A Predictive Clinical Model for Moderate to Severe Intraventricular Hemorrhage in Very Low Birth Weight Infants

Abstract

Importance:

Intraventricular hemorrhage (IVH) occurs in 15–45 percent of all very low birth weight (VLBW) preterm infants. Despite improvements in the perinatal care, the incidence of IVH remains high. As more preterm infants survive, there will be a larger burden of neurodevelopmental abnormalities borne by former preterm infants.

Objective:

The objective of this study was to develop a predictive clinical model of IVH risk within the first few hours of life in an effort to augment perinatal counseling and guide the timing of future targeted therapies aimed at preventing or slowing the progression of disease.

Design:

This is a prospective observational cohort study of VLBW infants born in the NICU at John’s Hopkins Children’s Center from 2011 to 2019. Presence and severity of IVH was defined on standard head ultrasound screening (HUS) using the modified Papile classification. Clinical variables were identified as significant using absolute risk regression from a general linear model. The model predictors included clinically meaningful variables that were not collinear.

Setting:

This study took place at the Johns Hopkins Children’s Center Level IV NICU.

Participants:

The study sample included VLBW infants treated in the neonatal intensive care unit (NICU) at John’s Hopkins Children’s Center from 2011 to 2019. A total of 683 infants included in the study had no or grade I IVH, and 115 infants had grades II through IV IVH. Exclusion criteria included admission to the JHH NICU after 24 hours of age, BW > 1500 g, and failure to consent.

Main Outcome:

The main outcome of this study was presence of grades II-IV IVH on standard head ultrasound screening using the modified Papile classification.¹

Results:

A total of 798 VLBW infants were studied in this cohort and 14.4 percent had moderate to severe IVH. Fifty four percent of the cohort was black, 33 percent white, and half of the cohort was male. A higher gestational age, 5-minute Apgar score, hematocrit and platelet count were significantly associated with decreased incidence of IVH in a multi-predictor model (ROC 0.826).

Conclusion and Relevance:

In the face of continued lack of treatments for IVH, prevention is still a primary goal to avoid long-term developmental sequela. This model can be used for perinatal counseling and may provide important information during the narrow therapeutic window for targeted prevention therapies.

Introduction:

Intraventricular hemorrhage (IVH) occurs in 15–45 percent of all very low birth weight (VLBW) preterm infants.² Despite a decline in the incidence of IVH since the 1980’s, the absolute number of cases remains high due to advances in the care of preterm infants and the resulting lower mortality.^{3, 4} As more preterm infants survive, there will be a larger burden of neurodevelopmental abnormalities borne by former preterm infants.^{4, 5} Infants with severe periventricular/intraventricular hemorrhage (grades III and IV) have a higher odds of moderate to severe neurodevelopmental impairment (NDI) at 18 to 24 months of age compared to those with mild to no IVH.³ Newer research suggests that grades I and II of IVH are not as benign as previously thought, as these children have higher odds of moderate to severe NDI than their counterparts without IVH.^{3, 6} These neurodevelopmental consequences extend into childhood and adolescence.^{5, 7, 8} There is an increased risk of mild CP for infants with grades II and III IVH compared to infants with grade I to no IVH as well as a greater risk of overall mortality in infants with grades II-IV of IVH.⁹ This new research posits the inclusion of grade II IVH in the classification of moderate to severe or clinically significant IVH. Clearly there is a different prognosis for infants across this spectrum for moderate to severe IVH.¹⁰

Damage to the germinal matrix occurs very early after birth. As many as 50 percent of cases of IVH in VLBW infants occur in the first 6 to 8 hours of life.^{11, 12, 13} The standard of care for diagnosis of IVH is based on post-natal head ultrasound screening 3–7 days after birth.¹⁴ Head ultrasound is specific but lacks sensitivity and only demonstrates the presence of injury. We lack both diagnostic modalities to detect brain injury risk early in the injury continuum and therapeutic means to successfully intervene in the course of the disease.

The aim of this study was to develop a clinical prediction model for IVH that focuses on information known to the physician in the neonate’s first few hours after birth. Republishing of a risk model is especially prudent at this time as new therapies are being used in translational models of neonatal IVH.^{15, 16} If successful prevention therapies can be found, it may afford time for intervention before progression to later complications, such as periventricular hemorrhagic infarction and progressive post-hemorrhagic hydrocephalus, which are major determinants of neonatal morbidity and mortality.¹⁷

Methods:

Patient Population

The study sample was a prospective observational cohort of VLBW infants treated in the neonatal intensive care unit (NICU) at John’s Hopkins Children’s Center from 2011 to 2019. The study sample was drawn from a larger data set of babies at risk for brain injury. Variables previously associated with risk of neonatal brain injury were included in the original data set. Exclusion criteria included admission to the JHH NICU after 24 hours of age, BW > 1500 g, and failure to consent.

Diagnosis of IVH

Presence and severity of IVH was defined on standard head ultrasound screening using the modified Papile classification.¹ Head ultrasounds were read by pediatric radiology attending physicians and data for the study was extracted from the reports. Head ultrasounds were obtained within the first week of life with a mean age of 3–4 days followed by weekly. If IVH is detected, it is followed weekly until determined clinically stable. ca extends into adulthood than their counterparts without IVH.^{3, 5, 6, 7, 8 18 19}

Development of Predictive Model

Clinical variables were selected as candidate predictors based on bivariable association analysis with moderate to severe IVH (rank sum test for continuous variables, chi-squared test for categorical variables, at p < 0.05). The associations amongst all these candidate variables, listed in table 1, was assessed pairwise to identify possible collinearity (Spearmen’s rank correlation coefficient between two continuous variables > 0.8 or < −0.8, lack of overlap of the interquartile range of continuous variable by dichotomous predictor variable, chi-square test for a pair of dichotomous variables). For each set of mutually collinear variables identified, the variable with a stronger correlation to incidence of IVH was retained, thus generating a reduced list of non-collinear variables. The final list of variables was generated using multivariable absolute risk regression in a general linear model framework (link identity, family Gaussian, with robust standard errors to allow for non-normal residuals) selecting those that retained p < 0.10 in the final model. In their paper comparing methods, Pedrosa et al state, “Our results support the use of a binomial or Poisson GEE model with identify link and robust variance estimates. In cases where these models fail to run either a logistic regression, log Poisson regression, or linear regression GEE model with exchangeable correlation and robust standard errors.” ²⁰ Stata 15.1, with no add ins, was used to run the general linear models.

Table 1.

Subject characteristics and outcomes

Clinical Factor		Grade 0/I IVH	Grade II-IV IVH	p-value
n		683	115
Gender
	Male	334 (48.9%)	59 (51.3%)	0.63
	Female	349 (51.1%)	56 (48.7%)
Gestational Age (week)				<0.001
	<28	284 (41.6%)	95 (82.6%)
	28 to <32	336 (47.7%)	20 (17.4%)
	≥ 32	73 (10.7%)	0 (0.0%)
5-minute Apgar score (Median (IQR))		7 (5, 8)	6 (4, 7)	< 0.001
Acidosis on blood gas*		16 (3.2%)	5 (6.0%)	0.2
Birthweight (gram)
	<500	14 (2.0%)	2 (1.7%)	<0.001
	500–749	113 (16.5%)	42 (36.5%)
	750–999	171 (25.0%)	46 (40.0%)
	≥1000	385 (56.4%)	25 (21.7%)
Multiple gestation		218 (31.9%)	31 (27.0%)	0.29
Antenatal steroids		642 (94.0%)	101 (87.8%)	0.016
Type of delivery				0.029
	Vaginal delivery	209 (30.6%)	47 (40.9%)
	Cesarean section	474 (69.4%)	68 (59.1%)
Race				0.12
	White	227 (33.2%)	36 (31.3%)
	Black	362 (53.0%)	72 (62.6%)
	Hispanic	32 (4.7%)	3 (2.6%)
	Asian	38 (5.6%)	1 (0.9%)
	Other	24 (3.5%)	3 (2.6%)
Magnesium		544 (79.6%)	84 (73.7%)	0.15
Respiratory distress syndrome		608 (89.0%)	111 (96.5%)	0.013
Clinical chorioamnionitis		76 (11.1%)	20 (17.4%)	0.056
Histologic funisitis		103 (15.5%)	29 (28.4%)	0.001
Histologic chorioamnionitis		185 (27.9%)	40 (39.2%)	0.019
Blood infection		63 (9.2%)	31 (27.0%)	<0.001
CSF infection		6 (0.9%)	9 (7.8%)	<0.001
Urine infection		49 (7.2%)	19 (16.5%)	<0.001
Hct median [p25, p75]		44.3 [39.1, 49.4]	39.6 [36.3, 43.7]	<0.001
Platelet count > 150,000 median [p25, p75]		67 [37, 116]	55.5 [30.5, 97]	0.071

IQR interquartile range, IVH intraventricular haemorrhage

^*Blood gas with pH < 7 or BE ≥ −12

Based on the final list of non-collinear clinical variables with cutoffs for the variables based on age-appropriate norms, we developed a reference patient, compared to whom each the absolute risk imposed by each predictor is presented. The reference patient for the prediction score had a hematocrit of 45 percent, 5-minute Apgar of 6, Gestational age of 28 weeks and platelet count of 150,000. Prior to fitting the prediction model, we did exploratory analysis to determine if there was any non-linearity or threshold for the association of continuous variables with IVH and added a spline term to the predictive model to incorporate the non-linear relationship.

Comparison to Previous Predictive Models

Given the lack of adequate sized cohorts for a true validation study, we generated another perinatal prediction model for moderate to severe intraventricular hemorrhage applying the predictors chosen previously by Lee and colleagues to our cohort of neonates. ²¹ The reference patient for the Lee cohort is a male infant, who did not receive antenatal steroids, had a 5-minute Apgar score < 7, BE < −12 on initial blood gas and birth weight of < 1000 grams. The model with Lee predictors was given the highest sample size possible, i.e., even if babies were missing measurements for our new model, they were included for this analysis if the Lee et al. model could be applied with the variables in our dataset.²¹ For direct comparisons between models, the model with Lee predictors was also evaluated on a smaller overlapping cohort where all variables needed to obtain both predictions were available.

Statistics:

We performed all analyses using Stata (Version 15.1, College Station, TX) software. All reported P-values were two-sided. A p < 0.10 was considered to be statistically significant for all potential univariate variables and a p < 0.05 was used for the multivariable regression.

Results:

The biodemographic characteristics and outcomes for the study population are presented in Table 1. From the 798 VLBW infants treated in the Johns Hopkins NICU from 2011 to 2019, 115 infants suffered from moderate to severe IVH. Fifty four percent of the cohort was black, while 33 percent was white, and half of the cohort was male. Ninety percent of the cohort was born at < 32 weeks gestation. All infants were less than 1500 g with 27.2 percent between 750 to 999 g and 51.4 percent greater than 1000 g. The median 5-minute Apgar score was 7. Ninety three percent of infants received antenatal steroids and 90 percent developed respiratory distress syndrome (RDS).

Upon univariable analysis, several clinical factors including gestational age, 5-minute Apgar score, birthweight, antenatal steroids, magnesium, RDS, histologic funisitis, histologic chorioamnionitis, blood infection, CSF infection and urine infection were significantly correlated with IVH. When quantified, platelet count displayed a bi-modal distribution. Platelet count > 150,000 was significantly associated with no/grade I IVH. The final variables in the clinical predictive model, gestational age, 5-minute Apgar score, hematocrit and platelet count, were not collinear and had the highest correlation with moderate to severe IVH (Table 2). Table 2 includes 776 infants (663 with grade I to no IVH and 113 infants with grades II-IV IVH), because some predictor variables were missing in 22 patients. For every additional week of gestation beyond 28 weeks, the incidence of IVH decreased by 3.7 percent with corresponding increases for gestation less than 28 weeks. Every increase in Apgar score beyond 6 was associated with 1.1 percent less chance of moderate to severe IVH with corresponding decreases for lower Apgar scores. Incidence of IVH decreased by 0.5 percent for every percentage of hematocrit above 45 with corresponding increases for hematocrit below 45. IVH decreased by 0.04% for every 1,000 platelets above 150,000, but there was no association below this threshold. A model using the predictors for IVH identified by Lee et al was generated for our patient cohort (Tables 3).²¹ In our cohort, only 5-minute Apgar score, antenatal steroids and a cord BE > −12 were statistically significant among the Lee predictors. Comparison of our IVH risk model with the Lee model revealed improved IVH prediction with a ROC area under the curve of 0.826 (95% confidence interval 0.711 to 0.810) vs 0.761 (95% confidence interval 0.771 to 0.861), respectively (P = 0.005) (Figure 1).

Table 2.

Absolute risk model for moderate to severe IVH

n=785
Variables	Coefficient	95% Confidence Interval	p value
Gestational age (week)	−3.696	−4.7--2.7	<0.001
5-minute Apgar score	−1.215	−2.5–0.0	0.057
Hematocrit (%)	−0.485	−0.8--0.1	0.008
Platelet count	−0.037	−0.1–0.0	0.043

Reference patient is 28 weeks gestational age, 5-minute Apgar score of 6, hematocrit of 45%, and platelet count of 150,000

All coefficients are multiplied by 100 to produce absolute percentages

Platelet count is reported per thousand

Table 3.

Absolute risk model using Lee predictors with overlapping cohorts

n = 573
Variables	Coefficient	95% Confidence Interval	p-value
Sex	−4.322	−9.6–1.0	0.111
Multiple gestation	−2.205	−7.9–3.5	0.449
Antenatal steroids	−14.522	−29.8–0.70	0.062
5-minute Apgar score	10.500	4.2–16.8	0.001
Cord BE > −12	19.342	−0.1–38.8	0.051
BW 500–749	16.466	−5.3–38.2	0.138
BW 750–999	5.499	−15.2–26.2	0.603
BW ≥ 1000	−3.807	−23.8–16.1	0.708

Reference patient is a male infant, no antenatal steroids, 5-minute Apgar score <7, BE < - 12 on initial blood gas, and birth weight < 1000 g

All coefficients are multiplied by 100 to produce absolute percentages

BW birthweight (grams)

Figure 1.

ROC curves comparing the Lee model to the new proposed model.^{33, 34} Lee model has an AUC of 0.761 (95% CI: 0.711–0.810). The new model has an AUC of 0.816 (95% CI: 0.771–0.861). P-value is 0.0045.

Discussion:

We developed a clinical prediction model for IVH in VLBW infants that focuses on the antenatal, intrapartum and early postnatal periods. Emphasis on this early period is critical, as most IVH occurs in the first few hours of life and the window for possible prevention is narrow.^{11, 12, 13} The most important findings of the present study were that increased gestational age, 5-minute Apgar score above 6, initial hematocrit above 45% and platelet count above 150,000 are associated with a decreased incidence of moderate to severe IVH in VLBW infants.

IVH remains a major cause of morbidity and mortality in VLBW infants with an incidence of 15 to 45 percent.² IVH begins with injury to the germinal matrix, which houses developing cortical neurons and glial cells. Damage to the germinal matrix can lead to impaired myelination and cortical development as well as neurodevelopmental delay.^{22, 23} Severe IVH is often associated with worse developmental delay.²⁴ The germinal matrix is susceptible to injury due to the abundance of immature friable capillaries as well as a decreased capacity of the premature brain to autoregulate cerebral blood flow.^{22, 23}

We chose to include grade II IVH in our model, a priori, based on new outcome data showing that children with grade II IVH have a higher likelihood developing moderate to severe NDI that extends into adulthood than their counterparts without IVH.^{3, 5, 6, 7, 8} There is also an increased risk of mild CP for infants with grades II and III IVH compared to infants with grade I to no IVH as well as a greater risk of overall mortality in infants with grades II-IV of IVH.⁹ Clearly damage to the germinal matrix has occurred injuring developing cells in this region as well as exposure of the ependymal lining to the toxic products of blood breakdown within the ventricle.¹⁵

Predictive models for IVH exist. A multicenter retrospective cohort study by He and colleagues evaluated factors that influence severe intraventricular hemorrhage (SIVH) in the first 5-days of life.²⁵ In the 516 infant cohort, absence of antenatal steroid therapy, GA < 28 weeks, BW < 1000g, 1-minute Apgar score < 8, mechanical ventilation and hypotension were associated with increased SIVH.²⁵ A retrospective analysis by Poryo and colleagues of 765 inborn infants at 5 level 3 perinatal centers found higher gestational age, antenatal steroids and cesarean section without uterine contraction were associated with less IVH. ²⁶ RDS, pneumothorax and catecholamine use, all postnatal variables, were associated with increased IVH. Only 44 percent of the infants in this study were VLBW compared to 100 percent in our cohort.²² Without gestational age and birth weight in the model, early onset sepsis and patent ductus arteriosus were associated with increased IVH.²²

In our study, antenatal steroids were not included in the multivariate model. In our dataset, antenatal steroids were collinear with multiple clinical variables, including gestational age and platelet count. In order to remove redundancy, the variables with the strongest univariate association with IVH, gestational age and platelet count, were retained, and antenatal steroids was removed from the model.

He and Poryo did not report on potential collinearity in their studies.^{25, 26} In our model, gestational age and birth weight were highly collinear with one another, while RDS, IUGR and pre-eclampsia were collinear with the 5-minute Apgar score. For each set of mutually collinear variables, we chose one predictor for our final model. We thus attempted to represent the correlated predictive information adequately but without redundancy. By focusing on the very early postnatal period, we hoped to reduce problems of time sequencing and produce a more concrete biologic understanding of the causes of IVH.

Singh²⁷ and Luque²⁸ et al created predictive models for IVH to examine the role of prophylactic indomethacin. From a cohort of 2917 infants, infants with SIVH had a lower mean gestational age, birth weight, 5-minute Apgar score, antenatal steroid administration rate, were more frequently outborn and delivered by C-section.²⁷ With multivariate regression, gestational age was the strongest contributor to IVH followed by mechanical ventilation, lack of antenatal steroids, 1-minute Apgar score, birth weight, cesarean section, male gender and RDS.²⁸ These studies, like ours, focus on variables known to the clinician in the first few hours of life. However, neither study explicitly mentions avoiding collinearity of predictors in their models and the study by Singh et al only includes univariable analysis.²⁷

We compared the ability of the present model to predict IVH in VLBW infants to the model of Lee.²¹ This model prioritizes early predictors of IVH with a focus on acidosis on blood gases within the first hour after birth and addresses the problem of collinearity of birth weight and gestational age. In the Lee model, birth weight is the strongest predictor of IVH. Lower 5-minute Apgar score, male gender, multiple gestation, lack of antenatal steroids and presence of a base deficit ≥ 12 is associated with increased IVH. After applying these predictors to our cohort, male gender and multiple gestation were not significant. Importantly, our proposed model better predicted IVH in VLBW infants with an improved area under the ROC.

This study is one of a few to show the importance of hematocrit and platelet count at birth as predictors of IVH. The incidence of IVH decreased by 0.5 percent for every percentage of hematocrit above 45% with corresponding increases for hematocrit below 45. Dekom et al found that initial hematocrit was significantly associated with IVH.²⁹ After adjusting for gestational age, the odds of IVH was higher if the hematocrit was less than 45 percent.²⁹ In a meta-analysis, delayed cord clamping was associated with higher hemoglobin levels and lower rates of IVH.²⁹ A higher hematocrit at delivery may indicate higher intravascular volume, improved cerebral perfusion and hemodynamic stability with decreased risk of hypoperfusion of the germinal matrix.

The role of platelets in neonatal brain injury has not been well studied and has not been previously included in multivariable models of IVH prediction. In this study, IVH decreased by 0.04% for every 1,000 platelets above 150,000/dl. These results show additional information that higher platelet counts may be protective and add to the information to the meta-analysis showing that a platelet count < 100 × 10⁹/L was associated with increased risk of IVH.³⁰ It is biologically plausible that increased platelet counts protect against an increased risk of bleeding. Alternatively, these results may indicate that overall good health of the fetus and lack of stressors known to cause fetal thrombocytopenia, protects against IVH by supporting normal platelet counts at birth.

Single factors, such as acidosis and cesarean delivery, have been proposed to influence IVH; however, these variables are often collinear with other important variables. Collinear variables provide the dataset with large amounts of the same information and inflate standard errors of coefficients in the prediction model.³¹ By removing collinear variables and focusing on proximal clinical predictors with biologic plausibility, we improved the accuracy of our model.²⁸ Additional strengths include the moderately large cohort collected over the last decade and the ability to interpret risk for individual infants. The results are applicable to similar level IV NICU’s.

We feel republishing of a risk model is especially appropriate at this time as new therapies are being used in translational models of neonatal IVH.^{15, 16} A recent phase 2 randomized controlled trial revealed that treatment of infants born at 23–27 weeks’ gestation with a continuous infusion of recombinant human (rh) IGF-1/rhIGFBP-3 resulted in a decreased incidence of severe IVH.¹⁶ Further study using a preterm rabbit pup model showed that treatment with rhIGF-1/rhlGFBP-3 led to an upregulation of choroid plexus genes, which play an important role in vascular maturation and structure.¹⁶ By enhancing vascular maturation in the choroid plexus, IGF-1 may decrease the vulnerability of the preterm brain to fluctuations in cerebral blood flow, which contribute to the development of IVH.¹⁶ Using rats with chorioamnionitis and early post-natal IVH, Robinson et al found that treatment with erythropoietin and melatonin prevented macrocephaly, developmental delay (poor cliff aversion performance) and ventriculomegaly.¹⁵ This study suggests that the inflammatory milieu caused by chorioamnionitis coupled with early brain injury from IVH damages the ependymal motile cilia leading to posthemorrhagic hydrocephalus of prematurity (PHHP).¹⁵ Our model could be used to identify infants with a high risk of IVH, before ultrasound detection, to identify patients who may benefit from treatment, such as erythropoietin and melatonin, to prevent PHHP.

Two, important points should be made about the findings related to hematocrit and platelets. Even though these two factors are acquired postnatally, it is appropriate to include these values, as the data will be available in the first few hours after birth, which is when the risk of post-natal IVH will be assessed and any preventative therapies will be administered, e.g. indomethacin. Importantly, the present study is one of risk prediction for moderate to severe IVH, and the findings about hematocrit and platelet counts that affect our model should in no way be extrapolated to guide red blood cell or platelet transfusion thresholds.³²

A limitation of this study was that we were unable to distinguish between risks of inborn vs outborn patients. Delayed cord clamping became standard of care during this study, thus subsequent analyses will need to take this into account. We did not validate our model with a separate cohort. However, we compared our predictors against the model published by Lee et al,²¹ and demonstrated that our proposed model is a better predictor of IVH in VLBW infants. Future research includes testing this model in a validation cohort and incorporating concurrently obtained biomarkers to determine if serum biomarkers at birth increase the value of this early prediction model for IVH. Finally, we cannot exclude the possibility that a small fraction of infants in this study suffered IVH in the immediate peripartum period; however, none of those diagnosed with IVH had ultrasound evidence that the IVH occurred at a remote time during gestation prior to birth and diagnosis post birth.

Conclusion

We developed a novel and easily applied clinical prediction model for IVH in the very early postnatal period. In the face of continued lack of treatments for IVH, prevention is still a primary goal to avoid the severe long-term consequences of this disease process. Our study offers neonatologists the ability to provide more accurate perinatal counseling to families regarding brain injury risk and may provide important information during the narrow therapeutic window for future targeted prevention therapies.

Key Points.

Question:

What clinical factors, known to the clinician in the first few hours of life, are associated with an increased incidence of grades II-IV intraventricular hemorrhage in very low birth weight infants?

Findings:

In this prospective cohort study of 115 infants with grades II-IVH intraventricular hemorrhage, a higher gestational age, 5-minute Apgar score, hematocrit and platelet count were significantly associated with decreased incidence of IVH in a multi-predictor model.

Meaning:

This model can be used for perinatal counseling and may provide important information during the narrow therapeutic window for future targeted prevention therapies.

Data Availability

Request for data sharing should be submitted to the corresponding author for consideration.