EHR Phenotyping of Depressed Patients: A Hierarchical Bayesian Latent Variable Modeling Approach

• Activity Number: 66 - Novel Bayesian Methodology with Health Applications
• Type: Contributed
• Date/Time: Monday, August 3, 2020 : 10:00 AM to 2:00 PM
• Sponsor: Mental Health Statistics Section
• Author(s): Wenna Xi* and Samprit Banerjee
• Companies: Weill Medical College, Cornell University and Weill Medical College, Cornell University
• Keywords: Bayesian modeling; hierarchical modeling; latent variable modeling; electronic health records; phenotyping

Abstract

Early diagnoses and interventions of major depression disorder (MDD) are important to prevent the development of its sequelae. Unfortunately, MDD is highly underdiagnosed because of social stigma towards mental health and a lack of routine screening in the primary care setting. The underdiagnosis of MDD biases prevalence estimates and other downstream analyses of data captured in the electronic health records (EHR) of healthcare systems. In this study, we propose a hierarchical Bayesian latent variable model to identify potentially depressed patients who remain undiagnosed in the EHR data. Specifically, we identify the latent severity level of depression for each patient and the difficulty of getting a diagnosis of MDD, and assume that the probability of getting a diagnosis is a function of the two. We illustrate our model via simulations and apply our model to the outpatient EHR data at a large New York City-based academic medical center. On the subset of patients who have been screened by the patient health questionnaire (PHQ-9), we compare their posterior latent depression severities with their PHQ-9 scores as a validation of our method.