A Biological Signature for the Early Differential Diagnosis of Psychosis: Unveiling the Differences Between Mood Disorders and Schizophrenia With Multimodal Machine Learning Techniques

Schizophrenia (SZ) and mood disorders (BD, MDD) are among the most disabling disorders worldwide, with a relevant social, functional, and economic burden. Although they are identified as distinct disorders, the potential overlapping symptomatology poses important challenges for the differential diagnosis. A consistent literature affirms that brain structure, and function reflect an intermediate phenotype of an underlying genetic vulnerability for the disorders, shaped by interaction with environmental experiences. Such experiences include early life stress and trauma which seem to characterize psychiatric patients and have been associated with brain abnormalities. Further, early life experiences have been associated with inflammation in a subpopulation of psychiatric patients However imaging, inflammatory, and genetic group-level differences, albeit consistent, do not impact clinical practice since they have not been translated into individual prediction. To address these issues, a rapidly growing body of scientific literature implemented computational techniques, such as machine learning (ML). In this project we will develop cutting-edge ML algorithms to predict the differential diagnosis between mood disorders and SZ from genetic, neuroimaging, inflammatory and environmental data in a unique cohort of 1850 patients and 1000 healthy controls recruited in 4 different centers in Italy. The project will address three different aims: in aim 1 we will develop algorithms for the differential diagnosis between SZ and MD combining multimodal neuroimaging and genetic data; in aim 2 we will predict the differential diagnosis between SZ and MD from immuno-inflammatory and environmental data; finally, with aim three we will exploit an animal model to identify the underlying mechanisms of brain alterations associated with exposure to early life stress. Machine learning analyses will include algorithms for data harmonization and feature reduction, as well as for generating normative models. Finally. different classifying models will be compared considering the specific features to achieve the best performance.The definition of reliable and objective biomarkers, combined with cutting-edge computational methodology, could help clinicians in providing more precise diagnoses and early interventions, also considering dimensional constructs & factors influencing outcomes such as affective vs non-affective psychosis and breadth of exposure to traumatic events