Subtyping schizophrenia based on symptomatology and cognition using a data driven approach

Schizophrenia is a highly heterogeneous disorder, not only in its phenomenology but in its clinical course. This limits the usefulness of the diagnosis as a basis for both research and clinical management. Methods of reducing this heterogeneity may inform the diagnostic classification. With this in mind, we performed k-means clustering with symptom and cognitive measures to generate groups in a machine-driven way. We found that our data was best organised in three clusters: high cognitive performance, high positive symptomatology, low positive symptomatology. We hypothesized that these clusters represented biological categories, which we tested by comparing these groups in terms of brain volumetric information. We included all the groups in an ANCOVA analysis with post hoc tests, where brain volume areas were modelled as dependent variables, controlling for age and estimated intracranial volume. We found six brain volumes significantly differed between the clusters: left caudate, left cuneus, left lateral occipital, left inferior temporal, right lateral, and right pars opercularis. The k-means clustering provides a way of subtyping schizophrenia which appears to have a biological basis, though one that requires both replication and confirmation of its clinical significance.