205. Machine Learning to Further Improve Classification of Psychotic Disorders Using Clinical and Biological Stratification: Updates From the Bipolar Schizophrenia Network for Intermediate Phenotypes (BSNIP)

Abstract

Background: Classifcation in psychiatry continues to suffer from challenges to validity of the distinctions between its diagnostic categories. The fundamental goal of this study is to delineate different types of psychosis using a stratifed approach involving both biological and clinical information. Methods: Patients with a psychotic disorder (n = 404) were recruited at 5 sites and underwent MRI scans, EEG scans, and comprehensive clinical and neuropsychological assessments. A stratifed machine learning (nonlinear K-means clustering) approach using clinical and biological information was tested against classifcations built using clinical or biological information alone. Results: The optimal number of clusters was determined to be 3. Using silhouette scores to evaluate the separation of clusters using different approaches, we observed that the stratifed approach-clinical + biological information- outperformed the others with a score of 0.523, improving on clinical or biological only, with scores of 0.230 and 0.290, respectively. The stratifed clusters also separated more on regional brain volumes and global assessment of function compared to clinically separated groups, with Group 1 showing the largest defcits-effect sizes ranged from 0.4 to 1.1. Conclusion: The fndings show promise for improved classifcations for psychotic disorders by leveraging existing clinical insights with newfound knowledge from biological psychiatry. The fndings also demonstrate that new methods such as machine learning can be instrumental in helping us deliver on this promise.