Prediction of brain maturity in infants using machine-learning algorithms

Abstract

Recent resting-state functional MRI investigations have demonstrated that much of the large-scale functional network architecture supporting motor, sensory and cognitive functions in older pediatric and adult populations is present in term- and prematurely-born infants. Application of new analytical approaches can help translate the improved understanding of early functional connectivity provided through these studies into predictive models of neurodevelopmental outcome. One approach to achieving this goal is multivariate pattern analysis, a machine-learning, pattern classification approach well-suited for high-dimensional neuroimaging data. It has previously been adapted to predict brain maturity in children and adolescents using structural and resting state-functional MRI data. In this study, we evaluated resting state-functional MRI data from 50 preterm-born infants (born at 23-29weeks of gestation and without moderate-severe brain injury) scanned at term equivalent postmenstrual age compared with data from 50 term-born control infants studied within the first week of life. Using 214 regions of interest, binary support vector machines distinguished term from preterm infants with 84% accuracy (p<0.0001). Inter- and intra-hemispheric connections throughout the brain were important for group categorization, indicating that widespread changes in the brain's functional network architecture associated with preterm birth are detectable by term equivalent age. Support vector regression enabled quantitative estimation of birth gestational age in single subjects using only term equivalent resting state-functional MRI data, indicating that the present approach is sensitive to the degree of disruption of brain development associated with preterm birth (using gestational age as a surrogate for the extent of disruption). This suggests that support vector regression may provide a means for predicting neurodevelopmental outcome in individual infants.

Keywords: Developmental neuroimaging; Functional MRI; Infant; Multivariate pattern analysis; Prematurity.

Fig. 1

Cortical, subcortical and cerebellar regions of interest (ROIs) used in the present analyses. Two hundred and fourteen gray matter ROIs assigned to 13 resting state networks in adults (Power et al., 2011) were selected based upon anatomic location from an ROI set derived from task data and cortical functional areal parcellations in adults. Anterior (A), dorsal (B), right (C) and left (D) lateral views presented. ROIs are overlaid on a neonate-specific atlas image.

Fig. 2

Regions of interest (ROIs) important for differentiating term-born and preterm-born subjects using binary SVMs. Node colors represent assignment to resting state networks in adults (Power et al., 2011). Black = cingulo-opercular, red = default mode, yellow = frontoparietal, dark blue = cerebellum, green = visual, cyan = somatomotor, yellow-green = lateral somatomotor, dark cyan = dorsal attention, purple = subcortical, gray = ventral attention, brown = unnamed 1, and white = unnamed 2. Note distribution of ROIs throughout the brain and across multiple networks. Anterior (A), dorsal (B), right (C) and left (D) lateral views presented. ROIs are overlaid on a neonate-specific atlas image.

Fig. 3

Consensus features important for differentiating term-born and preterm-born subjects using binary SVMs. One hundred and twenty-six consensus features (100% overlap across all cross-validation folds) distinguished term-born and preterm-born subjects with 84% accuracy. Features are scaled by their weights, which denote their relative importance in group differentiation. Green lines denote functional connections stronger in term-born subjects, whereas orange lines denote functional connections contributing to preterm classification. Note distribution of features throughout the brain. Anterior (A), dorsal (B), right (C) and left (D) lateral views presented. Features are overlaid on a neonate-specific atlas image.

Fig. 4

Support vector regression results depicting actual (x-axis) versus predicted (y-axis) gestational age (GA) for term-born and preterm-born infants determined using individual term equivalent rs-fMRI data sets (preterm-born infants = circles, term-born infants = squares). Note the delineation in predicted gestational age values between infants within each group, reflecting differences in functional connectivity measures between term and preterm subjects scanned at comparable postmenstrual age.