Quantitative Diffusion and Spectroscopic Neuroimaging Combined with a Novel Early-Developmental Assessment Improves Models for 1-Year Developmental Outcomes

BACKGROUND AND PURPOSE: Preterm infants are at risk for overt and silent CNS injury, with developmental consequences that are dif ﬁ cult to predict. The novel Speci ﬁ c Test of Early Infant Motor Performance, administered in preterm infants at term age, is indicative of later developmental gross motor and cognitive scores at 12 months. Here, we assessed whether functional performance on this early assessment correlates with CNS integrity via MR spectroscopy or diffusional kurtosis imaging and whether these quantitative neuroimaging methods improve predictions for future 12-month developmental scores. MATERIALS AND METHODS: MR spectroscopy and quantitative diffusion MR imaging data were acquired in preterm infants ( n ¼ 16) at term. Testing was performed at term and 3 months using the Speci ﬁ c Test of Early Infant Motor Performance and the Bayley Scales of Infant and Toddler Development, Third Edition, at 12 months. We modeled the relationship of MR spectroscopy and diffusion MR imaging data with both test scores via multiple linear regression. RESULTS: MR spectroscopy NAA ratios at a TE of 270ms in the frontal WM and basal ganglia and kurtosis metrics in major WM tracts correlated strongly with total Speci ﬁ c Test of Early Infant Motor Performance scores. The addition of MR spectroscopy and diffusion separately improved the functional predictions of 12-month outcomes. CONCLUSIONS: Microstructural integrity of the major WM tracts and metabolism in the basal ganglia and frontal WM strongly correlate with early developmental performance, suggesting that the Speci ﬁ c Test of Early Infant Motor Performance re ﬂ ects CNS integrity after preterm birth. This study demonstrates that combining quantitative neuroimaging and early functional movement improves the prediction of 12-month outcomes in premature infants.

P remature birth results in CNS dysmaturation characterized by altered microstructural WM and myelination not quantifiable on head sonography or qualitative MR imaging. [1][2][3] Interventions for developmental delays are typically not started until later in infancy on failure to sit or walk, squandering a period of high neuroplasticity. We developed the Specific Test of Early Infant Motor Performance (STEP) to address this need and have shown that STEP scores at term and 3 months can predict scores on the Bayley Scales of Infant and Toddler Development, Third Edition (Bayley-III) at 12 months. [4][5][6] A recent consensus statement emphasized combining neuroimaging with clinical assessment for improved diagnosis of cerebral palsy. 7 In this study, we determined whether neuroimaging (MR spectroscopy or diffusion MR imaging) reflects STEP performance at term and improves the ability of STEP to predict later development. MR spectroscopy quantifies mobile intracellular metabolomics in the basal ganglia (BG) and frontal WM regions affected by neonatal injury that relate to outcome. [8][9][10][11][12] Both diffusional kurtosis imaging (DKI) 13 and DTI 14 are diffusion MR imaging methods that quantify tissue integrity through the random movement of water molecules. Kurtosis quantifies the water movement deviation away from and not represented by the strict Gaussian distribution that DTI imposes. 13 DKI measures tissue heterogeneity (intra-or extracellular barriers) 15 and is acutely sensitive to pathologic processes. 16 While application of DKI to brain development is new, [17][18][19][20][21] studies in typically developing children show nonlinear relationships of kurtosis with age. [19][20][21] The kurtosis tensor provides higher-order CNS structural information on the organization of the microenvironment, which may directly impact development.
We hypothesized that lower metabolic or microstructural integrity identifies dysmaturation that manifests as functional developmental performance at term or 3 months by the STEP and at 12 months by the Bayley-III. 4,12 This investigation provides evidence that quantitative neuroimaging paired with early developmental assessment improves the prediction of later-stage development.

Sample
We enrolled infants (n ¼ 16) born at 24-34 weeks gestational age with institutional review board approval and parental informed consent, in accordance with the Declaration of Helsinki.

ROI Placement
Bilateral WM ROIs (n ¼ 8) were manually drawn using each subject's FA image for anatomic reference in the anterior limb and posterior limb of the internal capsule (PLIC), external capsule, uncinate fasciculus, inferior fronto-occipital fasciculus (IFOF), A clinical pediatric neuroradiologist verified correct ROI placement. Last, a total WM ROI was created by combining these individual ROIs.

Statistical Analysis
Multiple linear regression (SAS 3.8; SAS Institute) was performed with STEP at term or 3 months as the response variable and metabolite ratios from TE ¼ 30 or 270 ms or DKI-derived metrics as predictor variables. Gestational age at birth and MR imaging were included in all models to account for degree of prematurity.
For the DKI metrics, STEP was modeled in two ways: 1. For a single ROI, each diffusion parameter was averaged separately over all voxels and included as an independent covariate (eg, STEP term ¼ FA sCC 1 KFA sCC ). This approach yields a parsimonious model in which a single WM tract could be sampled with several different diffusion parameters to model prognosis in a clinical setting. 2. To assess whether global brain dysmaturity or injury is more important to functional outcome prediction, we used a single diffusion metric averaged within each ROI as the covariates in the models (eg, STEP term ¼ FA sCC 1 FA PTR 1 FA PLIC ). This allowed us to determine whether an individual diffusivity or kurtosis parameter in major WM tracts could improve model predictions for parsimonious sampling compatible with clinical settings.

Model Selection
We selected models for the best goodness of fit via larger adjusted R-squared (adj-R 2 ) and smaller Akaike Information Criterion (AIC). The difference in the AIC from full-versus-simpler models (D AIC ) was calculated with a P value, P ¼ exp(-D AIC /2). 40 The model that minimized the AIC 41 and maximized the D AIC between full and simple models with P , .05 was selected as the most parsimonious model. We found a strong linear relationship between diffusivity or kurtosis parameters and STEP scores (Online Supplemental Data). The Online Supplemental Data provide details of the b -value estimates for each WM ROI in the model fits. Of note are FA and KFA (Fig 2): KFA showed a notable improvement over FA in predicting concurrent functional performance, though FA models included fewer ROIs (Fig 2 and Online Supplemental Data). The KFA also outperformed all the other diffusion metrics at both STEP time points. Most metric models resulted in models with adj-R 2 . 0.70, indicating very strong correlations.

Demographics
For the ROI-centered approach, we found that many individual WM tracts had moderately strong linear relationships with STEP scores (adj-R 2 $ 0.60, Online Supplemental Data). IFOF modeled the STEP term (adj-R 2 = 0.86, AIC = 8.86), and gCC modeled the STEP 3-month (adj-R 2 = 0.74, AIC = 25.38), similar to the total WM ROI (adj-R 2 = 0.74, AIC = 25.47) across diffusion metrics (Online Supplemental Data). Again, kurtosis metrics contributed more than diffusivity to the models (Online Supplemental Data). Plots showing measured STEP scores at term or 3 months versus modeled STEP scores using FA for diffusivity (left columns) or KFA for kurtosis (right columns) to predict STEP. KFA and FA models show the best fit to measured STEP scores primarily from combining metrics in the external capsule (yellow) and the PTR (purple). Goodness-of-fit metrics of adj-R 2 and AIC are noted with each model. While both FA and KFA predict STEP function during 3 months of development, KFA performs better in the model than FA (higher R 2 and lower AIC); for STEP at term: DAIC ¼ 6.22, P , .05; and for STEP at 3 months: DAIC ¼ 18.31, P , .001. All models are corrected for gestational age at birth and at MR imaging.

Both MR Spectroscopy and DKI Improve STEP Prediction of Bayley-III Scores
Combining MR spectroscopy with STEP scores predicted Bayley-III motor and cognitive scores markedly better than STEP scores alone (Online Supplemental Data). For TE ¼ 270 ms data, improved motor score prediction models were found for term STEP (D AI C ¼ 33.02, P , .001) and for 3-month STEP (D AIC ¼ 22.63, P , .001). In modeling cognitive scores with TE ¼ 270 ms, (NAA/Cho) BG improved predictions beyond STEP alone (term: D AIC ¼ 17.09; 3 months: D AIC ¼ 32.05; both P , .001). MR spectroscopy at TE ¼ 30 ms improved STEP predictive models of cognitive scores (term: D AIC ¼ 15.36; 3 months: D AIC ¼ 27; both P , .001) but included multiple metabolite ratios that detracted from a parsimonious model, unlike the TE ¼ 270 ms models, which included a single ratio. Model b -value estimates for each metabolite ratio included are detailed in the Online Supplemental Data.
We found that FA and radial and mean kurtosis were the best diffusion metrics to combine with STEP scores in predicting Bayley-III scores (Fig 3). For gross motor scores, FA improved the STEP-alone model at term and 3 months (both D AIC . 41, P , .001) as did radial kurtosis (both D AIC . 26, P , .001). For cognitive scores, FA improved the STEP-alone model at term and 3 months (both D AIC . 4, P , .001), as did mean kurtosis (both D AIC . 27, P , .001). Other DKI-derived metrics also predicted Bayley-III scores with STEP time points. (Online Supplemental Data). Included WM ROIs and b -value estimates for each are listed in the Online Supplemental Data.
For the ROI-centered approach, various combinations of kurtosis and diffusivity metrics within individual ROIs improved the predictions of Bayley-III gross motor and cognitive scores compared with the STEP-only models. For parsimony and simplicity, models using either PTR and IFOF with STEP scores had the lowest AICs for predicting Bayley Motor scores (Fig 4) and improved the model fit over the STEP alone (at term: D AIC . 30; P , .001; at 3 months: D AIC . 28; P , .001). The gCC and external capsule with STEP scores improved prediction in Bayley cognitive scores (Online Supplemental Data) over STEP alone (at term: D AIC . 18; P , .001; and at 3 months: D AIC . 14; P , .001). Model fits and b -value estimates for each diffusion metric covariate included in the selected models for each WM ROI are provided in the Online Supplemental Data.

DISCUSSION
During early life after preterm birth, brain injuries as well as alterations in gray and WM maturation occur that influence later development of motor and cognitive skills. Influencing neuroplasticity to improve outcomes requires understanding the neuroimaging representations of brain injuries and their relation to functional movements during early infancy. Using MR spectroscopy and DKI parameters at term-age equivalent, we show that metabolomics for healthy neurons and microstructural integrity correspond to concurrent performance on the STEP early developmental assessment and that neuroimaging combined with the STEP improves prediction of future motor and cognitive scores compared with the STEP alone. Better predictive models would allow testing interventions at a very early period to optimally harness neuroplasticity. Our data highlight the potential for prognostication using the STEP and term-age quantitative neuroimaging and serve as a guide for metric selection. Our data provide proof of concept that quantitative MR imaging of the developing brain is vitally important in prognosis.
MR spectroscopy at TE ¼ 270 ms in either the frontal WM or BG strongly related to STEP and NAA, and Cho ratios improved STEP predictive models for both gross motor and cognitive Bayley-III scores (Online Supplemental Data). However, we observed no relation or benefit with NAA or other metabolite ratios obtained from point-resolved spectroscopy sequence spectra at TE = 30 ms. This discrepancy may be due to a more reliable quantification of NAA and Cho at a long TE, in which there are fewer overlapping metabolite peaks, as has previously been reported. 26 As markers of neuronal metabolic health 42 and myelination, NAA and Cho are both complementary and additive in our models. 1,[9][10][11] Lower concentrations of NAA in the deep gray nuclei of the BG reflect metabolic impairment, lower neuronal volume or density, and possibly insufficient dendritic arborization in the thalamus, responsible for sensory-input integration and regulation of voluntary movement. 43 In the WM, lower NAA is associated with decreased axonal integrity in neonatal WM disease 10,44 and worse developmental outcomes in both preterm and term infants. 9,11,12 ROIs (IFOF, PTR, and gCC) were also strongly correlated with STEP performance (Online Supplemental Data), indicating that visuomotor and sensorimotor integration pathways are critical in the early development of motor skills. IFOF and PTR with the STEP also predicted later motor function, while the gCC and the external capsule contributed strongly to STEP prediction for later cognitive function (Fig 4 and Online Supplemental Data). The best WM tracts for prognostication from our study (IFOF, PTR, and gCC) are readily identified, and ROIs are easily placed by clinical neuroradiologists on standard anatomic MR images or ROIs registered with an appropriate WM atlas.
Most quantitative clinical investigations use diffusion tensorderived FA that measures the directional dependence of water diffusion in a Gaussian distribution. FA increases nonlinearly with axon myelination, packing, and maturation that plateaus around 2 years of age. 19,21 However, the diffusion tensor neither accounts for non-Gaussian cellular factors nor fully characterizes microstructure across the spectrum of development. In estimating the kurtosis tensor, we captured the non-Gaussian spin displacement that corresponds to cellular and tissue heterogeneity within the WM microstructure, different from DTI. 13,16,45 Kurtosis metrics show a nonlinear relationship with age 19,21 and increase with intracellular complexity, cell density, and myelination due to further restrictions on water movement. Radial kurtosis, in particular, increases more quickly during development than its diffusivity counterpart because myelin restricts diffusion across the axon. 21 KFA is mathematically analogous to FA, but the kurtosis tensor provides complementary, more detailed directional variation on anisotropy information. 37,38 Thus, unlike FA, KFA measurements are sensitive to higher-order angular variations in WM fiber orientations, 36,38 also found in fiber-crossing regions. 37 In our study, KFA outperformed FA in representing early function via STEP performance (Figs 2 and Online Supplemental Data), suggesting that complex WM cellular barriers or fibercrossings are important in the early ability to move with normal tone. Although not widely used, kurtosis metrics may be better markers of WM tract injury and health in the largely unmyelinated brain of infants than diffusivity metrics such as FA. When predicting future gross motor and cognitive performance (Online Supplemental Data), however, FA provided better results, followed closely by the KFA, radial kurtosis (gross motor), and mean kurtosis (cognitive). Taken together, our data support kurtosis information possibly significantly adding to conventional FA in understanding the complexity of WM development at term age.
As a standard protocol on clinical MR imaging machines, MR spectroscopy can yield high-quality results in neonatal studies due to high water content of brain tissue and low iron deposition. MR spectroscopy can be challenging in a clinical environment in unsedated neonates. Therefore, clinicians and MR imaging technicians should attempt to schedule MR spectroscopy scans after a feeding to encourage sleep and should be cognizant of when to repeat scans (ie, motion artifacts) or modify the imaging setup (eg, better swaddling/head immobilization). Furthermore, on the basis of our findings, long-TE spectral quantification is robust (ie, a more stable baseline and less confounding peaks) in quantifying a large concentration of metabolites (namely, NAA, Cho, and Cr). Conversely, shorter-TE MR spectroscopy will have a better SNR and more information on a small concentration of metabolites compared with long-TE sequences, with a trade-off in difficulties in quantification of overlapping peak shapes.
For more advanced diffusion imaging, DKI is easily performed within 4-6-minutes by adding the b-value 2000s/mm 2 to standard DTI protocols with 30 directions per nonzero b-value. Numer- and cognitive scores (right columns) plotted against the modeled Bayley scores using STEP at term (upper rows) or 3 months (lower rows) 1 diffusivity and kurtosis metrics within individual WM ROIs to predict function. By means of this ROI-centered approach, the PTR and IFOF metrics combined with the STEP score result in the best predictive models, adjusted for gestational age at birth and MR imaging. Compared with STEP-alone predictive models (Online Supplemental Data), diffusivity and kurtosis in individual WM ROIs greatly improve the STEP model fit for both Bayley-III gross motor and cognitive test scores at 12 months. b -value estimates are provided in the Online Supplemental Data.
Limitations of our study include our sample size, which precluded assessing whether brain dysmaturity or specific sites of injury were more important in neuroimaging models and any contributions of sex. Also, 12 months is a typical time for general developmental delays to manifest but is not optimal for diagnosing cerebral palsy. Another limitation is the combination of single-voxel and CSI data for the TE = 270 ms MR spectroscopy data modeling, which had slightly different TR values as well as differing spectral and shim quality. Last, MR spectroscopy voxel placement and WM ROI voxels were selected to reflect "apparently healthy" WM, but it is possible that lesions (Online Supplemental Data) may not have been readily visible during data gathering and analysis.

CONCLUSIONS
Our neuroimaging data provide proof of concept that the STEP reflects preterm brain dysfunction, dysmaturation, and/or injury. These data suggest that either MR spectroscopy or DKI can augment the STEP at term age in the prediction of long-term motor and cognitive development. If validated in a larger cohort, quantitative neuroimaging and STEP assessments may be used as surrogate end points in clinical trials and may facilitate earlier intervention for infants at high risk of developmental delays.
Disclosure forms provided by the authors are available with the full text and PDF of this article at www.ajnr.org.