Using Machine Learning to Classify Schizophrenia Based on Retinal Images

VII.ObjectivesThinning of retinal layers has been documented in patients with chronic schizophrenia using standard metrics of optical coherence tomography (OCT) devices. We demonstrate the effectiveness of machine learning (ML) techniques to differentiate between schizophrenia patients and healthy controls using OCT images. MethodsFeatures extracted from a convolutional neural network (CNN) designed to segment retinal layers from OCT images represented abstracted data from the OCT images of 14 first episode (FEP) and 18 chronic schizophrenia patients, and their respective 20 and 18 age-matched controls. The abstracted data and OCT machine metrics were used separately to train support vector classification (SVC) models to differentiate between control and schizophrenia samples and test them. ResultsSVCs operating on OCT machine metrics did not classify unseen samples of FEP schizophrenia patients and controls with performance better than chance, while those looking at chronic schizophrenia did, paralleling results obtained using parametric statistics. In contrast, SVCs operating on OCT image data extracted from the CNN classified unseen samples from both populations with performance greater than chance. ConclusionThese results suggest that ML techniques can detect patterns in patients with FEP schizophrenia with greater performance using features extracted from OCT images than metrics provided by OCT machines.