Building connectome-based predictive models in infancy and toddlerhood

• In Event: Using Connectomics and Machine Learning to Examine Brain Development

Abstract

Human neuroimaging studies of functional brain connectivity traditionally described networks common across individuals or different between groups such as patients and controls or adults and adolescents. Recent work, however, has demonstrated that individual differences in functional connectivity patterns reflect individual differences in consequential cognitive abilities including working memory, attention, and executive control. Although this work holds promise for the development of brain-based biomarkers of behavior, functional connectivity models have primarily been applied to predict behavior from data observed during adulthood and child and adolescent development. The reliability and predictive power of functional connectivity observed earlier in life is less well understood. Furthermore, the majority of existing connectome-based models are postdictive rather than predictive in nature, forecasting outcomes that have already been observed. Here, to characterize the degree to which functional connectivity patterns predict phenotypes in infancy and toddlerhood—and thus to lay the groundwork for building models that predict longitudinal change in abilities and behavior across infant and child development—we analyzed functional MRI data collected during natural sleep from the Baby Connectome Project sample. In fully cross-validated analyses, we trained and tested support vector regression models to predict the age of infants and toddlers between 8 and 24 months old from whole-brain patterns of functional connectivity. Models successfully predicted the age of held-out infants with approximately 4 months error. Extensive control analyses confirmed that model predictions were not driven by differences in head size or motion. Examining predictive network anatomy revealed that functional connections within canonical resting-state networks, such as the default mode network, predicted age as well or better than the whole-brain pattern. Thus, functional connectivity patterns observed in infancy reliably predict chronological age. Looking ahead, these findings suggest the possibility that infant and toddler functional connectivity patterns may also predict aspects of cognitive development, such as language and executive control skills, as well as changes in these abilities over time.

Author

• Monica Rosenberg, University of Chicago
  Presenting Author