Machine learning for novel phenotyping in schizophrenia

Heterogeneity among people diagnosed with schizophrenia-spectrum disorders (schizophrenia) and high prevalence of co-occurring disorders makes identification of optimal treatments difficult. This study identified behavioral health phenotypes using machine learning with Medicaid claims of adults with schizophrenia. We compared the phenotypes' clinical outcomes and psychotropic medication prescription patterns for clinical validity.

Using national Medicaid claims from January 2010 - December 2012 we identified 249,006 adults ages 18-64, with ≥ 1 inpatient and/or ≥ 2 outpatient claims with principal or secondary diagnoses of schizophrenia (ICD9 295.xx) in 2010. Latent Dirichlet Allocation (LDA) incorporated their behavioral health co-occurring disorders in 2010 to identify behavioral health phenotypes, validated using 5-fold cross validation. Pairwise comparisons among each phenotype of psychotropic medication types, and likelihoods of any behavioral health inpatient admission or emergency department (ED) visit in 2011 were conducted.

LDA with 5-fold cross validation identified 5 behavioral health phenotypes we labeled depression, substance use, mania-mixed mood, anxiety-paranoid, and conduct disorder-developmentally delayed; a sixth phenotype had no co-occurring disorders. Likelihoods of behavioral health inpatient admissions and ED visits were significantly different between the phenotypes. Psychotropic medications prescribed to the phenotypes were distinct. Post-hoc analyses using the same methods with 2017 Medicaid claims of 383,849 adults identified comparable phenotypes.

This study demonstrated the feasibility of using machine learning with claims data to identify behavioral health phenotypes among individuals with schizophrenia. Future pharmacoepidemiologic investigations addressing confounding bias will compare effectiveness of treatments for each phenotype, informing efforts to identify personalized treatments for people with schizophrenia.