Can MRI brain networks reproducibly distinguish psychosis patients from control subjects?

Thursday, September 13, 2018

12.00 p.m

ISI seminar room 1st floor

Dr. Sarah Morgan University of Cambridge

Algorithms to classify individual psychosis patients from control subjects using MRI scans have proved elusive, in large part because it is often difficult to compare results between studies which use different datasets, machine learning methods and pre-processing techniques. We tackle these challenges by using two independent datasets and three imaging modalities to assess which MRI features show reproducible control/patient differences and how accurately those differences can be used to distinguish the groups. Given the importance of understanding the biological basis for diagnoses, we also examine which features drive the classification accuracies and how reproducible their selection is between datasets. We find that rs-fMRI connectivity matrices give high classification accuracies (83% in the one dataset and 72% in the other dataset), based on a remarkably reproducible regional pattern of control/patient connectivity differences. Some structural measures, e.g. CT, are also predictive, although importantly the underlying pattern of differences is less reproducible between datasets.
Interestingly, the absolute classification accuracies vary between datasets, perhaps due to differing symptom levels or data quality, although which modalities and types of data give higher accuracies is broadly consistent.