Prediction of In-Hospital Mortality After 24 Hours in Very Low Birth Weight Infants

We retrospectively analyzed the data in the KNN registry, a national prospective Web-based registry for very low birth weight (VLBW) infants who were born between 2013 and 2017 in South Korea. Among 10 425 patients, we excluded infants who (1) died within 24 hours after birth (n = 277), (2) were transferred to another hospital during hospitalization or hospitalized for >365 days (n = 542), and (3) were born at ≤22⁺³ weeks’ gestation whose weight percentile based on the 2013 Fenton preterm growth chart^12,13  could not be determined (n = 36). We also excluded infants born at hospitals that were not registered for the KNN and then transferred to KNN-registered hospitals (n = 322).

We used multiple perinatal factors analysis obtained within 1 to 2 hours after birth as risk factors in the analysis. Factors used for the analysis included maternal age, parity, amount of amniotic fluid, maternal and/or paternal education and nationality, artificial fertilization, maternal disease (ie, diabetes mellitus [DM], hypertension), preterm premature rupture of membrane (PPROM), and steroid use. Neonatal factors used for the analysis included birthplace (inborn or outborn), gestational age, sex, Apgar score at 1 and 5 minutes, severe congenital anomaly, the need for resuscitation during delivery through oxygen supply, positive pressure ventilation, intubation or cardiac compression, multiple gestation, birth weight, height at birth, head circumference at birth, body temperature at admission, and initial pH and base excess obtained from blood gas analysis.

Oligohydramnios and polyhydramnios were defined as the index of amniotic fluid <5 and >24 cm, respectively. DM on maternal history was categorized as gestational DM and overt DM on the basis of the obstetric medical record. Hypertension on maternal history was categorized as pregnancy-induced hypertension, including eclampsia and preeclampsia regardless of proteinuria or edema, and chronic hypertension, including superimposed hypertension. Artificial fertilization only included in vitro fertilization with or without embryo transfer and did not include intrauterine insemination or ovulation induction. Severe congenital anomaly included life-threatening anomalies such as tracheal obstruction. The percentile of birth weight was determined by a Web-based calculator on the basis of the 2013 Fenton growth curve.^12,13  Small for gestational age was defined as birth weight less than the 3rd percentile or less than the 10th percentile. The serial process of initial resuscitation was performed on the basis of the neonatal resuscitation program of the International Liaison Committee on Resuscitation guideline.¹⁴  The initial pH and base excess were obtained from venous, arterial, or capillary blood within 1 hour after birth. Body temperature measured within 1 hour after admission to the NICU was used.

We evaluated the statistical properties (eg, potential outliers, normality, and missing data) of the baseline and follow-up measurements using summary statistics and graphical tools. Categorical variables are presented as counts and percentages, and continuous variables are presented as means and SDs or medians and interquartile ranges depending on their distribution. The missing data mechanism was speculated to be missing completely at random or at least missing at random; as such, we used the multiple imputation method.

To build the model for predicting the in-hospital mortality, we investigated each potential covariate using graphical tools and the univariate logistic regression model. For continuous predictors, we checked the linearity of the relationship between each predictor and the log odds of the outcomes and considered appropriate transformation if the relationship was not linear. The covariates that were either significant in the univariate model or considered clinically important regardless of statistical significance were entered into the multivariable logistic regression model. The final model was selected by backward elimination after removing variables with collinearity. This model building process was performed on the basis of careful considerations on collinearity, parsimony, goodness of fit, and clinical interpretation to develop an accurate, precise, and useful prediction model.

For internal validation of the model, we performed the receiver operating characteristics (ROCs) analysis and calculated the area under the curve (AUC). A total of 500 bootstrap samples were generated to estimate the optimism of the AUC, and the optimism-corrected AUC was calculated as a measure of the performance of our prediction model.

The KNN registry was approved by the institutional review board of each participating hospital. Standard informed consent was obtained from the parents of the infants before registration in the KNN.

A total of 9248 VLBW infants were included in this study, among whom 1105 (11.9%) died during hospitalization. The mean gestational age at birth was 29.0 ± 2.9 weeks and the mean birth weight was 1095.9 ± 280.1 g. Compared with mothers of the survived infants (survival group), mothers of the dead infants (death group) had lower proportions of DM and hypertension and higher proportions of PPROM and amount of amniotic fluid (polyhydramnios or oligohydramnios) (Table 1).

Compared with the infants in the survival group, those in the death group had a higher proportion of congenital anomaly, lower gestational age at birth, lower birth weight, and lower Apgar score at 1 and 5 minutes; other differences between the groups are described in Table 2. The proportion of infants who were small for gestational age did not significantly differ between the 2 groups when using the 3rd percentile cutoff or the 10th percentile cutoff.

In univariate logistic regression, the neonatal factors that were significantly associated with in-hospital mortality are described in Table 3. In the logistic regression model, the relationships between the log odds of the in-hospital mortality and gestational age at birth, body temperature at admission, and pH and base excess in blood gas analysis within 1 hour after birth reached plateaus after the following threshold values: 196 days, 38.0°C, 7.35, and 0, respectively. These variables revealed linear relationships with mortality at values lower than the aforementioned thresholds; therefore, we used the truncated values at the thresholds for these variables. We eliminated the variables that have no significant association with in-hospital mortality or unreliable association with death.

The variables selected for predicting in-hospital mortality after multivariable analysis and their odds ratios are shown in Table 4, and the final equation derived for prediction of in-hospital mortality (the KNN prediction formula) is as follows:

where xb = 8.994 + 0.964 × (polyhydramnios)^* + 0.185 × (oligohydramnios)* – 0.111 × Apgar score at 1 minute^† – 0.047 gestational age (day)^‡ – 0.003 × birth weight (g)^‡ + 0.679 × intubation^§ – 0.058 × base excess.^‖

The ROC analysis revealed that the apparent AUC for the prediction of in-hospital mortality was 0.870. The optimism-corrected AUC of our prediction model based on 500 bootstrapping samples was 0.867 (Fig 1). The calibration plot, which delineated the agreement between the observed and predicted risks of in-hospital mortality, revealed that the prediction model had a good calibration (Fig 2).

For external validation of the prediction model, we chose the study groups for infants who were born at a different time period (January 2018–December 2018) at 2 tertiary centers of KNN that have different treatment policies. After applying the same exclusion criteria used for the initial study population, a total of 143 infants were included for external validation. As a result, the KNN prediction formula produced an AUC of the ROC curve of 0.876 (Fig 3). The calibration plot of the external validation group also revealed that the prediction model had a good calibration (Fig 4).

In this study, we used a nationwide, prospective, Web-based registry to gather data on VLBW infants and their variables obtained within 1 to 2 hours after birth to build a formula for predicting the in-hospital mortality rates of such high-risk infants. After multivariable analysis, neonatal factors such as polyhydramnios, Apgar score, gestational age, birth weight, and the need for intubation were included in the formula. Notably, our KNN prediction formula had high AUC values of 0.867 for the prediction of in-hospital mortality in the study population and 0.876 in an independent external validation population.

The currently available scoring systems for predicting early death in neonates include the clinical risk index for babies (CRIB) II score, Score for Neonatal Acute Physiology Perinatal Extension II (SNAPPE-II) score, and Eunice Kennedy Shriver National Institute of Child Health and Human Development (NICHD) calculator for neonatal condition or outcomes.^5,7–10  Out of these, the CRIB II score⁹  and the SNAPPE-II score¹⁰  use variables that are collected at 12 hours after birth to predict mortality. In multiple studies, researchers have tried using perinatal factors and early physiologic factors that are obtained soon after birth (at least in the first 24 hours) to predict the probability of death or severe impairments.^5,8–10,15  In this respect, we believe that our KNN formula would be also clinically useful because it uses neonatal measurements that could be obtained within 1 to 2 hours after birth.

Similar to the KNN prediction formula, the NICHD calculator⁵  and the CRIB II score⁸  use variables that are available soon after birth. Accordingly, birth weight and gestational age were identified as independent predictors of mortality in all 3 scoring systems. However, whereas the NICHD calculator included the antenatal steroid as a predictor, the KNN prediction formula did not. Boghossian et al¹⁶  reported that the effect of antenatal steroid on mortality was significant only in infants born between 24 and 25 weeks’ gestation and infants born between 22 and 25 weeks’ gestation from the NICHD calculator.⁵  In our study cohort, the mean gestational age of the infants as a whole was 29.04 weeks, which may explain the exclusion of antenatal steroid as a significant factor. Similar to our study, the study on the CRIB II score included infants <32 weeks’ gestation and identified birth weight, gestational age, and base excess soon after birth as significant predictors of mortality.⁸  Singleton was also a predictive factor in the NICHD equation but not in the KNN prediction formula. Multiple gestations was associated with a greater increased risk of mortality in more immature infants born at 26 weeks’ gestation or earlier^17,18 ; however, multiple gestations were not significantly associated with mortality in the study from the Vermont Oxford Network on the basis of infants of birth weights ranging from 500 to 1500 g.¹⁹ 

Abnormal amount of amniotic fluid is observed in ∼1% to 7% of pregnancies and is associated with congenital anomalies or maternal disease and increased risk of adverse perinatal outcomes.^20–24  Besides congenital anomalies, oligohydramnios is also associated with preterm birth, intrauterine growth restriction, and PPROM^20,23 ; moreover, longer duration of oligohydramnios or anhydramnios accompanied with pulmonary hypoplasia is significantly associated with mortality.²⁵  Polyhydramnios and oligohydramnios are associated with low Apgar scores at 1 and 5 minutes and the need for extensive resuscitation.^23,26,27  The perinatal mortality rate in polyhydramnios is reported to be 30% to 41% even in cases of idiopathic polyhydramnios without congenital malformation and up to 76% in cases with oligohydramnios.^20,23,27,28  In our current study, both polyhydramnios and oligohydramnios revealed significant associations with in-hospital mortality.

In previous reports,^29,30  low Apgar scores were associated with poor neonatal conditions in the delivery room and with early neonatal death. In our study, Apgar scores at 1 and 5 minutes revealed similar relationships for in-hospital mortality, with the Apgar score at 1 minute revealing a more linear relationship to in-hospital mortality than that at 5 minutes, especially in scores of 0 to 3. Despite other authors reporting the association between low Apgar scores at 5 minute and neonatal mortality,^29,31–34  we chose the Apgar score at 1 minute for the formula. Shankaran et al¹⁸  also reported that low Apgar scores at 1 minute were associated with the risk of mortality in infants of birth weights ranging from 501 to 1000 g.

In this study, the risk of in-hospital mortality decreased with higher gestational age until 28 weeks’ gestation. A review article³⁵  summarized the gestational age-specific survival in extremely preterm infants and suggested that the effect of gestational age on mortality was most obvious in infants born between 22 and 25 weeks’ gestation. Lower gestational age was associated with higher mortality in the early period after birth in infants of birth weights from 501 to 1000 g.¹⁸  In the study¹⁹  from the Vermont Oxford Network, a 10 g increase in birth weight was associated with a 3% decrease in odds of mortality. Accordingly, in our study, higher birth weight was associated with lower in-hospital mortality. However, birth weight less than the 3rd percentile or less than the 10th percentile were not predictive factors for death unlike the results from the study on the SNAPPE-II score.¹⁰ 

In the study by Shankaran et al,¹⁸  intubation in the delivery room was associated with a decreased risk of mortality to discharge among infants with birth weights ≤750 g and increased risk of mortality among infants with birth weights >750 g. In our study, intubation was associated with a higher risk of mortality, which may be because our study population included more infants with birth weights ≥750 g.

We used the value of pH and base excess obtained from any arterial or venous or capillary blood samples because of good correlation.^36–40  In previous report,^41,42  lower pH and base excess was associated with adverse outcomes, but pH was not chosen in the final KNN prediction formula. Base excess was associated with mortality after controlling for birth weight, gestational age,⁴³  and acidosis.⁴⁴ 

KNN formula could predict the mortality of VLBW infants within 1 to 2 hours after birth using only 6 factors that reflect a wide range of perinatal conditions from prenatal to postnatal, including 1 maternal factor (amount of amniotic fluid), 2 neonatal biological factors (gestational age and birth weight), and 3 postnatal conditions (Apgar score at 1 minute, intubation at birth, and base excess). Because the KNN database was recently established compared with the currently available scoring systems, the KNN prediction formula more accurately reflects the benefits of recent improvements in neonatal care. Although we built the KNN formula with retrospectively analyzed perinatal data, the data themselves were prospectively collected and revealed a high AUC value for the prediction of in-hospital mortality in both external and internal validation. Therefore, we believe that the KNN formula, which was derived through a prospective study, can be used to reliably predict the mortality in the clinic and be readily applied across a wide range of clinical situations because only 6 variables are needed for its calculation.

Despite its strengths, the KNN prediction formula, similar to the CRIB II score, has limited applications in infants born at outside hospitals. Although, the KNN prediction formula was based on data solely obtained from Koreans, because Tyson et al⁵  showed that race and ethnic groups did not have significant associations with neonatal outcomes, the KNN prediction formula may be investigated for potential usage in predicting the in-hospital mortality of preterm infants of other ethnicities. In consideration of potential future studies on predictive ability of KNN prediction formula and comparison with other scoring systems, we excluded the infants who died within 24 hours after birth, which is similar to the SNAPPE-II score developed for the infants who survived at least 24 hours after birth. Although the exclusion of infants who died within 24 hours may be considered as a limitation of this study, the mortality rate within 24 hours after birth was only 2.9% in the population including the deaths within 24 hours, which was lower than we expected. Moreover, KNN prediction formula revealed a good discrimination (AUC: 0.870; H.W.P., S.Y.P., E.A.K., unpublished observations) even with the infants with 24-hour mortality included.

In this study, we derived a formula for predicting the in-hospital mortality in VLBW infants using perinatal factors available soon after birth. The formula revealed high AUC values in both the internal validation cohort (0.867) and an independent external validation population (0.876). The KNN prediction formula could be a useful tool for counseling parents by providing predictions for the in-hospital mortality rates of their VLBW infants.

AUC

area under the curve

CRIB

clinical risk index for babies

DM

diabetes mellitus

KNN

Korean Neonatal Network

NICHD

Eunice Kennedy Shriver National Institute of Child Health and Human Development

PPROM

preterm premature rupture of membrane

ROC

receiver operating characteristic

SNAPPE-II

Score for Neonatal Acute Physiology Perinatal Extension II

VLBW

very low birth weight