Towards Precision Psychiatry: Gray Matter Development and Cognition in Adolescence

Precision Psychiatry promises a new era of optimized psychiatric diagnosis and treatment through comprehensive, data-driven patient stratification. Among the core requirements towards that goal are: 1) neurobiology-guided preprocessing and analysis of brain imaging data for noninvasive characterization of brain structure and function, and 2) integration of imaging, genomic, cognitive, and clinical data in accurate and interpretable predictive models for diagnosis, and treatment choice and monitoring. In this thesis, we shall touch on specific aspects that fit under these two broad points. First, we investigate normal gray matter development around adolescence, a critical period for the development of psychopathology. For years, the common narrative in human developmental neuroimaging has been that gray matter declines in adolescence. We demonstrate that different MRI-derived gray matter measures exhibit distinct age and sex effects and should not be considered equivalent, as has often been done in the past, but complementary. We show for the first time that gray matter density increases from childhood to young adulthood, in contrast with gray matter volume and cortical thickness, and that females, who are known to have lower gray matter volume than males, have higher density throughout the brain. A custom preprocessing pipeline and a novel high-resolution gray matter parcellation were created to analyze brain scans of 1189 youths collected as part of the Philadelphia Neurodevelopmental Cohort. This work emphasizes the need for future studies combining quantitative histology and neuroimaging to fully understand the biological basis of MRI contrasts and their derived measures. Second, we use the same gray matter measures to assess how well they can predict cognitive performance. We train mass-univariate and multivariate models to show that gray matter volume and density are complementary in their ability to predict performance. We suggest that parcellation resolution plays a big role in prediction accuracy and that it should be tuned separately for each modality for a fair comparison among modalities and for an optimal prediction when combining all modalities. Lastly, we introduce rtemis, an R package for machine learning and visualization, aimed at making advanced data analytics more accessible. Adoption of accurate and interpretable machine learning methods in basic research and medical practice will help advance biomedical science and make precision medicine a reality