Real‑World Clinical Characterization of Major Depressive Disorder and Treatment‑Resistant Depression Supported by Natural Language Processing: Multicenter Observational Study From the MOOD Project

AI Summary

AI-supported clinician-oriented methodology integrating structured EHR data and NLP-extracted free text to profile MDD and TRD across two hospital sites.
Cohort of 1147 adults with MDD; 46% met TRD criteria; 62.9% female; mean age 57.8 years; overall mortality 13.3%.
High prevalence of CNS and heart comorbidities; anxiety and dementia more common in TRD; severity often undocumented and rating scales rarely recorded.

Interact J Med Res. 2026 May 22;15:e86448. doi: 10.2196/86448.

ABSTRACT

BACKGROUND: Major depressive disorder (MDD) and treatment-resistant depression (TRD) are heterogeneous conditions in which key clinical details are split across structured fields and free-text notes in electronic health records (EHRs), constraining population-level insight and timely audit of care quality.

OBJECTIVE: This study aims to present a clinician-oriented, artificial intelligence-supported real-world evidence (RWE) methodology integrating structured and unstructured EHR data to profile MDD and TRD, and report comorbidity patterns from a 2-site pilot. This analysis reports the first objective of the MOOD project, which is to characterize the real‑world clinical and disease severity profile of patients with MDD and treatment‑resistant depression, providing a necessary foundation for subsequent evaluations of treatment patterns and outcomes.

METHODS: We conducted a retrospective study in 2 Belgian hospitals (September 2021-June 2023). Adults (aged ≥18 years) with MDD were identified via DSM-IV (Diagnostic and Statistical Manual of Mental Disorders [Fourth Edition]) and ICD-10 (International Statistical Classification of Diseases, Tenth Revision) codes or natural language processing-detected note mentions; bipolar depression was excluded. TRD was defined as initiation of a third distinct antidepressant, supplemented by explicit mentions of TRD in notes. Structured data (demographics, diagnoses, medications, and hospitalizations) were harmonized in an Observational Medical Outcomes Partnership warehouse. Free-text notes were processed with a natural language processing pipeline to capture symptoms, psychiatric comorbidities, and contextual events.

RESULTS: We identified 1147 adults with MDD, of which 46% (524/1147) met TRD criteria. Females comprised 62.9% (722/1147) and mean (SD) age was 57.8 (18.4) years. Mortality was 13.3% (152/1147) overall (57/1147, 10.9% TRD vs 95/1147, 15.2% non-TRD). Common medical comorbidities were central nervous system diseases (477/1147, 41.6%) and heart diseases (349/1147, 30.4%). Dementia was more frequent in TRD (42/1147, 8% vs 32/1147, 5.1%), whereas obesity was higher in non-TRD (70/1147, 11.2% vs 46/1147, 8.8%). Anxiety disorder occurred in 35.4% (406/1147) overall and was more prevalent in TRD (229/1147, 43.7% vs 177/1147, 28.4%); personality and panic disorders also trended higher. Severity was sparsely documented (severe MDD 170/1147, 14.8%) and standardized scales were rarely recorded.

CONCLUSIONS: We present a step-by-step artificial intelligence-supported methodology tailored for clinicians, discussing challenges in integrating RWE into psychiatry, and identifying opportunities to enhance data collection with minimal workflow changes, which emphasizes the transformative potential of RWE systems in mental health research.

PMID:42172608 | DOI:10.2196/86448

Document this CPD