P495. Using Excitation/Inhibition Ratio to Optimize the Classification of Autism Spectrum Disorder and Schizophrenia

Background Significant genotypic and phenotypic overlap between Autism Spectrum Disorder (ASD) and Schizophrenia (SZ) has been widely documented. Overlap in impaired neural activity, including in excitation/inhibition (E/I) ratio, has also been suggested. Nevertheless, direct comparisons are scarce, which can hinder accurate diagnosis. We used machine learning approaches to assess the potential for differential diagnosis of the E/I ratio, with and without phenotypic measures. Methods We collected resting-state fMRI data from 55 healthy controls/HC (24 ± 3.73 y.o.; 29 females), 30 ASD (22 ± 3.74 y.o.; 5 females), and 39 SZ (26 ± 3.58 y.o.; 8 females) plus phenotypic data: Positive and Negative Syndrome Scale (PANSS), Autism Diagnostic Observation Schedule (ADOS), Bermond–Vorst Alexithymia Questionnaire (BVAQ), Empathizing Quotient (EQ), and IQ. Using independent component analysis, we identified 53 components and, for each component’s time-course and for each participant, estimated the Hurst exponent as a proxy for E/I. Using Optimal Classification Trees (OCT), we ran classification analyses on the two clinical groups using five incremental feature models: (1) PANSS and ADOS; (2) PANSS; ADOS, BVAQ, EQ and IQ; (3) Hurst only; (4) Hurst plus PANSS and ADOS; (5) Hurst plus PANSS; ADOS, BVAQ, EQ and IQ. Results E/I was overall decreased in SZ compared to ASD (T > 2.0, p < .05). We observed a consistent increase in classification accuracy over the five models: (1) 50%; (2) 64%; (3) 78%; (4) 78%; (5) 85%. Conclusions Incorporating the E/I ratio alongside phenotypic data results in the best accuracy when disentangling the overlap between ASD and SZ. Supported By This work was supported by the National Institutes of Health (R01 MH095888 and R01 MH119069; M. Assaf) and the National Alliance for Research in Schizophrenia and Affective Disorders (Young Investigator Award 17525; S. Corbera)