with \ngestational age at delivery <28 and \n≥ 46 weeks, and a further 266 (9.4%) pregnancies whose \nfollow-up did not go beyond 28th gestational week. A total of 2493 pregnancies were included in \nthe study. Of those, 138 (5.5%) women were lost to follow-up before delivery or did not have a \nrecorded gestational age at delivery, so they were treated as censored observations in the time-to-\nevent models and excluded from the binary classification model. A total of 968 (38.8%) women \nwere enrolled in the cohort after 28 weeks of gestation (left-truncation), thus, time-varying \npredictors were considered missing for them since no information on those factors was available \nat the time of prediction. These women were also excluded from the binary classification model. \nAmong the 2355 women who were included in the study and followed until delivery, 14% had a \npreterm delivery. There was no difference in some background characteristics like age, body \nmass index, parity or history of previous preterm births among women with term deliveries \ncompared to women with preterm deliveries (Table 1). However, the two groups showed \nsignificant differences in literacy (43.3% of women with term delivery were illiterate, compared \nto 50.5% of those who delivered prematurely), geographic location (42.8% of term deliveries \noccurred in the highland district within Birhan field site, compared to 52.4% of preterm \ndeliveries), and multiple gestation (1.1% of term deliveries were multiple, compared to 3.4% of \npreterm deliveries). \nSome predictors had high levels of missing data, particularly where our study relied on facility \nvisits. Approximately 25% of participants did not attend any antenatal care visit in the study \nhealth facilities after being enrolled in the cohort, thus, variables collected at antenatal care \nvisits, such as current infections or concomitant diseases were missing for these women. Further, \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted December 24, 2022. ; https://doi.org/10.1101/2022.11.04.22281948doi: medRxiv preprint \n\n8 \n \nover 70% of women who attended at least one antenatal care visit had missing data on lab and \npoint-of-care results such as white blood cell counts, proteinuria or bacteriuria. \nThe predictive performance of all models was generally poor (Table 2). The c-statistic and AUC \nwere highest for the XGB classification model, with an AUC of 0.60. The receiver operating \ncharacteristic curves (ROC) depict the trade-off between the false and true positive rates \nachieved by varying the threshold for classifying delivery as preterm or term (Figure 1). As an \nexample, at the point on this curve corresponding to a 90% true positive rate, all models had a \nfalse positive rate of at least 75%, indicating a lack of specificity in picking out women who are \ntruly at higher risk. \nThere was substantial heterogeneity in the factors that were ultimately retained in the five models \n(Table 3). Both biological and socio-demographic factors were among the top contributors of \nstandard time-to-event models. Regarding decision tree models, the top five predictors are \nmainly biological, with neonatal sex being the predictor with the greatest importance. \nThe performance of several individual models improved when simulated measurements of CL \nand FFN were included as features for each individual, particularly the accelerated failure time \nand LTRCART decision tree models (Table 4). However, no model exceeded an estimated AUC \nof 0.60 indicating that the overall predictability of preterm delivery did not change substantially \nfrom the inclusion of these additional predictors. \n \nDISCUSSION \nOur study shows that risk prediction of preterm delivery remains a challenge in the absence of \ndata on biomarkers. Despite using a wide range of methodological approaches to adjust for \nmissing data, competing risk of stillbirths, and late cohort enrollments, both traditional \nepidemiological and machine learning models performed poorly and had low specificity in \nidentifying women who delivered before 37 weeks of gestation. This low predictive performance \ndiffers with existing models with higher predictive ability that were designed to predict preterm \nbirth among women at an already high risk due to obstetric conditions such as twin pregnancy \n[7,8], short cervix, cervical insufficiency [10,32], or hospital admission due to preterm labor \n[6,33,34]. Other higher performing algorithms used predictors that were not available or \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted December 24, 2022. ; https://doi.org/10.1101/2022.11.04.22281948doi: medRxiv preprint \n\n9 \n \napplicable in low-resource settings such as amniotic and cervical fluids [33], inflammatory \nmarkers [35], or method of conception [5,8]. Lack of a fixed prediction time point and competing \nrisk of stillbirth are methodological gaps of most published studies. \nTo our knowledge, our study is among the few in developing risk prediction models in a low-\nresource setting. Only one published study presented the development of a model for preterm \nbirth prediction in Ethiopia [36]. Despite reporting good model performance, this study presents \nan important limitation; it predicts preterm birth retrospectively using all information available in \nhindsight (e.g. events such as premature rupture of membranes), while our aim is to assess \nwhether early prediction is possible to inform preventive interventions, leading us to fix a time \npoint for prediction (28 weeks of gestation). Moreover, their study was conducted using data \nfrom a hospital-based cohort, likely to be composed of women at higher a priori risk of adverse \noutcomes. \nStudies carried out in Ethiopia identified risk factors for preterm delivery including obstetric \nconditions, socio-demographic characteristics, urinary and vaginal infections, and hypertensive \ndisorders [37,38]. We included all these factors to build the most accurate models with the \navailable data. Among all predictors, neonatal sex was assigned high importance in the decision \ntree models despite the small difference in prevalence between boys and girls. This is consistent \nwith higher rates of preterm birth for male fetuses in other studies [39,40]. Nevertheless, \nprediction studies like this aim to characterize prognosis and to anticipate or forecast an \noutcome. \nWe simulated the conditional distribution of CL and FFN to test whether access to these \nmeasurements may improve models’ predictive ability. However, the improvement of the models \nwas negligible, indicating that women who are identified as high-risk via CL or FFN could also \nbe identified as high risk via other predictors. Although both CL and FFN are among the most \nused indicators to identify high risk pregnancies for preterm delivery in clinical practice, their \nmeasurement is not always recommended in a priori low-risk populations [15,41]. Our results do \nnot support the allocation of resources to CL and FFN measurement in low-resource settings to \npredict risk of preterm birth. Similarly, other studies observed a poor predictive power of CL and \nFFN in the absence of additional maternal predictors [42].\n \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted December 24, 2022. ; https://doi.org/10.1101/2022.11.04.22281948doi: medRxiv preprint \n\n10 \n \nThe results of this study have important research-related and public health implications. Given \nthe poor performance of all available predictive models, it is fundamental to continue research on \nthe underlying causes of preterm delivery, better understanding the pathways between different \nrisk factors and preterm birth, in order to predict and prevent preterm births in the future. Most \ncases still occur among women without any known risk factor [14,15]. It is crucial to look for \nnew indicators and biomarkers of preterm delivery. Genetic factors, immunological biomarkers \nand protein expression are showing promising results [43,44]. There may be value in exploring \nthe use of ‘omics’ since no biomarkers predictive of preterm birth have yet been identified [45]. \nWhile all settings can benefit from such technologies, from an equity perspective it may be \nespecially important to ensure availability in low-resource settings where the survival of preterm \ninfants is lower and identifying high-risk women can enable targeted preventative interventions. \nPredictive algorithms with modest performance could be used to identify pregnancies at a very \nlow risk of preterm delivery, thus excluding them from interventions. However, a large \nproportion of women at a low risk will still be targeted in those interventions due to the low \nspecificity of the models. In Ethiopia, recommending pregnant women to stay in maternity \nwaiting homes is part of the birth preparedness strategy, though it has not been demonstrated to \nimprove pregnancy outcomes [46,47]. Targeting the recommendation of staying in maternity \nwaiting homes to a reduced number of individuals would increase the cost-effectiveness of the \nintervention and improve the pregnancy experience of some women.\n \nOur findings should be interpreted considering some limitations. Similar to most longitudinal \nstudies, there was study attrition. To address loss to follow-up, we adjusted time-to-event models \nfor censoring and created a “missing” category for all predictors with missing data. The use of a \ncomposite outcome that included all preterm deliveries regardless of vital status of the fetus or \nneonate did not enable the models to separately predict the risk of having a live preterm baby \nfrom the risk of having a preterm stillbirth. However, the use of a combined outcome allows us \nto address competing risks of preterm stillbirths, a common limitation of other available \nprediction models. Despite these limitations, our study fills an evidence gap by exploring \nprediction of preterm delivery during the first 28 weeks of gestation in a resource-limited setting \nwith important restrictions in data availability. We considered a comprehensive selection of >70 \npredictors and tested five different algorithms: Accelerated Failure Time, Cox, LTRCART and \ntwo decision tree ensembles. We acknowledge that the classification tree ensemble is not \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted December 24, 2022. ; https://doi.org/10.1101/2022.11.04.22281948doi: medRxiv preprint \n\n11 \n \nrecommended for data with censoring or truncation. However, we fit this model together with the \nother four algorithms for the purpose of being fully exhaustive in our effort to explore all \npotential methodological options to achieve our study aim of developing an accurate algorithm. \nThe results of all of them show that the difficulty of predicting preterm delivery is robust to \npotential variation in the process of constructing risk models. \nIn settings with low coverage of antenatal care and limited resources to perform ultrasound and \nbiomarker measurements, predicting risk of preterm delivery remains a major challenge. New \nindicators of preterm delivery may be necessary to enable targeted interventions.\n \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted December 24, 2022. ; https://doi.org/10.1101/2022.11.04.22281948doi: medRxiv preprint \n\n12 \n \nREFERENCES \n1 Chawanpaiboon S, Vogel JP, Moller A-B, Lumbiganon P, Petzold M, Hogan D, et al. \nGlobal, regional, and national estimates of levels of preterm birth in 2014: a systematic \nreview and modelling analysis. The Lancet Global Health. 2019;7:e37-e46. \n2 Cobo T, Kacerovsky M, Jacobsson B. Risk factors for spontaneous preterm delivery. \nInternational journal of gynaecology and obstetrics: the official organ of the International \nFederation of Gynaecology and Obstetrics. 2020;150:17-23. \n3 Goldenberg RL, Culhane JF, Iams JD, Romero R. Epidemiology and causes of preterm \nbirth. Lancet (London, England). 2008;371:75-84. \n4 Kim JI, Lee JY. Systematic Review of Prediction Models for Preterm Birth Using \nCHARMS. Biol Res Nurs. 2021;23:708-22. \n5 Meertens LJE, van Montfort P, Scheepers HCJ, van Kuijk SMJ, Aardenburg R, \nLangenveld J, et al. Prediction models for the risk of spontaneous preterm birth based on \nmaternal characteristics: a systematic review and independent external validation. Acta \nobstetricia et gynecologica Scandinavica. 2018;97:907-20. \n6 Stock SJ, Horne M, Bruijn M, White H, Boyd KA, Heggie R, et al. Development and \nvalidation of a risk prediction model of preterm birth for women with preterm labour \nsymptoms (the QUIDS study): A prospective cohort study and individual participant data \nmeta-analysis. PLoS Med. 2021;18:e1003686. \n7 van de Mheen L, Schuit E, Lim AC, Porath MM, Papatsonis D, Erwich JJ, et al. \nPrediction of Preterm Birth in Multiple Pregnancies: Development of a Multivariable \nModel Including Cervical Length Measurement at 16 to 21 Weeks’ Gestation. Journal of \nObstetrics and Gynaecology Canada. 2014;36:309-19. \n8 Zhang J, Pan M, Zhan W, Zheng L, Jiang X, Xue X. Two-stage nomogram models in \nmid-gestation for predicting the risk of spontaneous preterm birth in twin pregnancy. \nArch Gynecol Obstet. 2021;303:1439-49. \n9 Lee KJ, Yoo J, Kim YH, Kim SH, Kim SC, Kim YH, et al. The Clinical Usefulness of \nPredictive Models for Preterm Birth with Potential Benefits: A KOrean Preterm \ncollaboratE Network (KOPEN) Registry-Linked Data-Based Cohort Study. Int J Med \nSci. 2020;17:1-12. \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted December 24, 2022. ; https://doi.org/10.1101/2022.11.04.22281948doi: medRxiv preprint \n\n13 \n \n10 Lee SM, Park KH, Jung EY, Cho SH, Ryu A. Prediction of spontaneous preterm birth in \nwomen with cervical insufficiency: Comprehensive analysis of multiple proteins in \namniotic fluid. J Obstet Gynaecol Res. 2016;42:776-83. \n11 Zhu YZ, Peng GQ, Tian GX, Qu XL, Xiao SY. New model for predicting preterm \ndelivery during the second trimester of pregnancy. Scientific reports. 2017;7:11294. \n12 Gravett MG, Rubens CE, Nunes TM, Group TGR. Global report on preterm birth and \nstillbirth (2 of 7): discovery science. BMC pregnancy and childbirth. 2010;10. \n13 Lawn JE, Gravett MG, Nunes TM, Rubens CE, Stanton C, Group GR. Global report on \npreterm birth and stillbirth (1 of 7): definitions, description of the burden and \nopportunities to improve data. BMC pregnancy and childbirth. 2010;10:S1. \n14 Mercer BM, Goldenberg RL, Moawad AH, Meis PJ, Iams JD, Das AF, et al. The preterm \nprediction study: effect of gestational age and cause of preterm birth on subsequent \nobstetric outcome. National Institute of Child Health and Human Development Maternal-\nFetal Medicine Units Network. Am J Obstet Gynecol. 1999;181:1216-21. \n15 Rozenberg P. Universal cervical length screening for singleton pregnancies with no \nhistory of preterm delivery, or the inverse of the Pareto principle. BJOG : an international \njournal of obstetrics and gynaecology. 2017;124:1038-45. \n16 Chan GJ, Hunegnaw BM, Van Wickle K, Mohammed Y, Hunegnaw M, Bekele C, et al. \nBirhan maternal and child health cohort: a study protocol. BMJ Open. 2021;0. \n17 Bekele D, Hunegnaw BM, Bekele C, Van Wickle K, Tadesse F, Goddard FGB, et al. \nCohort Profile: The Birhan Health and Demographic Surveillance System. Int J \nEpidemiol. 2022;51:e39-e45. \n18 WHO. Preterm birth. 2018. Available: https://www.who.int/news-room/fact-\nsheets/detail/preterm-birth. Accessed: 16 December 2022. \n19 Blencowe H, Cousens S, Jassir FB, Say L, Chou D, Mathers C, et al. National, regional, \nand worldwide estimates of stillbirth rates in 2015, with trends from 2000: a systematic \nanalysis. The Lancet Global Health. 2016;4:e98-e108. \n20 Chan GJ, Goddard FGB, Hunegnaw BM, Mohammed Y, Hunegnaw M, Haneuse S, et al. \nEstimates of Stillbirths, Neonatal Mortality, and Medically Vulnerable Live Births in \nAmhara, Ethiopia. JAMA Netw Open. 2022;5:e2218534. \n21 Therneau TM, Grambsch PM. Modeling Survival Data: Extending the Cox Model. New \nYork: Springer; 2000. \n . CC-BY 4.0 International licenseIt is made available under a \n is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.(which was not certified by peer review)preprint \nThe copyright holder for thisthis version posted December 24, 2022. ; https://doi.org/10.1101/2022.11.04.22281948doi: medRxiv preprint \n\n14 \n \n22 Jackson C. flexsurv: A Platform for Parametric Survival Modeling in R. Journal of \nStatistical Software. 2016;70:1-33. \n23 Fu W, Simonoff JS. Survival trees for left-truncated and right-censored data, with \napplication to time-varying covariate data. Biostatistics. 2017;18:352-69. \n24 Chen T, Guestrin C. XGBoost. Proceedings of the 22nd ACM SIGKDD International \nConference on Knowledge Discovery and Data Mining2016. p. 785-94. \n25 Hastie T, Tibshirani R, Friedman J. The Elements of Statistical Learning. Data Mining, \nInference, and Prediction. 1st edition ed. New York: Springer Science; 2001. \n26 Iams JD, Casal D, McGregor JA, Goodwin TM, Kreaden US, Lowensohn R, et al. Fetal \nfibronectin improves the accuracy of diagnosis of preterm labor. American Journal of \nObstetrics & Gynecology. 1995;173:141-5. \n27 Kagan KO, To M, Tsoi E, Nicolaides KH. Preterm birth: the value of sonographic \nmeasurement of cervical length. BJOG : an international journal of obstetrics and \ngynaecology. 2006;113 Suppl 3:52-6. \n28 Society for Maternal-Fetal Medicine . Electronic address pso, McIntosh J, Feltovich H, \nBerghella V, Manuck T. The role of routine cervical length screening in selected high- \nand low-risk women for preterm birth prevention. Am J Obstet Gynecol. 2016;215:B2-7. \n29 Goldenberg RL, Iams JD, Mercer BM, Meis PJ, Moawad AH, Copper RL, et al. The \nPreterm Prediction Study: The Value ofNew vs Standard Risk Factors in Predicting Early \nand All Spontaneous Preterm Births. American Journal of Public Health. 1998;88:233-8. \n30 Goldenberg RL, Mercer BM, Meis PJ, Copper RL, Das A, McNellis D. The Preterm \nPrediction Study: Fetal Fibronectin Testing and Spontaneous Preterm Birth. Obstetrics & \nGynecology. 1996;87:643-8. \n31 Iams JD, Goldenberg RL, Meis PJ, Mercer BM, Moawad A, Das A, et al. The length of \nthe cervix and the risk of spontaneous premature delivery. New England Journal of \nMedicine. 1996;334:567-72. \n32 Yoo HN, Park KH, Jung EY, Kim YM, Kook SY, Jeon SJ. Non-invasive prediction of \npreterm birth in women with cervical insufficiency or an asymptomatic short cervix \n(