Objectively-Diagnosed Diffuse White Matter Abnormality at Term is an Independent Predictor of Cognitive and Language Outcomes in Infants Born Very Preterm

Abstract

Objective:

To externally validate the independent value of objectively-diagnosed diffuse white matter abnormality (DWMA; aka, diffuse excessive high signal intensity) volume to predict neurodevelopmental outcomes in very preterm infants (31 weeks gestational age or younger).

Study design:

Prospective, multicenter, regional population-based cohort study in 98 very preterm infants without severe brain injury on magnetic resonance imaging (MRI). We diagnosed DWMA volume objectively on structural MRI at term-equivalent age using our published algorithm. Multivariable linear regression was used to assess the value of DWMA volume to predict cognitive and language scores on the Bayley Scales of Infant & Toddler Development, Third Edition (Bayley-III) at 2 years corrected age.

Results:

Of the infants that returned for follow-up (N=74), the mean (SD) gestational age was 28.2 (2.4) weeks and 42 (56.8%) were boys. In bivariable analyses, DWMA volume was a significant predictor of Bayley-III cognitive and language scores. In multivariable analyses, controlling for known predictors of Bayley-III scores (i.e., socioeconomic status, gestational age, sex, and global brain abnormality score), DWMA volume remained a significant predictor of cognitive (P<.001) and language (P=.04) scores at 2 years. When dichotomized, objectively-diagnosed, severe DWMA was a significant predictor of cognitive and language impairments, whereas visual, qualitative diagnosis of DWMA was a poor predictor.

Conclusions:

In this multicenter, prospective cohort study, we externally validated our previous findings that objectively-diagnosed DWMA is an independent predictor of cognitive and language development in very preterm infants. We also demonstrated again that visually-diagnosed DWMA is not predictive of neurodevelopmental outcomes.

Despite ongoing improvements in perinatal care, the prevalence of neurodevelopmental impairments (NDI) remains very high in very preterm infants (≤31 weeks gestational age), with up to 40% experiencing long-term risk of cognitive, language, motor, and behavioral abnormalities^{1, 2}. Unfortunately, it may take 2–5 years before cognitive and language abilities can be accurately diagnosed and impairments can be identified. Tools for earlier detection are needed so that early intervention therapies can be more precisely targeted to those at highest risk when neuroplasticity is at its peak. Brain magnetic resonance imaging (MRI) screening at term-equivalent age has been increasingly adopted in hopes of addressing this need^{3, 4}. However, accurate prediction of cognitive and behavioral abnormalities with neonatal MRI remains challenging, partly because most preterm infants now exhibit more subtle pathology, including microstructural, metabolic, and maturational abnormalities that are not visible on structural MRI^5–7. Furthermore, neonatal structural MRI readings frequently suffer from insufficient diagnostic reliability^{8, 9}.

Qualitative diagnosis is least reliable for diffuse excessive high signal intensity (DEHSI)^{10, 11}, the most prevalent abnormality observed in 50–80% of very preterm infants on term-equivalent age T2-weighted MRI^12–16. This imaging signal abnormality is believed to be either a delay in white matter maturation or subtle pathology, possibly on the milder end of the spectrum of periventricular leukomalacia^{17, 18}. Based in part on the unreliability of qualitative diagnosis, few studies have identified risk factors for development of DEHSI^19,20. Conversely, quantitative studies have identified metabolite abnormalities, as well as axonopathy and myelinopathy, in regions of DEHSI^{14, 19, 21, 22}. We previously reported development of an objective algorithm to quantitatively diagnose DEHSI^{20, 23, 24}. Using this algorithm, we identified a strong association between DEHSI and standardized cognitive and language scores at age 2 in a small cohort study of extremely preterm infants^{24, 25}. Two studies that queried this region using other objective, quantitative measures also reported an association with cognitive and language development^{26, 27}. In contrast, most studies that relied on a visual, qualitative diagnosis of DEHSI have not found an association with NDI^{5, 14–16, 28–30}. Considering the above, henceforth we will use the term diffuse white matter abnormality (DWMA) in place of DEHSI. Our objective in this study was to externally validate our previous findings and confirm the independent value of objectively-diagnosed DWMA for predicting cognitive and language development in very preterm infants without severe injury. We also hypothesized that visual, qualitative diagnosis of DWMA would not be predictive of developmental outcomes.

Methods

All very preterm infants born at ≤31 weeks gestational age and consecutively cared for in one of four level III neonatal intensive care units (NICUs) at Nationwide Children’s Hospital (NCH), Ohio State University Medical Center, Riverside Hospital, or Mount Carmel St. Ann’s Hospital in Columbus, Ohio were eligible for inclusion from 11/2014–3/2016 (consecutive sample of 110 infants prospectively enrolled). These NICUs cared for approximately 80% of very preterm infants in the Columbus, Ohio region. We excluded infants with congenital or chromosomal anomalies that affected the central nervous system (i.e., expected poor outcome; N=10) and infants that remained hospitalized at 44 weeks postmenstrual age (PMA) (N=7) unless cared for at NCH, the sole site of imaging. Data collection occurred between 1/2015 and 7/2018. We excluded 12 infants (1: excessive motion artifacts; 11: severe brain injury because automated quantification of DWMA in the centrum semiovale was not possible). The Institutional Review Board of NCH approved the study. We obtained written informed consent from a parent or guardian of every eligible infant.

Magnetic Resonance Imaging Acquisition and Post-Processing

All infants underwent a structural MRI on a 3T Siemens Skyra MRI scanner and 32-channel phased array head coil at NCH. Infants cared for at NCH typically were imaged while they were still hospitalized in the NICU and infants from the other three sites were imaged at NCH after NICU discharge. All scans were completed between 39 and 44 weeks PMA. Inpatients were transported to MRI by a neonatal nurse and neonatologist. All imaging was performed during natural sleep and without sedation by feeding infants immediately prior to the scan, providing hearing protection and using an immobilization device. We assessed the following sequences: axial T2-weighted: echo time 147, repetition time 9500 ms, flip angle 150°, resolution 0.93 × 0.93 × 1.0 mm ³, scan time 4:09 min.; 3D MPRAGE³¹: echo time 2.9, repetition time 2270 ms, echo spacing time 8.5 ms, flip angle 13°, resolution 1.0 × 1.0 × 1.0 mm³, time 3:32 min.; axial SWI: echo time 20, repetition time 27 ms, flip angle 15°, resolution 0.7 × 0.7 × 1.6 mm³, time 3:11 min.

We quantified DWMA using our published objective algorithm (Figure)²⁴. To summarize, brain tissue segmentation was achieved by unified segmentation that included intensity inhomogeneity correction³² with spatial priors obtained from a neonatal probabilistic atlas³³. This produced voxel labels and volumes for the three main tissue classes (white matter, gray matter, and cerebrospinal fluid) and total brain volume. Voxels with signal intensity values greater than 2.0 standard deviation above the mean of cerebral tissues (white and gray matter) were considered to be DWMA. This cut-off yielded very few cases with DWMA volume above zero. Therefore, we used a slightly lower cut-off of 1.8 standard deviation. Partial volume artifacts were controlled by labeling pixels with high probability (≥95%) of gray and white matter membership to be cerebral tissue. Last, we manually removed erroneous, randomly isolated voxels, to enhance detection accuracy (Figure). Volume of DWMA was calculated as the product of voxel volume and total number of voxels in the detected DWMA region. We only examined DWMA volume from the centrum semiovale because this region is not confounded by the normal occurrence of high signal intensity as seen in periventricular developmental crossroads regions that can adversely impact diagnostic reliability^8–11 and our prior study showed DWMA from the centrum semiovale predicted outcomes better than from other white matter regions²⁴. More details are available in our prior publications^{24, 25}. We determined the normalized volume of DWMA by dividing DWMA volume by total cerebral white matter volume.

Figure 1.

Illustration of DWMA identified by algorithm in the centrum semiovale in two very preterm infants with no structural injury on MRI but Bayley-III cognitive scores <80. Left panels show raw axial T2-weighted MRI images through the centrum semiovale and right panels show the same images after objective DWMA segmentation. Human visual assessment of DWMA boundaries likely disagrees with the algorithm’s selection of DWMA regions.

Three pediatric neuroradiologists performed all structural MRI readings using a standardized published scoring system qualitative graded for severity of structural brain injury/maturation and the objective quantitative biometric measurements were all made separately by a single trained expert, as described by Kidokoro et al.³⁴ This approached yielded a global brain abnormality score. Visual, qualitative assessment of DWMA was made by a single reader with >10 years of experience in interpreting neonatal MRI scans and masked to all clinical and quantitative DWMA information. DWMA scoring was based on severity and extent as described by Kidokoro et al¹⁴. Briefly, infants with no DWMA or high signal intensity only in the periventricular crossroads were graded as 0; grade 1 if DMWA only visible in one region; grade 2 if DWMA was visible in two regions; and grade 3 if 3 or more regions were involved in addition to the normal signal intensity observed in the crossroads. The reader also assessed whether the margins of the posterior crossroads were visible (invisible posterior crossroads). Intra-rater agreement on the DWMA grade was assessed by using kappa (κ) statistic following reevaluation of 20 randomly selected MRI scans by the same rater three weeks later. Of the 20 subjects, complete agreement was seen in 60.0% of cases (expected agreement 31.0%) for a κ of 0.42. This represents a moderate strength of agreement³⁵.

Neurodevelopmental Assessment

Participating infants were evaluated at 2 years of corrected age in the NCH High-Risk Follow-up Clinic. We assessed standardized cognitive and language development using the Bayley Scales of Infant and Toddler Development, Third Edition (Bayley-III)³⁶. The composite scores for the two Bayley-III subscales (i.e., cognitive and language) are scaled to metric with a mean of 100, SD of 15, and range from 40 to 160. We assigned a score that was 3 SD below the normative mean to children that could not complete the test because it was too difficult (required for two study infants). All assessments were performed by assessors who were unaware of the DWMA diagnosis and were certified by the National Institute of Child Health and Human Development Neonatal Research Network.

Statistical Analyses

We examined the relationships between the normalized volume of DWMA and Bayley-III cognitive and language scores at 2 years corrected age by linear regression fitted using maximum likelihood estimation that allows for correlations between twins in the study to be explicitly modeled³⁷. To evaluate the independent prognostic value of normalized volume of DWMA, we performed multivariable regression models by adding known perinatal predictors of Bayley-III scores, including socioeconomic status (maternal education, income, and insurance status), gestational age, sex and global brain abnormality score. We also tested the effects of adding center and PMA at MRI to both final models to assess if controlling for the effects of differences in clinical characteristics between sites and age at MRI scan had any effects on the final models. Neither of these variables had a measurable effect on the coefficients in either model and they were removed from the final models. We also examined the value of non-normalized DWMA volume in predicting Bayley-III scores.

In secondary analyses, we used the Fisher exact test to examine prognostic test properties and the association between objectively-diagnosed severe DWMA and visually-diagnosed severe DWMA (grade 3 – pre-specified) and moderate injury on structural MRI with cognitive impairments (Bayley-III cognitive score <80) and language impairments (Bayley-III language score <80). Objectively-diagnosed severe DWMA was defined as the normalized volume of DWMA >75^th percentile (pre-specified²⁵). In post-hoc exploratory analyses, we also tested a higher threshold of >90^th percentile. We used the traditional two-sided P value of <0.05 to indicate statistical significance. All analyses were performed using STATA 15.1 (Stata Corp., College Station, TX).

Results

Of the 98 infants with high quality images suitable for DWMA quantification, 74 (76%) returned for neurodevelopmental testing. Infants who returned for follow-up were not significantly different from those that did not in their baseline characteristics (Table I). Additionally, the global brain abnormality scores (P = .77) and DWMA volumes (P=0.29) were not significantly different between infants with and without follow-up. The Bayley-III mean (SD) cognitive and language scores at 24.2 (1.5) months corrected age were 99.1 (13.8) and 96.4 (15.4), respectively. Structural MRI was performed at a mean (SD) PMA of 40.3 (0.5) weeks. Based on the global brain abnormality score, 53 infants were classified as having no injury (71.6%), 16 with mild injury (21.6%), and 5 with moderate injury (6.8%) on their structural MRI at term-equivalent age. Punctate white matter lesions were observed in 14 infants (14.3%) and cerebellar hemorrhage in 5 infants (5.1%). Neither abnormality on its own or in combination was predictive of Bayley-III cognitive or language scores at 2 years corrected age (P>0.2). As stated above, all infants with severe injury were excluded from the study. Of the 74 infants with follow-up, visually/qualitatively-defined DWMA was diagnosed as grade 3 in 10 infants (13.5%), grade 2 in 21 infants (28.4%), and grade 0 or 1 in 43 infants (58.1%). Only one infant was diagnosed with invisible posterior crossroads.

Table 1.

Baseline characteristics of very preterm infants with neurodevelopmental follow-up at 2 years corrected age and those without follow-up.

Clinical Variables	Infants with Follow-up (N=74)	Infants without Follow-up (N=24)	P Value
Male, n (%)	42 (56.8%)	11 (45.8%)	0.48
Lower socioeconomic status, n (%)	34 (45.9%)	6 (25.0%)	0.10
Multiple births, n (%)	23 (31.1%)	10 (41.7%)	0.46
Chorioamnionitis (histologic or clinical), n (%)	12 (16.2%)	5 (20.8%)	0.76
Antenatal steroids (complete course within 7 days), n (%)	37 (50.0%)	9 (37.5%)	0.35
Gestational age at birth (weeks), mean (SD)	28.2 (2.4)	28.5 (3.0)	0.73
Birth weight (grams), mean (SD)	1133 (395)	1129 (409)	0.96
Transitional hypotension, n (%)	6 (8.1%)	2 (8.3%)	1.00
Sepsis (culture positive), n (%)	9 (12.2%)	6 (25.0%)	0.19
Postnatal steroids for bronchopulmonary dysplasia, n (%)	6 (8.1%)	0	0.33
Bronchopulmonary dysplasia (oxygen supplementation at 36 weeks postmenstrual age), n (%)	37 (50%)	11 (45.8%)	0.82

Normalized volume of DWMA was a significant predictor of Bayley-III cognitive and language scores in bivariable analyses (Table 2). This association remained significant even when raw DWMA volume was tested with tissue or total brain volumes in the regression analyses, suggesting that the normalization did not have a significant effect on the association with Bayley-III subscores. In multivariable analyses, controlling for other known predictors of Bayley-III scores, including socioeconomic status, gestational age, sex, and global brain abnormality score, normalized volume of DWMA remained a significant predictor of cognitive and language development at 2 years corrected age (Table 2). A 10% increase in DWMA volume and low SES were significantly predictive of a 15-point (1.0 SD) and 8-point reduction (0.53 SD) in Bayley-III cognitive scores, respectively. DWMA volume and low SES were also predictive of language scores, with a 10% increase in DWMA volume and low SES predictive of an approximate 8-point and 7-point decrease in language scores, respectively. Non-normalized DWMA volume (i.e., DWMA volume not divided by total white matter volume) exhibited similar coefficients in predicting Bayley-III scores. Visual, qualitative diagnosis of DWMA was not significantly predictive of cognitive (P=0.73) or language scores (P=0.79). Inclusion of known predictors/covariates in the model did not substantially change these relationships.

Table 2.

Regression coefficients for bivariable and multivariable models for objectively-diagnosed, normalized volume of diffuse white matter abnormality (DWMA) on structural MRI at term-equivalent age as a predictor of Bayley-III cognitive and language scores at 2 years corrected age in very preterm infants.

Predictors	Bivariable Relationship Coefficients (95% CI)	P Value	Multivariable Models Coefficients (95% CI)	P Value
Cognitive Scores			Pseudo R2: 54.2%
DWMA volume*	–17.23 (–23.96, –10.49)	<0.001	–15.21 (–22.16, –8.26)	<0.001
Lower socioeconomic status			–8.36 (–13.80, –2.92,)	0.003
Gestational age			–0.01 (–1.29, 1.28)	0.993
Male sex			0.51 (–4.10, 5.12)	0.828
Global brain abnormality score			–0.88 (–2.09, 0.33)	0.155
Language Scores			Pseudo R2: 36.5%
DWMA volume*	–11.82 (–20.16, –3.48)	0.005	–8.04 (–15.88, –0.21)	0.044
Lower socioeconomic status			–6.51 (–12.46, –0.02)	0.049
Gestational age			1.46 (–0.16, 2.77)	0.080
Male sex			–5.75 (–10.71, –0.79)	0.023
Global brain abnormality score			–0.14 (–1.49, 1.20)	0.833

^*Regression coefficient shown for a 10% increase in normalized volume of DWMA.

In secondary analyses, objectively-diagnosed moderate-severe DWMA (75^th percentile threshold; N=19) significantly predicted cognitive impairments (P =0.003) but not language impairments. In post-hoc exploratory analyses, a 90^th percentile threshold resulted in higher accuracy for cognitive and language impairments (Table 3). Objectively-diagnosed severe DWMA (>90^th percentile; N=7) predicted cognitive impairments at age 2 with 100.0% sensitivity and 95.7% specificity (P<0.001). In comparison, moderate injury on structural MRI (N=5) exhibited lower predictive value for cognitive impairments and visually-diagnosed severe DWMA (grade 3; N=10) did not significantly predict cognitive or language impairments (Table 3). Combining visually-diagnosed, moderate-severe grades of DWMA (grades 2 and 3; N=31) did not improve prediction of cognitive (P=0.30) or language impairments (P=0.46).

Table 3.

Prognostic test properties for abnormal structural MRI, visually-diagnosed severe diffuse white matter abnormality (DWMA), and objectively-diagnosed severe DWMA in predicting cognitive and language deficits at 2 years corrected age in very preterm infants.

Predictors	Sensitivity (95% CI)	Specificity (95% CI)	Positive Likelihood Ratio (95% CI)	Negative Likelihood Ratio (95% CI)	P Value
Cognitive Deficits
Structural MRI abnormality	50.0% (6.8, 93.2)	95.7% (88.0, 99.1)	11.7 (2.7, 51.2)	0.52 (0.20, 1.39)	0.021
Visually-diagnosed DWMA (grade 3)	25.0% (0.6, 80.6)	87.1% (77.0, 93.9)	1.9 (0.3, 11.8)	0.86 (0.49, 1.53)	0.448
Objectively-diagnosed DWMA (75th %-ile)	100% (40.0, 100)	78.6% (67.1, 87.5)	4.7 (3.0, 7.3)	0.00 (0.01, 1.46)	0.003
Objectively-diagnosed DWMA (90th %-ile)	100% (40.0, 100)	95.7% (88.0, 99.1)	23.3 (7.7, 70.6)	0.00 (0.01, 1.46)	<0.001
Language Deficits
Structural MRI abnormality	25.0% (3.2, 65.1)	95.5% (87.3, 99.1)	5.5 (1.1, 28.1)	0.79 (0.52, 1.18)	0.087
Visually-diagnosed DWMA (grade 3)	0.0% (0.0, 36.9)	84.8% (73.9, 92.5)	0.0 (0.0, 5.6)	1.18 (0.93, 1.35)	0.588
Objectively-diagnosed DWMA (75th %-ile)	37.5% (8.5, 75.5)	75.8% (63.6, 85.5)	1.6 (0.6, 4.2)	0.82 (0.47, 1.44)	0.42
Objectively-diagnosed DWMA (90th %-ile)	37.5% (8.5, 75.5)	93.9% (85.2, 98.3)	6.2 (1.7, 22.8)	0.67 (0.39, 1.14)	0.024

Discussion

Using an objective, quantitative algorithm to diagnose DWMA, we were able to validate our study hypothesis and prior findings that increasing DWMA volume is predictive of lower cognitive and language scores at 2 years corrected age in very preterm infants without severe injury^{24, 25}. As also hypothesized, we did not observe a significant association between visual, qualitative diagnosis of DWMA and Bayley-III scores. Our findings help explain the conflicting findings from published DEHSI / DWMA prediction studies and should stimulate additional studies to understand the neuropathologic basis of DWMA^{5, 14–17, 24, 28–30, 38}. The findings may also enhance our ability to accurately counsel families early about their infant’s developmental risks early in life and improve our ability to target the highest risk infants for post-discharge aggressive early intervention therapies and novel neuroprotective interventions. For example, based on our results of a 23.3 positive likelihood ratio, a very preterm infant that has objectively-diagnosed severe DWMA, using a baseline prevalence (i.e., pretest probability) for cognitive impairment of 40%, would have a 94% post-test probability of developing cognitive impairment at 2 years of age. Conversely, based on the negative likelihood ratio of 0, infants without severe DWMA would not be at risk for cognitive impairment.

Our current study validates our prior findings of a significant negative correlation between objectively-diagnosed DWMA volume and Bayley-III scores^{24, 25} and extends them with our larger, multisite, regional population-based cohort of very preterm infants compared with our previous single-center cohort of extremely low birth weight infants The generalizability of our studyis increased by our use of a different MRI platform than used in our previous study (Philips) and inclusion of known neonatal predictors of neurodevelopment in our external validation model. Although 2 prior studies that used a visual DWMA diagnosis reported a significant association with developmental outcomes^{13, 39}, a large majority have found no association^{5, 14–16, 28–30, 38}.

Several factors might have contributed to these negative findings. First, the common presence of signal inhomogeneity and normal occurrence of developmental crossroads on term-equivalent age MRI can readily confound a DWMA diagnosis and reduce diagnostic reliability.^8–11 This lower reliability, which we and others have reported, can significantly increase measurement error, thereby reducing the likelihood of finding a significant association even where one truly exists.⁴⁰ In this study, our intrarater reliability was moderate at best,³⁵ similar to the intrarater reliability that we previously reported for DWMA, which was read by a pediatric neuroradiologist (k = .042 vs 0.46, respectively), as well as intrarater reliabilities reported by other investigators.^8–11 Second, even when qualitative DWMA diagnostic reliability is satisfactory, without a gold standard test to confirm DWMA diagnosis,¹⁴ what is being labeled as DWMA might not be true pathology. Finally, a qualitative diagnosis will inherently have lower study power than a quantitative diagnosis (categorical vs continuous variable) to find an association, thus reducing a study’s ability to find significant associations for a given sample size, especially when the sample size is small.^41,42

Several previous studies have corroborated the pathologic nature of DWMA. Compared with healthy term control infants, Wisnowski et al reported several metabolic abnormalities, including reduced n-acetyl aspartate and elevated myoinositol and lactate in the parietal white matter of preterm infants with moderate-severe DWMA, on term magnetic resonance spectroscopy^{21, 43}. Using diffusion MRI, microstructural markers of oligodendrocyte and axonal abnormality have been reported in periventricular regions of DWMA^{14, 19, 22}. We reported similar findings of axonopathy, myelinopathy, and diffuse white matter gliosis on postmortem neuropathology in two extremely preterm infants with DWMA on MRI¹⁷. Additionally, we did not identify any microscopic necrosis, which is more consistent with periventricular leukomalacia (PVL) and punctate white matter lesions^{18, 44}. Instead, regions of DWMA exhibited vacuoles that likely represented axonal loss³⁸. These preliminary findings suggest that DWMA on MRI is likely more akin to the diffuse white matter gliosis without focal necrosis that represented half of the cerebral white matter injuries identified in a large autopsy series, rather than the milder end of the PVL spectrum (microscopic focal necrosis) that represented the second largest group of white matter injuries^{18, 44}. If axonal loss is a predominant feature of DWMA, then the reduction in functional connectivity of important cognitive and attention networks, as we have observed on functional connectivity MRI⁴⁵, may explain the lower cognitive scores in infants with objectively-diagnosed DWMA. The common occurrence of DWMA in the centrum semiovale, a central white matter region that contains several association, projection, and commissural fibers, may also explain its association with cognitive impairments. Any inflammatory (eg, sepsis, bronchopulmonary dysplasia) or hypoxic-ischemic insult to this region is likely to adversely impact myelination and/or axonal development of more structural brain networks than insults in peripheral subcortical white matter regions, and consequently result in more functional consequences.

Additional supportive evidence of the neuropathological nature of DWMA comes from studies that have identified several perinatal/neonatal risk factors for DWMA, including ligation surgery for patent ductus arteriosus, retinopathy of prematurity, necrotizing enterocolitis, and prolonged mechanical ventilation. Typically, these illnesses precede the natural time course of DWMA development, as DWMA has not been reported much earlier than 36 weeks PMA or after 50 weeks PMA. The peak prevalence of DWMA is around term-equivalent age^{29, 46}. The significant association of DWMA with such antecedent risk factors and NDI suggests that DWMA lies intermediate on the pathway between some of these prevalent neonatal illnesses and NDI. Our findings bring us closer to using DWMA as an objective intermediate biomarker to assess the effects of neonatal neuroprotective interventions that are targeted at reducing the burden of DWMA and consequently long-term cognitive impairments. If validated, this will facilitate an earlier mechanistic target (i.e., DWMA), potentially accelerating the conduct and efficiency of phase I and II trials of neuroprotective agents. Additionally, these findings can be applied to very preterm infants for risk stratification at term-equivalent age for post-discharge interventions. Compared with the current approach of evaluating all very preterm infants for delays/impairments, this type of biomarker could accurately stratify risk and permit targeted, aggressive early intervention therapies and the design of new therapies for such higher risk infants and avoiding unnecessary therapies and harms from new treatments for infants identified at low risk. Because our software can automatically quantify and report DWMA volume quickly (within 5 minutes), this software could be readily integrated into clinical MRI platforms to provide immediate results at the point of care, following MRI acquisition.

Our study has several limitations. Our follow-up rate was 76%, which is less than optimal to avoid ascertainment bias^{47, 48}. However, the clinical characteristics of the children with and without follow-up did not differ significantly at baseline (Table 1). Additionally, this is the second prospective cohort study from our group (studying a different population) demonstrating a significant association between DWMA volume and 2-year Bayley-III scores²⁵. For secondary analyses, our study power was limited by conversion of the continuous Bayley-III scores to binary variables, thus resulting in wide confidence intervals for the prognostic test properties. This may also partially explain why the test properties for prediction of language impairments were less accurate. Another likely explanation is that prediction of language impairments in preterm infants has proved more challenging than other outcomes, irrespective of the biomarker studied, perhaps because the infant’s environment affects language more than cognition. Last, because lower socioeconomic status is difficult to capture accurately, it may play an even greater role in cognitive outcomes than we found in our multivariable model. A larger study will be needed before this algorithm can be translated for clinical use. Nevertheless, our results suggest that this algorithm may facilitate early detection of infants at high risk for cognitive impairments.

Abbreviations

Bayley

III- Bayley Scales of Infant and Toddler Development, Third Edition

DWMA

diffuse white matter abnormality

DEHSI

diffuse excessive high signal intensity

MRI

magnetic resonance imaging

NDI

neurodevelopmental impairments

NICU

neonatal intensive care unit

PMA

postmenstrual age