Application of conformal predictors to functional magnetic resonance imaging research

A blood glucose level is an essential measure used in diagnostic and treatment decisions in diabetes. However, currently there are no such measures used in clinical practice for the diagnosis or prediction of clinical response in psychiatric disorders. Brain imaging studies have revealed significant, and often wide-spread, abnormalities in several psychiatric disorders. An essential step towards a clinical application would require a measure of patterns of brain activation that distinguishes patients from healthy individuals. This entails a shift in the data analysis from focusing on differences between groups of patients with a psychiatric illness and groups of healthy individuals to identifying for a particular individual the likelihood that he or she would have a particular illness. The present project will develop statistical analysis methods which are essential for the investigation of such quantitative measures for psychiatric disorders.

Currently, there are no neurobiological markers for diagnosis or prediction of clinical response for psychiatric disorders used in clinical practice. Neuroimaging studies have revealed significant, and often wide-spread, functional and structural cerebral abnormalities in several psychiatric disorders. An essential step towards a clinical application would require a quantitative measure of patterns of brain activation that distinguishes patients from healthy individuals. This entails a shift in the data analysis from focusing on differences between groups to classifying whether an individual belongs to group A or group B. Machine learning methods provide a key analysis tool in developing such quantitative classification for psychiatric disorders. Our preliminary data show significant potential for the functional neuroanatomy of affective and cognitive features as a neurobiological marker for the diagnosis and prognosis in depression, bipolar disorder and schizophrenia. A major limitation of current analysis though is the lack of confidence intervals in the data output. In this project, we propose a new collaboration with the originators of support vector machine analysis and clinical neuroimaging researchers. A computer science post-doctoral researcher will hop into clinical neuroimaging research to help build and adapt machine learning methods to the analysis of neuroimaging data. The project is an essential step towards the development of neurobiological markers in psychiatric disorders.