493. A NOVEL GRAPH THEORY-DRIVEN APPLICATION OF 18F-FDG-PET WITH ANTIPSYCHOTIC DOSE CORRECTION AND A MACHINE LEARNING APPROACH TO CAPTURE DISCRETE BRAIN METABOLIC CONNECTIVITY IN TREATMENT-RESISTANT SCHIZOPHRENIA

Abstract Background Approximately one-third of patients diagnosed with schizophrenia do not respond to at least two antipsychotic medications administered at adequate doses and duration, and are thus considered treatment-resistant1. Dysfunction in functional connectivity has been linked to psychotic symptoms in schizophrenia patients and may shed light on the distinct neurobiological differences between responders (nTRS) and treatment-resistant schizophrenia (TRS) patients2. However, most existing research focuses on the hemodynamic signal from fMRI, with no available information on brain metabolic networks3. To address this gap, we explore for the first time the metabolic profile of brain connectivity using a novel graph analysis, incorporating antipsychotic-dose correction and a machine learning approach. Aims & Objectives To detect: i) Qualitative differences in group networks using graph analysis; ii) Permutation-based comparisons of edges, centrality measures, and global properties; iii) Associations between metabolic signals of brain regions and resistance to treatment using logistic regression. Method 18F-FDG-PET was performed in 70 subjects, 28 nTRS, 26 TRS, and 16 CTRL. Relative brain glucose metabolism maps were processed in the AAL-Merged atlas template, extracting data from 65 brain regions. Inter-subject connectivity matrices were derived using Gaussian Graphical Models with an arctangent-type penalty. A statistical comparison between brain networks was conducted through permutation testing. Additionally, a logistic model based on nested cross-validation was employed to estimate the associations between the metabolic signals of brain regions and treatment resistance. Results 1. Significant differences in global strength were observed between CTRL and patients’ networks (permuted p-values = 0.008 and 0.001 for CTRL vs. nTRS/TRS, respectively), with reduced connectivity in the schizophrenia groups. 2. Predominant alterations were localized in the frontal lobe, Default Mode Network (DMN), and dorsal dopamine pathway. 3. Nodes with differing centrality measures across groups included the frontal gyri, calcarine fissure, putamen, cingulate cortex, precentral gyrus, rectus, angular gyrus, fusiform gyrus, and lingual gyrus. 4. In TRS patients, reorganization of the DMN was observed, including aberrant connections between frontal regions and cingulate/temporal areas. Impaired connections included the left anterior cingulate with the posterior cingulate and inferior/superior frontal gyri. Additional disrupted edges were identified between the posterior cingulate, temporal regions, and superior frontal gyrus. TRS patients also exhibited connectivity alterations in the supplementary motor area, left postcentral gyrus, and right caudate. 5. The machine learning approach demonstrated an accuracy of 0.72–0.8 in detecting the TRS. Discussion & Conclusions Patients' networks exhibited a global hypoconnectivity pattern with altered modular and hub organization compared to CTRL. The metabolic connectivity analysis uncovered novel insights into potential neurobiological correlates of antipsychotic resistance, highlighting alterations in the anterior cingulate, dorsal striatum, frontotemporal, and cortical-subcortical connectivity.