Abstract
Data-driven phenotypes have the potential to accelerate biomedical research but must be vetted thoroughly for robustness, interpretability, and generalizability. This study sought to create and evaluate the predictive validity of empirical phenotypes derived from assessments of signs and symptoms of psychopathology collected in a large cohort of patients using a validated semi-structured clinical interview based on the OPCRIT system. Data was available for N = 17,719 individuals diagnosed with schizophrenia (n = 10,429; 30% female), bipolar I disorder (n = 4,520; 53% female), or schizoaffective disorder (n = 2,770; 44% female) using DSM-IV-TR criteria. The best-fitting latent class analysis (LCA) model was a 6-class solution. Three of the six class profiles replicated results from previous studies. Stratifying by sex did not alter class profiles. Latent classes were significantly associated with established DSM diagnoses for primary and sex-stratified results but not for the LCA solution based primarily on symptoms and lacking indicators related to illness course and relative prevalence of psychotic versus affective symptoms in the overall clinical presentation. All LCAs (i.e., primary, sex-stratified, and symptom-only analyses) produced latent classes that showed statistically significant associations with demographic, etiological, and clinical correlates, but the effect sizes were modest for most associations and comparable in magnitude to those observed for established DSM diagnosis. These results collectively suggest that symptom-based empirical phenotypes derived from LCA may not outperform DSM diagnoses and reaffirm the long-observed importance of illness course for differentiating schizophrenia and bipolar I disorder.