Classification of psychosis spectrum disorders using graph convolutional networks with structurally constrained functional connectomes

Abstract

To improve diagnostic strategies for psychiatric disorders, machine learning methods are being increasingly used. Recently, brain network features are being used as biological substrates to predict the diagnostic categories. Previous studies have predicted psychiatric disorders from controls with reasonable precision. Our goal was to classify a broad spectrum of psychotic disorders including those experiencing subsyndromal psychotic-like symptoms. Given the broader phenomenology, we used support vector machines (SVM) and graph convolutional networks (GCN) with a variety of edge selection methods and compared the accuracy of these models. Additionally, the MultiVERSE algorithm was used to generate network embeddings of the functional and structural networks for each subject as inputs for SVM. GCN yielded the highest classification accuracy of 70% compared with traditional and MultiVERSE inputs of SVM (about 50%) on all samples. For the GCN algorithm, the choice of edge selection was important in maximizing model accuracy. Investigation of network connectivity between patients and controls who were classified correctly identified the left and right Auditory 4 Complex as central regions of functional network communication. Our study shows the possibility of using deep learning methods to distinguish persons with subclinical psychosis symptoms and diagnosable disorders from controls which could help improve accuracy and reliability of clinical diagnosis and potentially help initiate early intervention or preventative strategies. Since GCN exploits graph structure and neighborhood relationships to classify and exceeded the accuracy of SVM, our findings suggest that graph structure and connectivity contribute to more accurate prediction of psychosis spectrum disorders from controls.