by Janaina Mourão-Miranda, Leticia Oliveira, Cecile D. Ladouceur, Andre Marquand, Michael Brammer, Boris Birmaher, David Axelson, Mary L. Phillips
There are no known biological measures that accurately predict future development of psychiatric disorders in individual at-risk adolescents. We investigated whether machine learning and fMRI could help to: 1. differentiate healthy adolescents genetically at-risk for bipolar disorder and other Axis I psychiatric disorders from healthy adolescents at low risk of developing these disorders; 2. identify those healthy genetically at-risk adolescents who were most likely to develop future Axis I disorders. Methods
16 healthy offspring genetically at risk for bipolar disorder and other Axis I disorders by virtue of having a parent with bipolar disorder and 16 healthy, age- and gender-matched low-risk offspring of healthy parents with no history of psychiatric disorders (12–17 year-olds) performed two emotional face gender-labeling tasks (happy/neutral; fearful/neutral) during fMRI. We used Gaussian Process Classifiers (GPC), a machine learning approach that assigns a predictive probability of group membership to an individual person, to differentiate groups and to identify those at-risk adolescents most likely to develop future Axis I disorders. Results
Using GPC, activity to neutral faces presented during the happy experiment accurately and significantly differentiated groups, achieving 75% accuracy (sensitivity?=?75%, specificity?=?75%). Furthermore, predictive probabilities were significantly higher for those at-risk adolescents who subsequently developed an Axis I disorder than for those at-risk adolescents remaining healthy at follow-up. Conclusions
We show that a combination of two promising techniques, machine learning and neuroimaging, not only discriminates healthy low-risk from healthy adolescents genetically at-risk for Axis I disorders, but may ultimately help to predict which at-risk adolescents subsequently develop these disorders.