T63. EXPLORING NEW APPROACHES TO PARSE THE CLINICAL HETEROGENEITY IN SCHIZOPHRENIA BY STRUCTURE EQUATION MODELING

Abstract Background The high heterogeneity in schizophrenia limits advances in treatment and biological investigation. Strategies to phenotypic refinement in schizophrenia involve dimensions, staging, and subtyping. Structural equation modeling (SEM) allows applications to such problems, but requires a model with good model fit, currently a major challenge in studies with the most used instruments to assess schizophrenia symptoms dimensions, i.e., the Positive and Negative Syndrome Scale (PANSS). Since consistent MRI studies in schizophrenia have shown thinner frontal and temporal cortices, SEM models associated with neuroimaging data can be considered as a possible neuroimaging biomarker. Thus, we aimed to explore clinical heterogeneity in schizophrenia by performing SEM approaches with the PANSS items to: 1) identify the best dimensional model including confirmatory factor analysis (CFA), bifactor CFA and Bayesian CFA; 2) evaluate the impact of clinical staging on PANSS dimensions, using Multiple Indicators Multiple Causes (MIMIC) Modeling (CFA with covariates); 3) Perform a latent class analysis (LCA) to generate groups of patients with similar PANSS response items; and 4) Validate the LCA model with distal outcomes (cortical thickness) to investigate neurobiological data on classes structure. Methods We analyzed data from 700 patients diagnosed with schizophrenia from four different centers in Brazil. Standard CFA, bifactor CFA and bayesian CFA, (an approach that accepts correlation between variables) were performed and compared according to the specific fit indices. The multilevel structure was included in all models. We tested three multilevel structures independently: centers, raters and staging (first episode schizophrenia, multi-episode and treatment-resistant schizophrenia). In addition, MIMIC modeling evaluated the impact of clinical staging of schizophrenia on the factor means. For the LCA investigation, the LCA best model was chosen based on the comparison of Akaike information criterion (AIC), Bayesian information criterion (BIC) and Log likelihood values. The final model was submitted to a LCA with a distal outcome, where classes predicted cortical thickness. Structural MRI scans in a 1.5T scanner were acquired from 143 participants. Images were processed using Freesurfer software. Results The standard, Bifactor and Bayesian CFA produced poor fit models. The PANSS CFA factorial solution achieved good fit indices when any multilevel structure