We have a new manuscript under peer review, available now on the OSF preprint archive! The study uses machine learning to sort participants by their clinical status on the spectrum of psychosis according to their MRI data. The MRI data includes a unique "multiplex" network representation of brain connectivity, meaning that the data format considers both the anatomical connections of the brain as well as the functional connectivity.
Machine learning methods are a rapidly growing area of research, and there are many types of machine learning approaches. We have focused on an approach called "graph convolutional networks," which has some similarities to the principles of neuro-biology and the way neurons interact. In this approach, "layers" of simulated activation units, much like neurons, work together to detect "features" in the data that may be useful for clinical classification. In this paper, we show that this approach is superior to simpler machine learning classification approaches such as support vector machines (SVM).
