Statistical Learning of Candidate Network Stratifications in Schizophrenia

Neuroimaging research is limited by 1) lacking consensus on a description system of mental operations, 2) inconsistent results across individual neuroimaging studies, and 3) scarcity of strong hypotheses to test in new experiments. These caveats challenge especially clinical neuroimaging in psychiatry but can probably be alleviated by combining the BrainMap neuroimaging database and pattern-recognition algorithms. An integrated methodological framework will derive hypotheses from multi-site schizophrenia samples (482 participants) and quantify their importance.

The neurobiological knowledge stored in the BrainMap database will be condensed into a set of meta-analytic priors that capture the diversity of human cognition. The meta-analytic priors will enhance statistical properties and interpretability of exploratory analyses based on structural (i.e., voxel-based morphometry) and functional (i.e., task-unrelated resting-state correlations and task-related meta-analytic connectivity modeling) brain properties. Machine-learning methods (e.g., support vector machines, logistic regression, random forests) will automatically identify the most relevant meta-analytic priors for the research questions regarding neurobiological as well as demographic and clinical predictors. Pattern recognition procedures can thus test whether and how biologically meaningful priors relate to schizophrenia pathophysiology. The ensuing candidate priors can readily motivate and improve future hypothesis-driven investigations in schizophrenia.

In sum, frequently diverging and hardly reconcilable research findings in schizophrenia call for new, strong hypotheses. The proposed approach can automatically formalize and predict complex relationships between the clinical exophenotype and neurobiological endophenotype of schizophrenia.