Immune Changes in Pregnancy: Associations with Pre-existing Conditions and Obstetrical Complications at the 20th Gestational Week - A Prospective Cohort Study

BackgroundPregnancy is a complex biological process and serious complications can arise when the delicatebalance between the maternal immune system and the semi-allogeneic fetal immune system isdisruptedor challenged. Gestational diabetes mellitus (GDM), pre-eclampsia, pretermbirth, andlowbirth weight, pose serious threats to maternal and fetal health. Identification of early biomarkersthroughanin-depthunderstandingofmolecularmechanismsiscriticalforearlyintervention. MethodsWe analyzed the associations between 47 proteins involved in inflammation, chemotaxis,angiogenesis,andimmunesystemregulation,maternalandneonatalhealthoutcomes,andthebaselinecharacteristics and pre-existing conditions (diseases and obstetric history) of the mother in aprospective cohort of 1,049 pregnant women around the 20th gestational week. Bayesian linearregressionmodels were usedtoexamine the impact of riskfactors onbiomarkerlevelsandBayesiancause-specific parametric proportional hazards models were usedtoanalyze the effect ofbiomarkerson maternal and neonatal health outcomes. Finally, we evaluated the predictive value of baselinecharacteristics and the 47 proteins using machine-learning models. Shapley additive explanation(SHAP) scores were used to dissect the machine learning models to identify biomarkers mostimportantforpredictions. ResultsAssociations were identified between specific inflammatory markers and existing conditions,including maternal age and pre-pregnancy BMI, chronic diseases, complications from priorpregnancies, and COVID-19 exposure. Smoking during pregnancy significantly affected GM-CSFand 9 other biomarkers. Distinct biomarker patterns were observed for different ethnicities. Inobstetric complications, IL-6 inversely correlated with pre-eclampsia risk, while acute cesareansectionandbirthweight togestational ageratiowerelinkedtomarkerssuchasVEGForPlGF.GDMwas associatedwithIL-1RA,IL-17D,andEotaxin-3.SeverePPHcorrelatedwithCRPandproteinsofthe IL-17family. Predictive modelingusingMSDbiomarkersyieldedROC-AUCvaluesof0.708and0.672for GDMandpre-eclampsia, respectively.SignificantpredictivebiomarkersforGDMincludedIL-1RA and Eotaxin-3, while pre-eclampsia prediction yielded highest predictions when includingMIP-1 β,IL-1RA,andIL-12p70. ConclusionOur study provides novel insights into the interplay between preexisting conditions and immunedysregulationinpregnancy. These findings contribute toourunderstandingofthepathophysiologyof 3 obstetric complications andthe identificationofnovelbiomarkersforearlyintervention(s)toimprovematernalandfetalhealth. 4

During pregnancy there is a complex interaction between the maternal immune system and thesemi-allogeneic fetus 1 . The maternal immune systemis highlyregulatedthroughout pregnancy, andsuccessful pregnancy requires a balance betweentolerance andsuppression. Obstetric complicationsare common, affecting more thanone infour pregnancies, andimmune dysregulationis involvedinthe pathogenesis of a range of complications, including gestational diabetes mellitus (GDM),pre-eclampsia,pretermbirth,andlowbirthweight 2,3 .Immunedysregulationmaybeduetopreexistingor subclinical diseases, complications from prior pregnancies, or viral infections. Indeed, duringpregnancy, the maternal immune system seems to be more challenged by viral infections such asinfluenza 4,5 , respiratory syncytial virus (RSV) 6 , severe acute respiratory syndrome (SARS-CoV) 7,8 ,andMiddle East RespiratorySyndrome (MERS-CoV) 9 . Nonetheless, manypregnancycomplicationsarise without prior knownriskfactors. Consequently, identifyingbiomarkers that are presentearlyinpregnancy prior tothe manifestationof the pathologyis crucial toenable timelyactionandimprovematernalandfetalhealthoutcomes. Pregnancy complications such as preterm birth and early onset pre-eclampsia increase the risk insubsequent pregnancies.Conversely,aprioruncomplicatedpregnancydecreasesriskofcomplicationsin future pregnancies, possibly due to enduring immune alterations favoring fetal tolerance 10 . Animportant aspect of the immune alteration in pregnancy involves the balance of pro- andanti-inflammatory responses; typically, the immune systemmaintains a balance betweenThelper-1(Th-1) cells, associatedwithcell-mediatedimmunityandinflammation, andThelper-2(Th-2) cells,supportinghumoral immunityandlinkedwithanti-inflammatoryresponses 11,12 .Duringpregnancy,thesystemic maternal immune response shifts towards a more anti-inflammatory state whichpromotesfetal tolerance.Disruptionsinthisbalance,particularlyaswitchtowardsapro-inflammatorystate,canlead to pregnancy complications such as pregnancy loss, pre-eclampsia, and pretermbirth 11 . Eventhoughit's important tounderstandthese risks, we don't knowmuchabouthowpreviouspregnanciesaffect certain biomarker levels related to immune regulation in the current pregnancy. However,promisingresearchhighlights the potential forearlydetectionofpretermbirthusingcell-freeRNAora proteinpanel,aswellaspre-eclampsiathroughacombinationofsFlt-1/PlGFandultrasound,amongother methods 13–15 . Adeeper understandingof the impact of previous pregnancies andthe biologicalchanges that occur before labor starts is crucial. It serves as the foundationfor the development ofdiagnostictestsandtoolsaimedatimprovingmaternalandfetalhealthoutcomes. Overall, the link between immune adaptation during pregnancy and obstetric complications iscomplexandmultifaceted. While immune changes are necessaryforasuccessfulpregnancyandfetaldevelopment, they can alsoincrease the riskof obstetric complications. Prophylactic treatment is an 5 option for many obstetric complications, such as low-dose aspirin or progesterone. However, thisrequires precise risk stratification of pregnant women. Early biomarkers of later complicationsfacilitateearlyinterventionsthatcanimprovefetalandmaternalhealth. Here, we present results that build towards a deeper understanding of the immune system, theinterplay between diseases and obstetric history and the immune system and its association withobstetriccomplicationsin1,049pregnantwomen 6

Ethicalapprovals The PREGCOstudy was approved by the Knowledge Centre for Data ProtectionandCompliance,The Capital Regionof Denmark(P-2020-255), andbythe ScientificEthicsCommitteeoftheCapitalRegion of Denmark (journal number H-20022647). All PREGCO participants provided writteninformedconsent.Danish legislation allows for register-based research to be conducted without the consent ofparticipants and without ethical committee approval. Registry data, i.e. the Danish Medical BirthRegistry,washeldatStatisticsDenmark,whichistheDanishnationalstatisticalinstitution. Studydesignandparticipants PREGCO is a prospective cohort of pregnant women from Copenhagen University Hospital,Hvidovre, Denmark, that took place duringthe first pandemic wave inDenmark, between4thApril2020 and 3rd July 2020 16,17 . Participants were invited to participate at the second trimestermalformation scan(gestational weeks 18-22). All pregnant womeninDenmarkare offeredthis scanandmore than90%accept. CopenhagenUniversityHospital Hvidovre serves approximately12%ofpregnant women in Denmark (~7,200 births/year). Participants filled out a questionnaire including,but not limited to, pre-pregnancy body mass index (BMI), smoking, prior pregnancyoutcomes andcomplications,andpre-existingchronicconditions.Baselinecharacteristicsandfollow-upinformationwere obtained fromthe electronic health record, available throughout the whole studyperiod. Thisincluded maternal age, gestational age at birth, sex of the child, multiple pregnancy, birth weight,results from the combined first trimester screening examination (e.g., nuchal fold thickness,pregnancy-associated plasma protein A (PAPP-A), -human choriogonadotropin (β hCG), andβcrown-rumplength), secondtrimestermalformationscan(e.g.,headcircumferenceandfemurlength),obstetric complications (e.g., GDM, pre-eclampsia, acute cesarean section), and mode of delivery(spontaneous delivery, induction of labor, or cesarean section). See Supplementary Table 5 for acomplete list. Furthermore, serumwas collectedatthe12thand20thgestationalweekscanstoscreenfor the presence of SARS-CoV-2 antibodies (IgGand IgM) using the iFlash1800assay 16,17 . Serumsamples from the second trimester malformation scan were also used to measure a panel of 47inflammatorymarkers (describedbelow). The studywas completedandallpregnanciesendedbeforetheapprovalandintroductionofSARS-CoV-2vaccinesinDenmark(27thDecember2020). 7 Comparisonwiththegeneralpopulation The Danish Medical Birth Register (DMBR) 18 was used to compare the PREGCOcohort with allbirths inDenmarkduringthe same period. DMBR, establishedin1973, includes detaileddata onallbirths in Denmark and primarily comprises data from the Danish National Patient Registrysupplementedwithinformationonpre-pregnancyBMI andsmokinginthe firsttrimestercollectedatthe combinedfirst trimester screeningexamination. We comparedmaternal age,pre-pregnancyBMI,smoking,parity,numberofpregnancylosses,andsexofthechild. MesoScaleDiagnosticsinflammatorymarkers The setup for measuring the inflammatory markers has been described by Kjerulff et al 19 .Inflammatory markers from the second trimester malformation scan serumsample were measuredusing the Meso Scale Diagnostics (MSD) V-PLEX Human Biomarker 54-Plex kit, measuring 47biomarkers involvedininflammation, chemotaxis, angiogenesis, andimmune systemregulation(dueto poor quality observed for the TH-17 panel 19 , the panel was excluded). The samples were spreadacross sixteen 96-well plates. We performed preprocessing and median normalization of measuredintensities to minimize batch variation, as described in detail in the Supplementary Text(SupplementaryFigure1). Exposuresandoutcomes We investigatedthe influence of pre-existingconditions (diseases andobstetric history) anddiseaseson biomarker levels at the second trimester malformation scan. This included smoking duringpregnancy, number of prior live births, number of prior pregnancylosses,polycysticovarysyndrome(PCOS), endometriosis, inflammatory bowel disease, GDMin the current or previous pregnancy,pre-eclampsia ina prior pregnancy, vaginal bleedinginearlypregnancy, use of assistedreproductivetechnology (ART, divided into intrauterine insemination (IUI) and in-vitro fertilization (IVF)), andCoronavirus disease 2019(COVID-19) infectioninthe current pregnancy. TheCOVID-19assayandcut-offs have been previously described by Freiesleben et al and Egerup et al 16,17 . Second, we alsoinvestigated the association between inflammatory markers and common obstetric complications,namely GDM, pre-eclampsia, gestational duration, acute cesarean section, and the ratio of birthweight to gestational duration (Supplementary Table 1). To reduce redundancy in the statisticalanalysis, we identifiedhighlycorrelatedmarkersusingtheHobohmIIalgorithm 20 .Wetestedmultiplethresholds and found that a Spearman correlation cut-off of 0.5 sufficiently removed redundantmarkers(SupplementaryFigure2).Forfurtherdetails,seeSupplementaryInformation. 8 Machinelearningmodelstopredictobstetricalcomplications We evaluated the predictive value of the clinical measures and the MSD biomarkers using twomachine learning models, namely a logistic regression model with an L1 penalty and LightGBM.Missing values were imputed using the mean and mode for continuous and categorical variables,respectively, andscaledforthelogisticregressionmodel.Imputationandscalingweredonestrictlyontraining data. LightGBM is a gradient-boosted model that natively handles missing values andcategorical data, and scaling is not needed. As features we used the 47 Meso Scale Diagnosticsinflammatorymarkers, age,BMI,previouslivebirths,fetalabdominalcircumference,fetalabdominaldiameter,fetalheadcircumference,fetalfemurlength,PAPP-AMoM,β hCGMoM,andthedifferencebetween gestational age measured fromcrown-rump length and last menstruation at the combinedfirsttrimesterscreeningexamination. The models’ generalizabilitywas evaluatedusinganestedcross-validation(CV)procedure.Here,theNested CVinvolved an outer loop for model evaluation (interval validation) and an inner loopforhyperparameter optimization(developmentdata).Theouterloopwasafive-foldstratifiedCV,andtheinner loop was a five-foldstratifiedCVrepeatedfive times. Hyperparameters were optimizedusingthe Optuna 21 frameworkwiththe Tree-structuredParzenEstimatoralgorithm.SupplementaryTable2provides the ranges of hyperparameters that were explored. A total of 500 hyperparametercombinations were evaluated across the inner CVloops toidentifythe most effective configurationfor the model. The models were optimizedtominimize the binarycross entropy(BCE).Weselectedthe model with the lowest average inner CV BCE. We evaluated the Area Under the ReceiverOperating Characteristic Curve (ROC-AUC) andthe area under the precision-recall curve (AUPRC)byaveragingthe scores fromtheouterCV.FeatureimportancewasevaluatedusingShapleyAdditiveExplanations (SHAP) on the hold-out outer fold. The ROC-AUCranges from0.5 (random) to 1.0(perfect). However, incases of severe class imbalance, the ROC-AUCmaybe biased.Therefore,wealso evaluated the AUPRC. The AUPRC ranges from 0 to 1. However, the baseline value(corresponding to a randomclassifier) is equivalent tothe prevalence. 95%Confidence intervals forthe ROC-AUCand AUPRCwere calculatedusinga bootstrapapproach, with1,000repetitions. Weused the linear (for logistic regression) or tree (for LightGBM) explainer algorithm. Results werevisualized as the mean absolute SHAP values. Machine learning pipelines were implemented as asnakemakeworkflow,usingelementsfromOptuna,scikit-learn,andSHAP. 21–23 Statisticalanalysis Baseline characteristics were summarizedas mean(standarddeviation, SD) or median(interquartilerange, IQR), where appropriate. Baseline characteristics were comparedtodata fromthe DMBRfor 9 all births in2020usinga Z-testor test.HeatmapswerecreatedbycalculatingpairwiseSpearman’sχ 2 correlation coefficients. Clustering was done using hierarchical complete clustering. Only MSDbiomarkerswereusedforclustering.The association between prior existing conditions and inflammatory marker levels and Birthweight-Gestational Age ratio was estimated using a Bayesian robust linear regression model toaccommodateoutliers 24 .The associationbetweeninflammatorymarkers andGDM, pre-eclampsia, gestational duration,acutecesareansection, severe postpartumhemorrhage (PPH), andanycomplicationwas estimatedusingacause-specific Bayesian time-to-event model, in which the baseline hazard was modeled using anM-spline 25 . Womenlosttofollow-upwerecensoredattheirlastcontactwiththehospital.Allanalyseswereadjustedforage,pre-pregnancyBMI,andgestationalageatenrollment.Theassociationbetweenprior conditions and inflammatory markers was further adjusted for the ultrasound estimatedgestational age at the first trimester risk assesment scan. Obstetric outcomes were adjusted foroutcome-specific variables that were identifiedbasedonexpert andliterature review. Missingvalueswere imputedusingmultiple imputationas implementedinMICE 26 , withpredictive meanmatching.Chains were runfor 100iterations andfive imputationdata sets were created. Eachimputeddatasetwas analyzedseparatelyinthe Bayesianmodelsandtheposteriorswerethencombined.Conservativeprior distributions were usedfor the Bayesianmodels, providinga regularizingeffect. Estimates arereported as the median and the 95% credible interval (bCI) unless otherwise stated. SeeSupplementaryMethods for a more detaileddescriptionof themethodsandmodels.Allmodelswerefittedusingrstanarm 25 orbrms 24 . 10

Cohortcharacteristics Atotal of 1,064womenwere enrolledinthis studyduringthesecondtrimestermalformationscanat20weeks. The cohortconstitutes75%ofallpregnantwomenaskedtoparticipate(flowchartshowninFigure 1A). Of these, 15were not includeddue tomultiple pregnancyand18were lost tofollow-updue to a change of hospital not using the EPIC electronic platform for medical records or homedelivery. The baseline characteristics for participants are showninTable 1. ComparedtoallbirthsinDenmark in 2020, maternal age, pre-pregnancy BMI, and the number of previous live births arewithin the expected range, albeit they are onaverage 1year older, 0.7BMI units lower, andhave ahigher frequencyof primiparaandpriorpregnancylosses(SupplementaryTable3).33women(3.1%)had elevated SARS-CoV-2 IgGantibody levels indicating COVID-19 infection prior to the secondtrimester malformation scan. During the study, 369 (34.7%) womenexperiencedat least one of thecomplications severe PPH (12.2%), preterm birth (11%), GDM(8.2%), pre-eclampsia (4.3%), oracutecesareansection(11.8%)afterenrollment(Table2). Biomarkerprofileclustering Clusteringanalysisdidnotrevealanystrongrelationshipsbetweenthewomen'sinflammatoryprofilesandtheir priorconditions,diseases,orlatercomplications(Figure2,SupplementaryFigure3a).Whenwe examined the pairwise correlations between markers, we found that they did not necessarilycluster by panel or marker group(SupplementaryFigure 3b). Instead, we observedone large clusterand some minor clusters. The largest cluster, with the strongest correlations, consisted of MCP-4,TARC, IL-8, VEGF-C, andIL-7. MCP-4, TARC, IL-8, VEGF-C, andIL-7areallcytokineslinkedtovariousinflammatorydiseases.The strongest anti-correlation observed was between Flt-1 and VEGF (-0.58, 95%CI -0.62; -0.53,Pearson correlation). By binding VEGF and blocking the membrane-bound receptors, soluble Flt-1functions as a naturally occurring antagonist of VEGF. Soluble Flt-1 also binds placental growthfactor (PlGF) that plays a crucial role in the growthanddevelopment of bloodvessels, particularlyduringpregnancyandfetal development 27 . Dysregulationof the interactionbetweenFlt-1andVEGFhasbeenlinkedtovariousdiseasesanddisorders,includingcancerandcardiovasculardisease. Pre-existingconditions,obstetricalhistory, andmaternalcharacteristics A number of previously existing conditions correlated to the measured inflammatory markers.However, there was no widespread pleiotropy, i.e., each exposure had its own biomarker signature(Figure 3). Amajorityof biomarkers were affectedbymaternal age (15/47) andpre-pregnancyBMI 11 (26/47). Conditions altering expression levels included chronic diseases (endometriosis, PCOS),complications inprior pregnancies (pre-eclampsia, GDM,pretermbirth),andCOVID-19.COVID-19was associatedwitha change inexpressionof twoinflammatorymarkers,IFN- γandIL-13.Smokinginthe current pregnancyhada profoundeffect ontenbiomarkers, includingGM-CSF( =0.05, 95%βbCI 0.02; 0.09). GM-CSFis a knownregulator of fetal growthandthe associationbetweensmokingand increased GM-CSF levels is hypothesized to be through an activation of the EGFRsignalingpathway 28 . GM-CSFwas alsoupregulatedinpregnancieswithafemalechild,albeittoalowerdegree( =0.02, 95%bCI 0.01; 0.03). Altered CRP levels were associatedwithcomplications occurringinβprior pregnancies, e.g.pre-eclampsia( =-0.45,95%bCI-0.89;-0.03)andpretermbirth( =0.76,95%β βbCI 0.28; 1.3). The effect of pre-eclampsia andpretermbirthis mostlikelymedicallyinducedduetopreventive treatment withaspirinandprogesterone, respectively. The number of previous live birthswas also associated with lower CRP levels, which could explain part of the mechanismbehindthereducedriskofpre-eclampsia( =0.16,95%bCI0.06;0.27).β TNF- αlevels were increasedinwomenwithGDMpriortothesecondtrimestermalformationscan(β=0.17, 95%bCI 0.01; 0.34) anddecreasedinwomenwithGDMinprior pregnancies ( =-0.14,95%βbCI -0.27;-0.01).TNF- αisamarkerofinsulinresistanceinpregnancy 29 andourfindingsindicatethatinsulinresistance is not affectedinthe longerterminGDMpregnancies.ThissuggeststhatincreasedTNF- αlevelscanbeattributedtotheacutepathogenicprocess. Ethnicity had a major influence on the variation of inflammatory marker levels, such as CRP,Eotaxin-3, IL-15, IL-16, IP-10, andVEGF-D,amongstothers(Figure3).Thesefindingshighlighttheneed for more research on the role of ethnicity in pregnancy complications and the underlyingmechanisms. This researchcouldinformthe development of personalizedinterventionstoreducetheriskof adverse pregnancyoutcomes bytakingintoaccountreferencerangesmaybespecifictoethnicgroups. Laterobstetriccomplications Following similarity reduction, we estimated the association between 41 markers and 11 differentoutcomes (Figure 4). We foundanumberofuniquemarkersthatwereassociatedwitheachcondition.Increasinglevels of IL-6decreasedthe riskof pre-eclampsia (hazardrate (HR)=0.59, 95%bCI0.31;0.97). Prior evidence clearlyindicatesanassociation,butthedirectionofeffectislessclear 30,31 .Acutecesarean section had four associated markers (IL-4, IL-5, MDC, MIP-1 β) none of which wereassociated with any of the other adverse outcomes. For the birthweight togestational age ratio, wefound five associated markers: bFGF, GM-CSF, PlGF, sICAM-1, and VEGF. The most profoundeffect onbirthweight togestationalageratiowasseenforvaginalbleedinginearlypregnancy,which 12 ledtoanestimateddecrease inpercentile of -0.16(95%bCI -0.06;-0.25).Thismeansthat,awomanexperiencing early hemorrhage wouldbe expectedtobe inthe 34thpercentile (95%bCI 25; 44). Incomparison, PlGF, the inflammatorymarker withthe largest effect, was associatedwithanincreasedpercentileof0.04(95%bCI0.01;0.07)perstandarddeviation.GDMwas associatedwithIL-1RA(HR=1.35, 95%bCI1.04;1.8),IL-17D(HR=0.66,95%bCI0.49;0.89), andEotaxin-3(HR=1.27, 95%bCI1.11;1.43).IL-1RAhaspreviouslybeenshowntoassociatewith GDMand has been suggested as a biomarker useful for diagnosingGDMas a complement toblood glucose measurements, as well as to identify GDMpatients who are at risk of developingpostpartum diabetes 32 . IL-17D is associated with incident type 2 diabetes and progression fromnormoglycemia totype 2diabetes 33 .GDMisassociatedwithseveralchemokinesintheproteinfamilywhereEotaxin-3/CCL26belongs 34,35 .CRP(HR=1.34, 95%bCI 1.04; 1.74),IL-13(HR=1.29,95%bCI1.06;1.54),IL-17B(HR=1.25,95%bCI 1.02; 1.53), and IL-17C(HR=1.23, 95%bCI 1.01, 1.49) were all associatedwithsevere PPH.However, none oftheeffectshadamagnitudesimilartoinvitrofertilization(IVF,HR=2.10,95%bCI1.18;3.66). Predictionofobstetriccomplications We evaluated the prognostic potential of clinical characteristics combinedwithMSDbiomarkers inpredicting five pregnancy-related conditions. The performance of our model was assessed usingROC-AUC values on the development data (inner CV) ranging from0.610 to 0.697, and similarvalues were obtainedinthe interval validation(outerCV),rangingfrom0.584to0.715.Theseresultsindicatethatourmodelgeneralizeswelltonewdata.TheAUPRCvalueswereconsistentlybetterthanrandomguessing,asshowninSupplementaryTable6. Among the pregnancy-related conditions analyzed, pre-eclampsia, GDMand PPH were predictedmost accurately using the MSD biomarkers. Specifically, GDM and pre-eclampsia displayedROC-AUC values of 0.708 (95%CI: 0.644-0.766) and 0.672 (95%CI: 0.580-0.758), respectively,withcorrespondingAUPRCvaluesof0.176(95%CI:0.126-0.238)and0.073(95%CI:0.042-0.115). The SHAP analysis further confirmed the classification of IL-1RA and Eotaxin-3 as potentialpredictive biomarkers forGDM(Figure6A).Additionally,theanalysishighlightedthesignificanceofIL-6andCRP, proinflammatorymarkers that were initiallyoverlookedinthepreliminaryassociationanalysisbuthavebeenpreviouslysuggestedaspredictiveforGDM 36–38 . Although the exploratory analysis yielded only a single probable biomarker for pre-eclampsia, ourpredictive modelingemphasizedthe combinedsignificanceofMIP-1 β,IL-1RA,andIL-12p70.These 13 biomarkers collectively had a greater impact than conventional risk factors such as age and BMI.Notably,thenumberofpriorlivebirthsremainedthemostinfluentialpredictor(Figure6B). Theremainingmodelshadalowabilitytodiscriminateandwerenotinvestigatedfurther. 14

Here we present the biomarker profile of alargepopulation-representativecohortofpregnantwomenaround the 20th gestational week. We show that pregnancy has long-lasting effects and that themolecularlevelrewiringtakesplaceatveryearlystagesinthepregnancy,priortothemanifestationofcomplications at a clinically detectable level. Early changes in biomarkers were associated withdevelopment of obstetric outcomes up to 22 weeks later. Moreover, prior live births, complicationsfromprior births, andsmokingaffect circulatingbiomarker levels, whichmayhaveadirecteffectoncomplications and fetal growth. Machine learning models highlighted the joint significance ofinflammatorymarkers,albeittheperformanceisnotyetadequateforclinicaldeployment. PREGCOis a prospective cohort recruitedatasinglehospital.Enrollmentwashigh,above61%.Thecohort was generally comparable to other pregnancies in Denmark in the same time period. Thedifference inthe frequencyof pregnancylosscanmostlikelybeattributedtoanunderreportingintheDanish National Patient Registry. The cohort was enrolled during a time of lowered activity inDenmarkdue torestrictionsonbothprofessionalandprivatesocialinteractions,hencetheCOVID-19prevalence was therefore lowat this time. Furthermore, enrollment tookplace whenonlythe alphaCOVID-19 variant was present, and later variants may have different effects on the immunesystem 39,40 . Furthermore, the testingregimeandwillingnesstoseekmedicalattentioncouldhavebeendifferent due to the COVID-19 pandemic. Relative to the sample size, the number of biomarkersinvestigated was large and some were highly correlated. We pruned highly correlated markers andutilized Bayesian models with conservative priors to provide regularization to the estimates tomitigate this and investigate issues of collinearity. We evaluated the generalization error of themachine learning models using a nested cross-validation approach, which yielded conservativeestimates. However, wecannotruleoutthatsomeeffectscouldnotbedetectedduetothesamplesize,thelargenumberofvariables,andconservativeBayesianpriors. Pregnancyis a complexprocess that involves manychanges inthe body, includingchallenges totheimmune system. These changes are necessary to allowthe developingfetus togrowandthrive, buttheycanalsoincreasetheriskofobstetriccomplications.Forinstance,GM-CSF,aknownregulatoroffetalgrowth,wasupregulatedinpregnancieswithafemalechildorsmoking.Theassociationbetweensmoking and increased GM-CSF levels is hypothesized to be through an activation of the EGFRsignaling pathway 28 . Furthermore, preterm birth or pre-eclampsia in a previous pregnancy wasassociated with lower levels of CRP. This is most likely medically induced due to preventivetreatmentwithaspirinandprogesterone,respectively. 15 The balance between pro- and anti-inflammatory cytokines is crucial for successful placentation.Pro-inflammatory cytokines, such as TNF- , IL-1, and IL-6, play a role in the recruitment andαactivation of immune cells at the implantation site, which is essential for the establishment of afunctional placenta 41 . IL-6is involvedinthe regulationofangiogenesis,decidualization,andimmunecell migration during earlypregnancy, suggestinga crucial role for IL-6inplacentation 42 . Likewise,elevatedlevels of IL-17Ahave beenassociatedwithpretermlabor andpretermpremature ruptureofmembranes (PPROM) 43 . One proposed mechanism for this is IL-17A's ability to promotepro-inflammatory cytokines and chemokines that can affect the development of the fetal-placentalinterface, especially together with TNF- 43 . In our study, the pro-inflammatory cytokines of theαIL-17-family were associated with disorders that have been linked with a negatively alteredplacentation,suchaspostpartumhemorrhageorpretermbirth.TheassociationofCRP,IL-13,IL-17B,andIL-17Cwithsevere postpartumhemorrhage suggests that theymaybeinvolvedintheregulationofhemostasisandlow-gradeinflammationearlyinpregnancy,leadingtolaterobstetriccomplications.CRP is an acute-phase reactant produced by the liver in response to inflammation, infection, andtissue damage. IL-13is aTh-2cellcytokineinvolvedintheregulationofimmuneresponseandtissuerepair, whereas IL-17BandIL-17Care membersoftheIL-17familyofcytokinesthatareinvolvedinthe regulationof inflammationandimmunity. Alterations inthese biomarkers mayaffect the normalbalance between pro- and anti-inflammatory factors and disrupt the physiological processes duringand after delivery, leading for example to an increased risk of postpartum hemorrhage. Moreimportantly, the biomarkers showed these associations to the outcome around gestational week20,makingthempromisingcandidatesforclinicaluseinearlydetectionandprevention. GDMis associated with increased risk of maternal and fetal morbidityandmortality, as well as anincreased risk of developing type 2 diabetes later in life. The precise mechanisms underlying thedevelopment of GDMare not yet fullyunderstood, but evidencesuggeststhatacomplexinterplayofgenetic and environmental factors is involved. In our study, Eotaxin-3, interleukin 1 receptorantagonist (IL-1RA), and IL-17D were associated with GDM. All three cytokines have beenimplicated in the regulation of immune responses, inflammation, and metabolic regulation. Forexample, elevated levels of Eotaxin-3 have been observed in pregnant women with pre-existingdiabetes 44 . Antagonizing the pro-inflammatory cytokine IL-1, IL-1RA is an anti-inflammatorycytokine 45 . The interplay of IL-1 and IL-1RAandtheir dysregulationappear tobe relatedtotype 2diabetes andGDM 32,46 . However,studyresultsarediverging,asforexampleonestudyhasshownthatthe IL-1RAlevels inGDMpatients are considerablylower thanthose of controls 32 , whereas anotherbiomarker study showed inconsistence with our results, that elevated IL-1RAare associated withGDM 35 . Insummary,ourdatashowthatthelevelsofthesecytokinesarealteredinwomendevelopingGDMlater inpregnancyandthat the balance of pro- andantiinflammatorycytokines is necessarytomaintainanormalglucosemetabolismthroughoutpregnancy. 16 Additionally, the studyhighlights the importanceofconsideringmultiplebiomarkersinassessingriskof obstetric complications, as eachbiomarker was associatedwithonlyoneorafewoftheoutcomes.This suggests that a combinationof biomarkers maybe necessarytoaccuratelyidentifywomenwhoareatanincreasedriskofdevelopingcomplications,supportedbythemachinelearningmodels. Once they arise, many obstetric complications cannot be reversed or treated. Perinatal medicinetherefore aims to identify high-risk populations early on, to ideally use therapies to reduceunfavorable maternalandfetaloutcomes 47 .DietaryadaptationsandinsulinforGDM,antepartumfetalmonitoring for stillbirth, aspirin for pre-eclampsia, and progesterone for pretermdeliveryare a fewexamples of suchtherapies that have beenproposedinhigh-riskpopulations 48–53 .Assuch,ithasbeenproposed that the typical care pyramidshouldbe reversed, withthe primaryattentionshiftingtotheearly rather than later stages of pregnancy 47 . The presentedbiomarkers andbiomarker combinationscanpotentiallybe usedtofurther improve this pyramidof care byprovidinghigh-qualitycare tothepatientsatrisk. 17

ThestudywassupportedbyfundingfromtheNovoNordiskFoundation(grantagreementsNNF14CC0001andNNF17OC0027594).TheworkwasalsocarriedoutasapartoftheBRIDGETranslationalExcellenceProgramme(bridge.ku.dk)attheFacultyofHealthandMedicalSciences,UniversityofCopenhagen,fundedbytheNovoNordiskFoundation(NNF18SA0034956). CompetingInterestStatement SB has ownerships in Hoba Therapeutics Aps, Novo Nordisk A/S, Lundbeck A/S, and managingboard memberships in Proscion A/S. HSNreceived personal payment or honoraria for lectures andpresentationsfromFerringPharmaceuticals,Merck,AstraZeneca,CookMedical,andIbsaNordic. 18 Literature 1LaRoccaC,CarboneF,LongobardiS,MatareseG.Theimmunologyofpregnancy:regulatoryTcellscontrolmaternalimmunetolerancetowardthefetus. Immunol Lett2014; 162:41–8.2OakleyL,PennN,PipiM,Oteng-NtimE,DoyleP.RiskofAdverseObstetricandNeonatalOutcomesbyMaternalAge:QuantifyingIndividualandPopulationLevelRiskUsingRoutineUKMaternityData. PLOSONE2016; 11:e0164462.3KalagiriRR,CarderT,ChoudhuryS, et al.InflammationinComplicatedPregnancyandItsOutcome. AmJPerinatol2016; 33:1337–56.4CallaghanWM,CreangaAA,JamiesonDJ.Pregnancy-RelatedMortalityResultingFromInfluenzaintheUnitedStatesDuringthe2009–2010Pandemic. Obstet Gynecol2015; 126:486–90.5ObohoIK,ReedC,GargiulloP, et al.BenefitofEarlyInitiationofInfluenzaAntiviralTreatmenttoPregnantWomenHospitalizedWithLaboratory-ConfirmedInfluenza. JInfect Dis2016; 214:507–15.6RiccòM,FerraroP,CorradoS,ZaniboniA,SattaE,RanzieriS.RespiratorySyncytialVirusinPregnantWomen:SystematicReviewandMeta-Analysis. Women2022; 2:147–60.7FathiM,VakiliK,DeraviN, et al.Coronavirusdiseasesandpregnancy:COVID-19,SARS,andMERS. Przegl Epidemiol2020; 74:276–89.8VillarJ,ContiCPS,GunierRB, et al.Pregnancyoutcomesandvaccineeffectivenessduringtheperiodofomicronasthevariantofconcern,INTERCOVID-2022:amultinational,observationalstudy. TheLancet2023; 401:447–57.9AlfarajSH,Al-TawfiqJA,MemishZA.MiddleEastRespiratorySyndromeCoronavirus(MERS-CoV)infectionduringpregnancy:Reportoftwocases&reviewoftheliterature. JMicrobiol Immunol Infect2019; 52:501–3.10 DeshmukhH,WaySS.ImmunologicalBasisforRecurrentFetalLossandPregnancyComplications. AnnuRevPathol MechDis2019; 14:185–210.11 MorG,AldoP,AlveroAB.Theuniqueimmunologicalandmicrobialaspectsofpregnancy.Nat RevImmunol2017; 17:469–82.12 AghaeepourN,GanioEA,McilwainD, et al.Animmuneclockofhumanpregnancy. SciImmunol2017; 2:eaan2946.13 RasmussenM,ReddyM,NolanR, et al.RNAprofilesrevealsignaturesoffuturehealthanddiseaseinpregnancy. Nature2022; 601:422–7.14 StepanH,HundM,AndraczekT.CombiningBiomarkerstoPredictPregnancyComplicationsandRedefinePreeclampsia. Hypertension2020; 75:918–26.15 LeowSM,DiQuinzioMKW,NgZL, et al.Pretermbirthpredictioninasymptomaticwomenatmid-gestationusingapanelofnovelproteinbiomarkers:thePredictionofPreTermLabor(PPeTaL) study. AmJObstet Gynecol MFM2020; 2:100084.16 EgerupP,FichOlsenL,ChristiansenA-MH, et al.SevereAcuteRespiratorySyndromeCoronavirus2(SARS-CoV-2)AntibodiesatDeliveryinWomen,Partners,andNewborns. ObstetGynecol2021; 137:49.17 laCourFreieslebenN,EgerupP,HviidKVR, et al.SARS-CoV-2infirsttrimesterpregnancy:acohortstudy. HumReprod2021; 36:40–7.18 BliddalM,BroeA,PottegårdA,OlsenJ,Langhoff-RoosJ.TheDanishMedicalBirthRegister. EurJEpidemiol2018; 33:27–36.19 KjerulffB.TheDanishBloodDonorStudy(DBDS)10KInflammatoryBiomarkerCohort:influenceofsex,age,BMI,andsmokingon47circulatinginflammatorybiomarkersin9,876healthyindividuals.2022; Inreview.20 HobohmU,ScharfM,SchneiderR,SanderC.Selectionofrepresentativeproteindatasets.ProteinSci1992; 1:409–17.21 AkibaT,SanoS,YanaseT,OhtaT,KoyamaM.Optuna:ANext-generationHyperparameterOptimizationFramework.In:Proceedingsofthe25thACMSIGKDDInternationalConferenceonKnowledgeDiscovery&DataMining.NewYork,NY,USA:AssociationforComputingMachinery,2019:2623–31.22 PedregosaF,VaroquauxG,GramfortA, et al.Scikit-learn:MachineLearninginPython. JMachLearnRes2011; 12:2825–30. 19 23 LundbergSM,LeeS-I.AUnifiedApproachtoInterpretingModelPredictions.In:AdvancesinNeuralInformationProcessingSystems.CurranAssociates,Inc.,2017.https://proceedings.neurips.cc/paper_files/paper/2017/hash/8a20a8621978632d76c43dfd28b67767-Abstract.html(accessedJuly14,2023).24 BürknerP-C.brms:AnRPackageforBayesianMultilevelModelsUsingStan. JStat Softw2017; 80:1–28.25 BrillemanSL,ElciEM,NovikJB,WolfeR.BayesianSurvivalAnalysisUsingtherstanarmRPackage.2020;publishedonlineFeb22.DOI:10.48550/arXiv.2002.09633.26 BuurenSvan,Groothuis-OudshoornK.mice:MultivariateImputationbyChainedEquationsinR. JStat Softw2011; 45:1–67.27 WewersTM,SchulzA,NolteI,PavenstädtH,BrandM,MarcoGSD.CirculatingSolubleFms-likeTyrosineKinaseinRenalDiseasesOtherthanPreeclampsia. JAmSocNephrol2021; 32:1853–63.28 FuY-Y,NergardJC,BarnetteNK,WangY-L,ChaiKX,ChenL-M.Proteasomeinhibitionaugmentscigarettesmoke-inducedGM-CSFexpressionintrophoblastcellsviatheepidermalgrowthfactorreceptor. PloSOne2012; 7:e43042.29 KirwanJP,Hauguel-DeMouzonS,LepercqJ, et al.TNF- αIsaPredictorofInsulinResistanceinHumanPregnancy. Diabetes2002; 51:2207–13.30 VinceGS,StarkeyPM,AustgulenR,KwiatkowskiD,RedmanCWG.Interleukin-6,turnournecrosisfactorandsolubleturnournecrosisfactorreceptorsinwomenwithpre-eclampsia. BJOGInt JObstet Gynaecol1995; 102:20–5.31 KaumaSW,WangY,WalshSW.Preeclampsiaisassociatedwithdecreasedplacentalinterleukin-6production. JSocGynecol Investig1995; 2:614–7.32 KatraP,DerekeJ,NilssonC,HillmanM.PlasmaLevelsoftheInterleukin-1-ReceptorAntagonistAreLowerinWomenwithGestationalDiabetesMellitusandAreParticularlyAssociatedwithPostpartumDevelopmentofType2Diabetes. PLOSONE2016; 11:e0155701.33 BrahimajA,LigthartS,GhanbariM, et al.Novelinflammatorymarkersforincidentpre-diabetesandtype2diabetes:theRotterdamStudy.DOI:10.1007/s10654-017-0236-0.34 LiuH,LiuA,KamingaAC,McDonaldJ,WenSW,PanX.ChemokinesinGestationalDiabetesMellitus. Front Immunol2022; 13:705852.35 HartPMB,StephensonNL,ScimeNV,ToughSC,SlaterDM,ChaputKH.Secondtrimestercytokineprofilesassociatedwithgestationaldiabetesandhypertensivedisordersofpregnancy.PLOSONE2022; 17:e0279072.36 Roleofinterleukin-6(IL-6)inpredictinggestationaldiabetesmellitus.https://www.ogscience.org/journal/view.php?doi=10.5468/ogs.20020(accessedJuly24,2023).37 DiFilippoD,WanniarachchiT,WeiD, et al.Thediagnosticindicatorsofgestationaldiabetesmellitusfromsecondtrimestertobirth:asystematicreview. ClinDiabetesEndocrinol2021; 7:19.38 JCM| FreeFull-Text| ProteomicsStudiesinGestationalDiabetesMellitus:ASystematicReviewandMeta-Analysis.https://www.mdpi.com/2077-0383/11/10/2737(accessedJuly24,2023).39 WardIL,BerminghamC,AyoubkhaniD, et al.Riskofcovid-19relateddeathsforSARS-CoV-2omicron(B.1.1.529)comparedwithdelta(B.1.617.2):retrospectivecohortstudy. BMJ2022; 378:e070695.40 DaviesNG,JarvisCI,EdmundsWJ,JewellNP,Diaz-OrdazK,KeoghRH.Increasedmortalityincommunity-testedcasesofSARS-CoV-2lineageB.1.1.7. Nature2021; 593:270–4.41 RaoVA,KurianNK,RaoKA.Cytokines,NKcellsandregulatoryTcellfunctionsinnormalpregnancyandreproductivefailures. AmJReprodImmunol2023; 89:e13667.42 VilotićA,Nacka-AleksićM,PirkovićA,Bojić-TrbojevićŽ,DekanskiD,JovanovićKrivokućaM.IL-6andIL-8:AnOverviewofTheirRolesinHealthyandPathologicalPregnancies.Int JMol Sci2022; 23:14574.43 ItoM,NakashimaA,HidakaT, et al.AroleforIL-17ininductionofaninflammationatthefetomaternalinterfaceinpretermlabour. JReprodImmunol2010; 84:75–85.44 SugimotoM,KondoM,KamimotoY, et al.ChangesinVEGF-relatedfactorsareassociatedwithpresenceofinflammatoryfactorsincarbohydratemetabolismdisordersduringpregnancy.PLOSONE2019; 14:e0220650.45 VolarevicV,Al-QahtaniA,ArsenijevicN,PajovicS,LukicML.Interleukin-1receptor 20 antagonist(IL-1Ra)andIL-1Raproducingmesenchymalstemcellsasmodulatorsofdiabetogenesis.Autoimmunity2010; 43:255–63.46 Böni-SchnetzlerM,DonathMY.IncreasedIL-1 βactivation,theculpritnotonlyfordefectiveinsulinsecretionbutalsoforinsulinresistance?Cell Res2011; 21:995–7.47 NicolaidesKH.TurningthePyramidofPrenatalCare. Fetal DiagnTher2011; 29:183–96.48 GinsbergNA,LevineEM.Impactofaspirinonpreeclampsia. AmJObstet Gynecol2021;224:544–5.49 FeduniwS,GolikD,KajdyA, et al.ApplicationofArtificialIntelligenceinScreeningforAdversePerinatalOutcomes—ASystematicReview. Healthcare2022; 10:2164.50 Conde-AgudeloA,RomeroR.Doesvaginalprogesteronepreventrecurrentpretermbirthinwomenwithasingletongestationandahistoryofspontaneouspretermbirth?Evidencefromasystematicreviewandmeta-analysis. AmJObstet Gynecol2022; 227:440-461.e2.51 ACOGPracticeBulletinNo.190:GestationalDiabetesMellitus. Obstet Gynecol2018; 131:e49–64.52 AmericanCollegeofObstetriciansandGynecologists’CommitteeonObstetricPractice,SocietyforMaternal-FetalMedicine.IndicationsforOutpatientAntenatalFetalSurveillance:ACOGCommitteeOpinion,Number828. Obstet Gynecol2021; 137:e177–97.53 ManagementofStillbirth:ObstetricCareConsensusNo,10. Obstet Gynecol2020; 135:e110–32.54 MerinoGA,FresnoC,NettoF,NettoED,PrattoL,FernándezEA.TheimpactofqualitycontrolinRNA-seqexperiments. JPhysConf Ser2016; 705:012003.55 JohnsonWE,LiC,RabinovicA.AdjustingbatcheffectsinmicroarrayexpressiondatausingempiricalBayesmethods. Biostatistics2007; 8:118–27.56 Datanormalizationandstandardization.Olink.https://olink.com/application/data-normalization-and-standardization/(accessedDec29,2022).57 BolstadBM,IrizarryRA,ÅstrandM,SpeedTP.Acomparisonofnormalizationmethodsforhighdensityoligonucleotidearraydatabasedonvarianceandbias. Bioinformatics2003; 19:185–93.58 HongM-G,LeeW,NilssonP,PawitanY,SchwenkJM.MultidimensionalNormalizationtoMinimizePlateEffectsofSuspensionBeadArrayData. JProteomeRes2016; 15:3473–80.59 BrillemanSL,ElciEM,NovikJB,WolfeR.BayesianSurvivalAnalysisUsingtherstanarmRPackage.2020;publishedonlineFeb22.DOI:10.48550/arXiv.2002.09633.60 GoodrichB,GabryJ,AliI,BrillemanS.rstanarm:BayesianappliedregressionmodelingviaStan.2020.https://mc-stan.org/rstanarm. 21 Figures Figure 1: (a) Flowchart for inclusion, (b) Timepoints for data collection. Data fromthe electronichealthrecordwasobtainedthroughoutthepregnancyandpostpartum. 22 Figure2:HeatmapofphenotypesacrossparticipantsusinghierarchicalclusteringbasedonMSDbiomarkers.Phenotypesaredividedintomaternalcharacteristics,pregnancy-relatedoutcomesforthecurrentpregnancy(observedbeforeoratinclusion),fetalcharacteristics,pregnancy-relatedoutcomesforpriorpregnancies,priordiseasestothecurrentpregnancy,andphenotypesobservedafterinclusion. BMI,bodymassindex;C-section,cesareansection;DM,diabetesmellitus;GA,gestationalage;GDM,gestationaldiabetesmellitus;IBD,irritableboweldiseases;IUI,intrauterineinsemination;IVF, 23 invitrofertilization;MoM,multiplesofmedians;PCOS,polycysticovarysyndrome;PE,Pre-eclampsia;PPH,postpartumhemorrhage;SGA,smallforgestationalage;UTI,urinarytractinfection 24 Figure 3:Previouspregnanciesandpre-existingconditionseffectsoninflammatorymarkers.Effectsizeis increaseordecreaseinstandard deviations.Onlyassociationswhere the95%BayesianCredibleIntervaldoesnotincludezeroareshown. BMI,bodymassindex;GA,gestationalage;GDM,gestationaldiabetesmellitus;IUI,intrauterineinsemination;IVF,invitrofertilization;MOM,multiplesofmedians;PCOS,polycysticovarysyndrome;PE,pre-eclampsia;PPH,postpartumhemorrhage 25 Figure 4:Associationsbetweenmarkersandlaterobstetricaloutcomes.Theeffectsizeisthelog-hazard rateforalloutcomes,exceptBirthWeight-GestationalAgeRatio(BWGA).Theeffectof biomarkerson BWGAis a changein percentilefromthe50thpercentile.Onlyassociationswherethe95%BayesianCredibleIntervaldoesnotincludezeroareshown. BMI,bodymassindex;GA,gestationalage;IUI,intrauterineinsemination;IVF,invitrofertilization 26 Figure 5:SummarisedSHAPvaluesforeachfeature category(red=clinicalmeasure,blue=MSDbiomarker)acrossalloutcomesinvestigatedintheprognosticmodel. PPH,postpartumhemorrhage 27 Figure 6:SHAPvaluesforallprognosticfeaturesforgestationaldiabetesmellitus(GDM)andpre-eclampsia,respectively. Thetoptenfeatureswiththehighestmedianvalueforeachoutcomeare indicatedwithlightcolors.OnlyIL-1RAandCRPwerepartofbothtoptenlists.Features are colored basedon thefeature categoryas clinicalmeasure (red)or MSDbiomarker(blue). BMI, pre-pregnancy body mass index;MoM, multiple of medians; SHAP, Shapley additiveexplanation 28 Tables Table1:Baselinecharacteristicsforthe1,049womenincludedinthePREGCOcohort. Variable PREGCO(n=1,049) Maternalage,years* 31.7(4.5) Gestationalageatinclusion,days** 141(140-143) Pre-pregnancyBMI,kg/m 2 * 24.1(4.8)(na=9) Smokinginpregnancy 40(3.8%) Parity 0 576(54.9%) 1 375(35.7%) 2 81(7.7%) 3+ 17(1.6%) Numberofpregnancylosses 0 810(77.2%) 1 194(18.5%) 2 29(2.8%) 3 16(1.5%) Sexofchild,male 521(50.5%)(na=18) HCG***β 1.2(0.8)(na=35) PAPP-A*** 1.3(0.8)(na=35) COVID-19antibodies 34(3.2%) Conceptionmethod Spontaneous 944(88.7%) AssistedReproductiveTechnology(IVForIUI) 120(11.3%) Priorexistingchronicconditions PCOS 23(2.2%) DiabetesMellitus(type1ortype2) 2(0.2%) Endometriosis 9(0.8%) Complicationsinpreviouspregnancy Pretermbirth 19(1.8%) Pre-eclampsia 24(2.3%) PPH(>500mLbloodloss) 20(1.9%) GDM 12(1.1%)Symbols: *mean(sd); **median(IQR); ***multipleofmedians(MoM).Missingvaluesdenotedinna.Ifnomentionofmissingvalues,thevalueiscomplete. 29 Abbreviations: BMI, body mass index, GDM, gestational diabetes; IUI, intrauterine insemination;IVF,invitrofertilization;PCOS,polycysticovarysyndrome;PPH,postpartumhemorrhage 30 Table2:Outcomecharacteristics Variable PREGCO Birthweight(g)* 3499(554) Gestationalage(days)** 280(273;287) Preterm(<37+0) 116(11%) Birthweight/Gestationalage** 12.6(1.8) Severepostpartumhemorrhage(PPH) 130(12.2%) Pre-eclampsia 46(4.3%) Gestationaldiabetesmellitus(GDM) 87(8.2%) Modeofdelivery Spontaneousvaginalbirth 634(60.7%) Acutecesareansection 126(11.8%) Electivecesareansection 95(8.9%) Inducedbirth 235(22.1%) Anycomplication (SeverePPH,pretermbirth,GDM,pre-eclampsia,acutecesareansection) 369(34.7%) Symbols: *mean(sd);**median(IQR).Abbreviations:GDM,gestationaldiabetesmellitus;PPH,postpartumhemorrhage 31 SupplementaryInformation MesoScaleDiagnosticsdatanormalization We measured47inflammatorymarkers usingtheV-PLEXHumanBiomarker54-PlexKitfromMesoScale Diagnostics (MSD). Due to low-qualityassayvalidation, we excludedall measurements fromtheTh17panel,resultingin47assaysonsixpanelstobeincludedintheanalysis. To address the batcheffect exertedbythe individual 96-well plate setup, we exploredthe correctiveeffects of three data pre-processing methods and four normalization methods with the aim ofremovingbatcheffectsatpanellevel.The MSDV-PLEXkit is basedonelectrochemiluminescence technology, where light emissionfromSULFO-TAGlabelsismeasuredaslightintensity(“signal”).Thesignalvaluesarewithinthedynamicrange of the assay linearly associated with the concentration of the measured target. While thisprinciple is used for concentration determination by a measured standard curve, we here used thesignalsdirectlytoavoidanynoiseintroducedbythemeasurementofastandardcurve.Due to the multiplex setup, plate-based batch effects were assessedbypanel, as technical variationwas assumed to be equal across the four to ten assays measured per panel. Our first aimwas toremove any observed variation based on the 16 plates the samples were run on. We comparedthecombination of three pre-processing methods: a) log2-transformation, b) method aplus removal ofoutliers based on principal component analysis (PCA) 54 , and c) method b plus ComBat batchcorrection 55 , and four normalization methods: i) no normalization, ii) median normalization 56 , iii)quantile normalization 57 , andiv)MAnormalization 58 .Toassesstheeffectivenessofnormalization,weperformedavisualinspectionofPCAplots,densityplots,andboxplots.Based on visual inspection of the PCA plots of the log2-transformed data, we defined individualoutlier limits for the six panels removing samples that exceeded these limits. Consequently, weremoved17outliers onAngio1, 13outliers onChem1, 11outliers onCyto1, 41outliers onCyto2,8outliersonPro1,and9outliersonVascu2(seelimitsonSupplementaryFigure2).We foundmediannormalizationtosufficientlyremoveallbatcheffectsobservedatPC1andPC2andyielded overlapping curves with normal or near-normal distributions for the individual assaysobservedondensityandboxplots. Similar results were foundfor MAnormalization,whiletheothermethodsyieldedvaryingpoorerresults. SimilarityreductionofmarkersusingHobohmII Hierarchicalclusteringshowedhighsimilaritybetweenmarkers,indicatingredundancyinthedataset.Toaddress this, we employedthe HobolmII algorithmas outlinedinHobohmet al,1992 20 .WeusedSpearman correlation as the correlationmetric andtestedcut-off levels at 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 32 0.9, and 1 (Supplementary Figure 3). We chose a cut-off of 0.5, resultingin41markers. Usingthiscut-off,wedidnotobserveissuesofcollinearityinanyofthestatisticalmodels. Bayesianregressionmodels For all statistical analyses we employed Bayesian regression models. All models were fit usingrstanarmor brms 24,59,60 . Unless otherwise specified, all models were runfor 10,000iterations (5,000warm-up and 5,000 sampling) with default settings for the sampler. Convergence was assessed bycalculating r-hat statistics, looking for divergences and making sure the sample did not exceedthemaximumtree depthafter warm-up. Amodel hadconvergedifandonlyif(1)allr-hatvalues<1.01,(2) no divergences, and (3) no iterations exceeding the maximumtree depth. All parameters werecentered prior to model fitting and MSD inflammatory markers were standardized so that theinterpretationfollows changes instandarddeviations. Furthermore, we alsoinspectedcorrelations oftheposteriordistributionstoidentifyissuesofcollinearity. Bayesianrobustlinearregression The Bayesian robust linear regression extends the classical ordinary least squares by assuming aStudent t-distribution, thereby accommodating outliers. The degree of freedomis estimateddirectlyfromthedata.Formally, 𝑦 ~ 𝑡(ν, µ, σ)inwhich isthesumovertheinterceptandcovariates,µ µ =𝑏0 + 𝑘 ∑β𝑘 𝑥𝑘 Tocompletethemodel,wespecifyasetofpriorsacrosstheparametersinthemodel, β0 ~𝑁(0, 2.5) β𝑘 ~𝑡(7, 0, 0.5) σ~𝑁(0, 1)ν~𝑁(0, 1) Posterior model checkingwas done bysimulating100draws fromthe posterior andcomparingwiththeobserveddistribution. 33 Cause-specificparametricproportionalhazardsmodels For analyzing the duration, or time to failure, we employed cause-specific parametric proportionalhazards survival models, as implemented in rstanarm 59 . This takes into account that some womenwere lost tofollow-upandarethuscensored,andthattheremaybecompetingoutcomes(e.g.inducedlabor, acute cesarean section, or scheduled cesareansectionis a competingoutcome tospontaneousvaginal birth). We compared two baseline hazards (cubic B-spline and M-spline) for eachoutcomeand visually inspected the estimatedbaseline hazardcurve, versus the observedcurve. Model priorsfollow default settings, except the regression coefficients ( ) which were assigned a moreβ𝑘 conservativeprior, β𝑘 ~ 𝑡(7, 0, 0.5) Thepriorinducesregularizationbyforcingcoefficientstowardszero,i.e.anulleffect. Birthweighttogestationalageratio For the birthweight togestational age ratio, the values were standardizedaccordingtothemeanandstandarddeviationestimatedfromthe DanishMedicalBirthRegistry(DMBR),includingallbirthsin2020(SupplementaryTable4).Fromtheposterior,wedefinedatransformation, 𝑃(𝑋 ≤ 𝑥) = 𝐹(𝑋|µ,σ,ν ) − 0.5 inwhichXis the estimatedvalueofthemarker,andFistheStudent-tcumulativedensitydistributionparametrized by the mean, standard deviation, and degrees of freedomestimated fromthe DMBR.The parameters were estimated using a Bayesian robust linear regression. The resulting valuerepresentsthechangefromthe50thpercentileforaone-unitchangeinthevariable. 34 SupplementaryFigures Supplementary Figure 1: PCAplots for log2-transformed signal values colored byplate for the sixMSDpanels.Greylinesindicateselectedlimitsforoutlierdetection. 35 Supplementary Figure 2: Number of markers remaining after pruning by the HobohmII algorithmaccordingtoeightcut-offlevelsfortheSpearmancorrelationcoefficient. 36 SupplementaryFigure3:(a)Heatmapofwomenxmarkers,(b)Heatmapsofmarkersxmarkers. 37 SupplementaryTables SupplementaryTable1:Outcomedescriptions,definitions,adjustmentsandexclusioncriteria. Outcome Definition Outcome-specificadjustments Exclusioncriteria Pre-eclampsia BloodPressure>140mmhgsystolicand/or90mmHgdiastolic&proteinuria Previousnumberoflivebirths,Smoking,PAPP-AMoM,β HCGMoM,ART(dividedintoIUIorIVF),Ethnicity Pre-gestationalhypertensionGestationalhypertensiondiagnosedbeforeinclusionpre-eclampsiainpreviouspregnancy Spontaneousvaginalbirth Birthnotassistedbycesareansectionorinduction Smoking,Hemorrhageinearlypregnancy,UTIincurrentpregnancy,Conisatio n.a. Gestationalageatbirth Gestationalagecalculatedfromultrasound. Smoking,Hemorrhageinearlypregnancy,UTIincurrentpregnancy,Conisatio n.a. Pretermbirth Gestationalage<37+0 Smoking,Hemorrhageinearlypregnancy,UTIincurrentpregnancy,ConisatioGestationaldiabetesmellitus(GDM) Basedonatwo-hour75gramoralglucosetolerancetest(OGTT);GDMwithglucose≥9.0mmol/lincapillarywholebloodorvenousplasma PCOS,GDMinpriorpregnancy GDMdiagnosedbeforeinclusionDiabetesMellitus Birthweight/gestationalageratio Birthweightrecordedimmediatelyafterbirthbymidwifeordoctor.SeverePostpartumhemorrhage 1000mLbloodlossormorewithin24hours ART(dividedintoIUIorIVF),Hemorrhageinearlypregnancy,CesareansectioninpriorpregnancyAcutecesareansection Deliverybyacutecesareansection, Cesareansectioninpriorpregnancy 38 anydegreeAnycomplication Pre-eclampsia,GDM,SeverePPH,Pretermbirth,oracutecesareansection. Previousnumberoflivebirths,Smoking,PAPP-AMoM,β HCGMoM,ART(dividedintoIUIorIVF),PEinpriorpregnancy,GDMinpriorpregnancy,PCOS,UTIincurrentpregnancy,Cesareansectioninpriorpregnancy,Conisatio,Ethnicity Pre-gestationalhypertensionGestationalhypertensiondiagnosedbeforeinclusionpre-eclampsiainpreviouspregnancyGDMdiagnosedbeforeinclusionDiabetesMellitus(type1or2) Abbreviations: ART, assisted reproductive technology; GDM, gestational diabetes mellitus; IUI,intrauterine insemination; IVF, in vitro fertilization; MoM, multiple of the median; n.a., notapplicable; OGTT, oral glucose tolerance test; PCOS,polycysticovarysyndrome;PE,Pre-eclampsia;PPH,postpartumhemorrhage;UTI,urinarytractinfection 39 SupplementaryTable2:Hyperparametervalues Model Definition Outcome-specificadjustmentsLogisticregression L1penalty 1e-8…1(logspaced) LightGBM Numberofestimator 1…200Learningrate 0.0001…1(logspaced)Maxdepth 3…30Subsample,individuals 0.2…0.8Subsample,feature 0.2…0.8Mininumchildsamples 50…200Maximumtreeleaves 6…50L1penalty [0,0.01,1,2,5,7,10,50,100]L2penalty [0,0.01,1,2,5,7,10,50,100] 40 SupplementaryTable3:ComparisonofPREGCOandtheDanishMedicalBirthRegistry(DMBR). Variable PREGCO(n=1,049) DMBR2020(n=60,573) Difference Maternalage,years* 31.7(4.5) 30.7(4.7) 1.0(0.72;1.27,p<0.001) Pre-pregnancyBMI,m/kg 2 * 24.1(4.8) 24.8(5.4) -0.7(-0.99;-0.41,p<0.001) Smokingduringpregnancy** 40(3.8%) 4,660(7.7%) 0.49(0.35;0.67,p<0.001) Parity** 0 576(54.9%) 22,644(37.3%) p<0.001 1 375(35.7%) 20,421(33.7%) 2 81(7.7%) 10,011(16.5%) 3+ 17(1.6%) 7,507(12.4%) NumberofPregnancyLosses** 0 810(77.2%) 52,145(86%) p<0.001 1 194(18.5%) 6,965(11.5%) 2 29(2.8%) 1,230(2.0%) 3+ 16(1.5%) 282(0.5%) Sexofchild,male** 521(50.5%) 30,983(51%) 0.96(0.86;1.07,p=0.44) *Z-test;** testχ 2 41 SupplementaryTable4:Birthweight-gestationaldurationratioparametersestimatedfromtheDanishMedicalBirthRegistry(DMBR). Parameter Median 95%bCI µ 12.61 12.59-12.62 σ 1.54 1.53-1.56 ν 6.88 6.54-7.24 42 SupplementaryTable5:Variablesandsources Group Name Timepoint measured/recordedandSource MaternalCharacteristics Maternalage Pre-pregnancy,EHR BMI Pre-pregnancy,EHR Smoking Atinclusion,EHR Ethnicity Pre-pregnancy,EHR CurrentPregnancy ART Pre-pregnancy,EHR Gestationalageatinclusion,basedonultrasound Atinclusion,EHR GDMincurrentpregnancy Atinclusion,EHR COVID-19antibodies 8-12thweekscanandatinclusion,EHR Earlyhemorrhage Atinclusion,EHR Fetalcharacteristics Sexofchild Atinclusion,EHR PAPP-A,MoM 8-12thweekscan,EHR HCG,MoMβ 8-12thweekscan,EHR Gestationalage,basedonultrasound 8-12thweekscan,EHR Abdominalcircumference Atinclusion,EHR Headcircumference Atinclusion,EHR Femurlength Atinclusion,EHR Priorpregnancies GDMinpriorpregnancy Atinclusion,Questionnaire Pre-eclampsiainpriorpregnancy Atinclusion,Questionnaire PPHinpriorpregnancy Atinclusion,Questionnaire Pretermlaborinpriorpregnancy Atinclusion,Questionnaire Previousnumberofcesareansections Atinclusion,Questionnaire Previousnumberoflivebirths Atinclusion,Questionnaire Previousnumberofpregnancylosses Atinclusion,Questionnaire Priordiseases Conisatio Atinclusion,Questionnaire Endometriosis Atinclusion,Questionnaire PCOS Atinclusion,Questionnaire 43 Abbreviations: ART, assisted reproductive technology; BMI, body mass index; EHR, ElectronicHealth record; GDM, gestational diabetes mellitus; MoM, multiples of medians; PCOS, polycysticovarysyndrome 44 Supplementary Table 6: Results from cross-validation of machine learning models. The highestranking model for each outcome is highlighted in bold, basedonthe binarycross entropyfromtheinnerCV. Outcome Model Development (innerCV) Interval validation(outerCV) BinaryCrossEntropy ROC-AUC AUPRC ROC-AUC AUPRC AcuteSectio LASSO 0.216 0.638 0.130 0.590(0.515-0.667) 0.086(0.057-0.126) LightGBM 0.216 Anycomplication LASSO 0.599 0.630 0.457 0.615(0.576-0.655) 0.421(0.367-0.477) LightGBM 0.602 GestationalDiabetes LASSO 0.268 LightGBM 0.266 0.697 0.219 0.708(0.644-0.766) 0.176(0.126-0.238) Pre-eclampsia LASSO 0.150 LightGBM 0.149 0.655 0.125 0.672(0.580-0.758) 0.073(0.042-0.115) Pretermbirth LASSO 0.172 LightGBM 0.171 0.610 0.118 0.564(0.460-0.66) 0.071(0.041-0.132) SeverePPH LASSO 0.349 LightGBM 0.349 0.628 0.198 0.587(0.530-0.638) 0.152(0.117-0.193)