Machine learning for novel phenotyping in schizophrenia.

Area of Science:

• Psychiatry
• Computational Psychiatry
• Health Services Research

Background:

• Schizophrenia exhibits significant heterogeneity and high comorbidity rates, complicating treatment optimization.
• Identifying distinct patient subgroups is crucial for developing personalized treatment strategies.

Purpose of the Study:

• To identify behavioral health phenotypes in adults with schizophrenia using machine learning on Medicaid claims data.
• To assess the clinical validity of identified phenotypes by comparing outcomes and psychotropic medication patterns.

Main Methods:

• Latent Dirichlet Allocation (LDA) was employed on national Medicaid claims data (2010-2012) for 249,006 adults diagnosed with schizophrenia.
• Phenotypes were identified based on co-occurring behavioral health disorders and validated using 5-fold cross-validation.
• Comparisons included psychotropic medication types and rates of inpatient admissions and emergency department visits.

Main Results:

• Five distinct behavioral health phenotypes were identified: depression, substance use, mania-mixed mood, anxiety-paranoid, and conduct disorder-developmentally delayed, plus a no-comorbidity group.
• Significant differences were observed in inpatient admission and emergency department visit likelihoods across phenotypes.
• Distinct psychotropic medication prescribing patterns were associated with each phenotype.

Conclusions:

• Machine learning effectively identified behavioral health phenotypes in individuals with schizophrenia using claims data.
• These phenotypes demonstrate clinical validity through differential outcomes and medication use.
• Further research is needed to compare treatment effectiveness for each phenotype to inform personalized medicine for schizophrenia.