Addressing Gaps in Real-World Data Sources through Automated Extraction of Psychiatric Comorbidities and Medication Side Effects from Clinical Notes

OBJECTIVES: Real-world data (RWD) sources are useful for research on depression treatment and outcomes, but RWD studies are often limited to data contained in structured fields. Some critical variables, such as psychiatric conditions that co-occur with depression, may be recorded in clinical notes rather than in structured fields. This study aimed to assess the feasibility of identifying and extracting psychiatric comorbidities from clinical notes drawn from a RWD source.

METHODS: Data for this study were drawn from the OM1 Real World Data Cloud (OM1, Inc, Boston, MA, USA), a deterministically linked, de-identified, individual-level dataset containing EMR with medical and pharmacy claims from 2013 to present. The study cohort was restricted to patients with a diagnosis of major depressive disorder. A natural language processing-based approach was used to extract mentions of psychiatric comorbidities using a collection of common linguistic patterns to identify comorbidities in clinical notes and categorize them as affirmations (presence) or negations (absence). Language models were then constructed and validated by subject matter experts to ensure reliability of the patterns.

RESULTS: Using these inclusion criteria, 2.9 million notes were identified for analyses. Comorbidities of interest were generalized anxiety disorder, obsessive-compulsive disorder, panic disorder, post-traumatic stress disorder, social anxiety disorder, substance use disorder, alcohol use disorder, bipolar disorder, schizophrenia, and insomnia. Of the included notes, 1,703,849 were classified as affirmations. Generalized anxiety disorder was mentioned most frequently, followed by bipolar disorder, anxiety disorder, and substance use disorder.

CONCLUSIONS: Psychiatric comorbidities that commonly co-occur with depression can be extracted from clinical notes from RWD using a computerized natural language processing-based approach. The approach used in this study may be applied to other concepts related to depression or other clinical areas. Use of this approach may improve characterization of depression patients in RWD sources and improve the usefulness of RWD for depression research.