Neuroanatomical Heterogeneity of Schizophrenia Quantified via Semi-Supervised Machine Learning Reveals Two Distinct Subtypes: Results From the PHENOM Consortium

Background The neuroanatomical heterogeneity of schizophrenia can lead to confounded interpretations of patient-control comparisons by averaging across possible subgroups exhibiting distinct patterns of gray matter volume (GMV) decreases and increases. Our aim was to decompose the neuroanatomical heterogeneity of schizophrenia using newly developed semi-supervised machine learning tools. Methods T1-structural images of 671 participants from a subcohort of the PHENOM consortium coming from three sites [controls = 364 (age = 29.5±7.0 years; 44.2% female); schizophrenia = 307 (age = 30.9±7.3 years; 35.2% female)] were analyzed. A recently published semi-supervised machine learning algorithm (HYDRA) was applied to simultaneously cluster and classify patients relative to corresponding control subjects. Voxel-wise analyses were carried out using regionally multivariate discriminative statistical mapping (MIDAS). Results We found two neuroanatomical subtypes of schizophrenia. Subtype1 had widespread cortical GMV deficits with larger alterations in medial temporal, medial prefrontal/frontal, and insular cortices (FDR-p < 0.05). Subtype2 had GMV increases in striatum but no cortical abnormalities (FDR-p < 0.05). Subtype1 had relatively higher negative symptoms and lower educational achievement when compared to Subtype2. GMV correlated negatively with illness duration in Subtype1 but not in Subtype2. Conclusions Our results demonstrate distinctly different neuroanatomical profiles of two schizophrenia subgroups. Subtype1 appeared to be more progressive. Strikingly, Subtype2 did not demonstrate cortical GMV decreases and instead showed subcortical increases, which questions current conceptions of the neuropathology of schizophrenia. Given their notable brain differences, these imaging-defined subtypes may require different clinical treatment planning and intervention approaches.