A Clinical Prediction Rule for Histological Chorioamnionitis in Preterm Newborns

Background Histological chorioamnionitis (HC) is an intrauterine inflammatory process highly associated with preterm birth and adverse neonatal outcome. HC is often clinically silent and diagnosed postnatally by placental histology. Earlier identification could facilitate treatment individualisation to improve outcome in preterm newborns. Aim Develop a clinical prediction rule at birth for HC and HC with fetal involvement (HCF) in preterm newborns. Methods Clinical data and placental pathology were obtained from singleton preterm newborns (gestational age ≤32.0 weeks) born at Erasmus UMC Rotterdam from 2001 to 2003 (derivation cohort; n = 216) or Máxima MC Veldhoven from 2009 to 2010 (validation cohort; n = 206). HC and HCF prediction rules were developed with preference for high sensitivity using clinical variables available at birth. Results HC and HCF were present in 39% and 24% in the derivation cohort and in 44% and 22% in the validation cohort, respectively. HC was predicted with 87% accuracy, yielding an area under ROC curve of 0.95 (95%CI = 0.92–0.98), a positive predictive value of 80% (95%CI = 74–84%), and a negative predictive value of 93% (95%CI = 88–96%). Corresponding figures for HCF were: accuracy 83%, area under ROC curve 0.92 (95%CI = 0.88–0.96), positive predictive value 59% (95%CI = 52–62%), and negative predictive value 97% (95%CI = 93–99%). External validation expectedly resulted in some loss of test performance, preferentially affecting positive predictive rather than negative predictive values. Conclusion Using a clinical prediction rule composed of clinical variables available at birth, HC and HCF could be predicted with good test characteristics in preterm newborns. Further studies should evaluate the clinical value of these rules to guide early treatment individualisation.


Introduction
Chorioamnionitis is an antenatal inflammatory state of the intrauterine environment strongly associated with prematurity [1]. Around 40% of infants born before 32 weeks gestation have been exposed to chorioamnionitis [1][2][3], which often is a clinically silent process. Exposure to chorioamnionitis is known to affect several organ systems in the fetus [4]. Its presence in placentas from preterm infants has been associated with decreased respiratory distress syndrome, but increased incidences of bronchopulmonary dysplasia (BPD), necrotising enterocolitis (NEC) and neurologic sequelae, including white matter damage and cerebral palsy [2,[4][5][6]. These effects have generally been shown to be more pronounced when additional signs of fetal inflammation including funisitis are present [2,4,5].
Recent studies have suggested differential effects of distinct treatments in infants with varying degrees of chorioamnionitis as opposed to non-exposed infants in the early neonatal period [7][8][9][10]. Thus, perinatal identification of chorioamnionitis-exposed infants could facilitate the development of subgroup-targeted early intervention strategies to improve outcome in this group. Evidence from randomised controlled trials and observational studies suggests that chorioamnionitis-exposed infants may distinctively benefit from increased surfactant dosing [8,10], restrictive use of invasive and prolonged ventilation [8,[11][12][13], as well as postnatal corticosteroid administration to improve BPD-free survival [9].
The gold standard for diagnosing chorioamnionitis is placental histological examination [14]. Unfortunately in current practice, the final results of placental pathology may take days if not weeks, hindering its use as an indicator to guide early postnatal therapy. To overcome this problem, the ability of biological markers to detect chorioamnionitis before or shortly after birth has been investigated. Indeed, several markers and marker patterns in amniotic fluid and umbilical cord blood have been shown to carry predictive value for histological chorioamnionitis [15][16][17]. However, to date their use is experimental, and sensitivity and specificity have generally shown to be at best moderate.
Several clinical variables are well known to be distributed in a differential pattern among preterm infants with and without chorioamnionitis [2,5,18]. In the current study we aimed to develop a clinical prediction rule for both histological chorioamnionitis (HC) and chorioamnionitis with fetal involvement (HCF) at birth in preterm infants, composed solely of clinical variables available at that time.

Ethics statement
The derivation cohort was part of a study approved by the Medical Ethics Committee for Research on Human Subjects of the Erasmus University MC. Written parental consent was obtained. According to Dutch law a waiver for ethical assessment and parental consent was provided by the local Medical Ethical Committee of the Máxima Medical Centre for development and use of the validation cohort, considering that retrospective and anonymised data collection was performed using routinely collected medical chart data solely.

Derivation cohort
The clinical prediction rule was developed in a prospective cohort described previously [2,8,19]. Pregnant women, who delivered between May 2001 and February 2003 in the Erasmus University MC-Sophia Children's Hospital in Rotterdam, Netherlands, at a gestational age of 32.0 weeks or less, were eligible for the study. Newborns were enrolled immediately after delivery when admitted to the neonatal intensive care unit (NICU; level III). Trained research nurses unaware of results of placental histology prospectively collected relevant clinical data. Multiple births were excluded from analysis, as were newborns with severe congenital abnormalities.
Placentas and membranes were fixed in formalin directly after delivery. Sampling was done according to a standard protocol and included at least two membrane rolls, two cross-sections of the cord, and three representative blocks of the placental disk. Tissues were embedded in paraffin until examination. A single pathologist (RRdK) examined all placentas in a blinded fashion for presence of chorioamnionitis and additional fetal inflammatory response, according to the Amniotic Fluid Infection Nosology Committee guidelines [14]. Accordingly, fetal involvement included any of the following: chorionic vasculitis, umbilical phlebitis or vasculitis, (subacute) necrotising funisitis, or concentric umbilical perivasculitis.

Predictors
Clinical variables were considered as potential predictors when they were deemed to be readily available in daily practice, were available in the prospectively collected database of the derivation cohort, and were potentially associated with HC and/or HCFeither positively or negatively -or had been associated with these entities in prior clinical studies. According to these criteria, the following clinical parameters were evaluated as potential predictors of HC and HCF: ethnicity (self-classified); maternal age; gravidity; parity; antenatal steroid administration (betamethasone 12 mg intramuscularly, repeated after 24 hours); preeclampsia (new onset hypertension [blood pressure .140/90 mm Hg or mean arterial pressure .105 mm Hg] with proteinuria); HELLP syndrome (clinical presentation of intravascular haemolysis, elevated liver enzymes, and a low platelet count); preterm premature rupture of membranes (PPROM); clinical chorioamnionitis (maternal temperature .38.0uC with no other focus, and two or more of the following: uterine tenderness, malodorous vaginal discharge, maternal leukocytosis [WBC.15,000 cells/mL], raised serum C-reactive protein [CRP.15 mg/L], maternal tachycardia [.100 bpm], and fetal tachycardia [.160 bpm]); mode of delivery; gestational age (preferably estimated by ultrasonography or otherwise by using the last menstrual period when reliable); gender; birth weight; being small for gestational age (SGA; birth weight ,10 th percentile of the gender-specific mean for gestational age); and placental weight.

Clinical prediction rule development
Clinical prediction rule development methodology was guided by published standards [20]. Univariable analyses were performed to identify relevant clinical variables that differed between the groups ('no HC' versus 'HC', and 'no HCF' versus 'HCF'), using x 2 -test, Student t test, or Mann-Whitney U-test where appropriate (alpha ,0.10). These were entered into a backward logistic regression model to predict either HC or HCF. Before entry, continuous variables were dichotomised based on the most discriminative cut-off value on the ROC curve. The final model was selected using the likelihood ratio method with an alpha level of 0.10. Next, for each predictor in the model the beta value was divided by the lowest beta value in the model and rounded off to the nearest integer. A weighted score was thus obtained for each predictor and used to develop a clinical prediction rule for HC or HCF. ROC curves were computed based on individual composite scores calculated for each patient to determine the optimum cutoff value for prediction, defined as the cut-off with a sensitivity of at least 0.80 and maximum specificity. Relevant test statistics were computed from two-by-two contingency tables. Internal model performance was estimated using leave-one-out cross-validation.

Validation cohort
The clinical prediction rule was externally validated in a cohort of inborn singleton newborns with a gestational age #32.0 weeks admitted to the NICU (level III) of the Máxima Medical Centre, Veldhoven, Netherlands between January 1, 2009 and December 31, 2010. Retrospective data retrieval from maternal and neonatal medical charts was performed for the variables included in both clinical prediction rules. Data were anonymised before storage. Similar to the derivation cohort, placenta sampling and fixation was standardised and presence of chorioamnionitis and additional fetal inflammatory response was scored according to international guidelines [14]. Individual composite scores were calculated for both prediction rules with positive test scores defined by the optimum cut-off values identified in the derivation cohort. Again, relevant test statistics were computed from two-by-two contingency tables. Analyses were performed using SPSS 16.0 software (SPSS, Inc, Chicago, IL).

Clinical prediction rule development
The derivation cohort was composed of all 216 singletons from a cohort of 301 newborns described previously [2,8,19]. Of these, 84 (39%) had HC, while 51 (24%) had additional signs of fetal involvement. Placental pathology and relevant clinical data were available for all mother-newborn pairs. Newborns with HC were more often born vaginally, were of lower gestational age but had a higher birth weight and placental weight as compared to those without HC (Table 1). Accordingly, they were less likely to be SGA. Their mothers more often experienced PPROM and clinical chorioamnionitis, and less often had preeclampsia and HELLP syndrome. The same differences were present between newborns with HCF and those without, with the exception of birth weight, which was not significantly different between these groups.

External validation
The validation cohort was derived from a birth cohort of 206 inborn singleton newborns. Placental pathology was available for 183 (89%) of these. Distribution of the predictor variables over both cohorts is shown in Table 5. Newborns with missing placental pathology were less often delivered by caesarean section (

Discussion
In a well-defined cohort of very preterm newborns reported previously [2,8,19], we developed a clinical prediction rule for histological chorioamnionitis and for histological chorioamnionitis with fetal involvement. Using a simple set of clinical variables generally available in the clinical setting, both HC and HCF could be predicted at birth with high accuracy, with an expected small drop in prediction rule performance at external validation. The prediction rules presented here carry future potential in facilitation of subgroup-targeted early intervention strategies.
The current study has several strengths. Development and validation were performed according to accepted standards [20,21]. A large, prospective cohort with complete data was used for prediction rule development. Predictors were well defined and outcome assessment was performed in a blinded fashion according to widely used published standards [14]. External validation was performed using a cohort of consecutively inborn singletons with a high rate of available placental pathology. The temporal and geographical spacing between the derivation and validation cohort increase generalisability of the results. Conversely, external validation is limited somewhat by the retrospective nature of the validation cohort. Furthermore the higher rate of vaginal deliveries in newborns without placental pathology indicates a minor selection bias, the effect of which is likely to be small given the low weight carried by mode of delivery in the prediction rules. It is important to note that some diagnostic performance was lost at external validation, a phenomenon observed almost invariably during any process of clinical prediction rule development [22]. During further evaluation of the decision rule it must be kept in mind that it is derived from a cohort of very preterm singleton newborns, and should therefore not be extrapolated to multiplets and more mature newborns without further testing. Application of our model to multiplets in the derivation cohort showed it had no discriminative value in these newborns (not shown), which is in agreement with the distinct mechanisms underlying preterm delivery in this subgroup. Finally, because several diagnostic criteria for histological chorioamnionitis exist [14,[23][24][25][26], the generalisability of our prediction rule for chorioamnionitis diagnosed by criteria other than those by Redline and co-workers [14], requires evaluation.
To date, chorioamnionitis remains a troublesome diagnostic entity. Most cases are clinically silent and there is no reliable way of determining its presence, let alone severity, in a non-invasive manner before delivery [27,28]. Results of placental pathology, the current gold standard for diagnosis, often take days to weeks to obtain. The prediction rules presented here underline the distinct clinical characteristics of affected patients as compared to other preterm newborns. As noted earlier, the primary causes of preterm birth among singletons -preeclampsia and chorioamnionitisrarely co-exist [3,29]. Therefore it is not surprising to see how factors known to characterise either of these groups -preeclampsia and SGA on the one hand, clinical chorioamnionitis, PPROM and vaginal delivery on the other -comprise the major part of the prediction models. The additional contribution of gestational age is explained by its known inverse relationship with chorioamnionitis incidence: lower gestational age increases the likelihood of chorioamnionitis [1,3]. The mutual exclusion of many of the aforementioned variables would potentially favour classification and regression-tree (CART) analyses rather than points-based prediction rules as the most distinctive way to identify chorioamnionitis-exposed newborns. CART analyses were evaluated simultaneously in the current study and were disregarded for their slightly worse test performance (data not shown). Similarly, it may seem appealing to restrict the analyses to subgroups of patients particularly at risk of histological chorioamnionitis, such as those with preterm labour or PPROM. However, PPROM for example was absent in 42% of newborns in the derivation cohort and in 47% in the validation cohort. Consequently, restriction of analyses to PPROM patients would not have allowed identifying over 80% of patients with histological chorioamnionitis in the cohort, and thus using a combination of risk factors in a modelbased approach as presented here is favourable.
Few previous studies have evaluated multivariable models for chorioamnionitis prediction. Using a combination of clinical and laboratory or ultrasound parameters, others were able to predict intra-amniotic inflammation and infection, but with test characteristics inferior to those presented here [30][31][32]. Balaguer and colleagues showed that an algorithm could improve classification of the alleged underlying causes of prematurity in a small cohort [33]. However, it is unclear whether placental histology or other tests were performed to confirm chorioamnionitis. Several biological markers have previously been evaluated as a diagnostic tool to identify the presence and severity of chorioamnionitis shortly after birth. The marker most extensively investigated in this regard is probably interleukin-6 (IL-6), predominantly measured in cord blood. Its sensitivity and specificity for diagnosing HC range  from 73% to 74%, and 77% to 85%, respectively [15,34,35]. Diagnostic properties reported for HCF are within the same range [16,17,34,35]. Several other molecular and haematologic markers have been evaluated, but to date none has shown to carry sufficient predictive ability to justify routine clinical use [17,34,36,37]. Recent studies showed that test accuracy can be enhanced by combining several molecular markers in amniotic or cervical fluid [38,39]. However, external validation was not performed, and obvious drawbacks remain the need for an invasive procedure (for amniotic fluid), as well as costs and time required for proteomic profiling.
To the best of our knowledge, the diagnostic properties of the models presented here are at least equivalent to those of any noninvasive biological marker reported to date, most of which have not been externally validated. Moreover, the prediction rules are easy to use, cheap, and their results are readily available. The models may be used to identify at birth newborns exposed to various degrees of antenatal inflammation. As such, they can facilitate the development of decision rules in future studies to optimise and individualise treatment of preterm newborns. In addition they may be used for predefined subgroup analyses in clinical trials to investigate their value in predicting differential therapeutic responses. Negative predictive value were particularly well preserved during external validation of the prediction rules presented here, making these predominantly useful to guide targeted therapies with large subgroup-associated benefits and little adverse effects at the population level. Whereas in the current study we have preferred high sensitivity to specificity the model obviously allows for tailored selection of a less conservative cut-off value to increase specificity over sensitivity.
Evidence from observational studies suggests that infants exposed to varying degrees of antenatal inflammation may particularly benefit from increased surfactant dosing [8,10] and restrictive use of invasive ventilation [8,[11][12][13]. Furthermore, a randomised controlled trial identified subgroup-associated benefits of postnatal corticosteroids in infants with chorioamnionitis [9]. Results of placental pathology are generally unavailable when decisions regarding the initiation of such interventions are required. Future studies should evaluate the potential value of the use of the clinical prediction rules presented here to guide such decision-making. Another potential field of interest is antibiotic prophylaxis for prevention of early onset sepsis in preterm newborns. Its use is widespread, although serious potential sideeffects are well recognised [40][41][42]. Individualised antibiotic prophylaxis based on prediction of sepsis risk may well improve outcome at the group level by reducing unnecessary and potentially harmful use of antibiotics. The clinical prediction rules presented here could guide such individualisation of therapy.

Conclusion
Development of prediction strategies to identify subgroups of preterm newborns that may benefit from a targeted therapeutic approach is essential to improve future outcomes in neonatology. Presented here is the development and external validation of a multivariable diagnostic prediction rule for histological chorioamnionitis and histological chorioamnionitis with fetal involvement at birth in preterm newborns. The models are composed of small sets of clinical variables readily available in everyday practice and are therefore broadly applicable and simple to use. The models' diagnostic properties are encouraging and appeal for further evaluation to assess their value in supporting clinical decisionmaking.