455. AI-Derived Neuroanatomical Subtypes of Schizophrenia and Their Expression During Disease Onset and in Unaffected Siblings

Background
Schizophrenia (SCZ) has a heterogeneous presentation of symptoms, clinical outcomes and response to treatment, which suggest the presence of different subtypes. Heterogeneity likely hinders the understanding of psychosis neurobiology and prevents progress from symptom-based diagnoses to biology-based diagnoses. We used a data-driven, deep-learning analysis of MRI data to examine neuroanatomical subtypes of psychotic disorders. We also examine the extent to which the neuroanatomical subtypes were expressed in first-episode patients (FEP) and in unaffected siblings of patients with SCZ (SIBS).

Methods
Subtypes were learned using a dual deep learning generative adversarial network and clustering model applied to regional gray matter volumes from 446 patients and 711 healthy controls (HC) from 7 sites in the PHENOM consortium. Next, the subtype model was applied to a separate cohort of 597 FEP and 282 SIBS. To characterize the subtypes, we compared gray and white matter maps and clinical variables.

Results
We identified two neuroanatomical subtypes. Subtype 1 had widespread reductions in gray and white matter volume and enlarged ventricles. Subtype 2 had preserved gray matter cortical volumes relative to HC. Subtype 2 was more commonly expressed in FEP and SIBS than subtype 1. In FEP, subtype 2 expression was associated with a lower total PANSS score compared to subtype 1-expressing FEP.

Conclusions
Using a novel deep learning method, we identified two homogeneous subgroups of patients that are undetectable in common case-control group approaches. Our findings support the notion that patients with psychotic disorders, regardless of disease course, are neuroanatomically heterogeneous.