The use of maching learning to diagnose first episode psychosis using T1 MRI scans
Poster B49, Friday, October 21, 11:30 am - 1:00 pm, Le Baron
Anthony Harris1,2, Richard Morris3, Fabio Ramos1; 1University of Sydney, 2The Westmead Institute of Medical Research, 3University of New South Wales
Neuroimaging is part of the assessment of many psychiatric disorders and is recommended for young people experiencing their first episode of psychosis. However in almost all cases this expensive investigation is used in a negative sense, to rule out gross pathological processes like neoplasia or cerebrovascular disease. This squanders the wealth of data that the MRI contains, information that could have significant positive diagnostic worth. Machine learning provides a mathematically principled procedure to process MRI scans and predicts the likelihood of a particular disorder directly from the images. It also estimates the uncertainty of the diagnosis. We have investigated the ability of machine learning, using a linear Support Vector Machine in routinely collected T1 MRI images, to correctly identify people presenting with first episode psychosis, schizophrenia, major depression, recovering from severely traumatic stress and healthy controls. High rates of sensitivity and precision in diagnosis were demonstrated particularly for individuals with schizophrenia and major depression. These results were able to take into account the multimorbidity that is commonly observed in psychiatric populations. These results have significant implications for the use of neuroimaging in psychiatric diagnosis.
Topic Area: Neuroimaging