Background: Hospitalization for major depressive disorder (MDD) is a burden for patients, their families, and healthcare systems. Understanding predictive factors of hospitalization could lead to better outcomes.
Objectives: To determine predictors of hospitalization among patients with a new diagnosis of MDD.
Methods: Retrospective cohort study of real-world behavioral health data derived from electronic health records (EHR) comprising structured and unstructured longitudinal clinical data. Adults (≥18 years) with ≥2 MDD diagnoses between 09/2009 and 06/2020 were identified from the NeuroBlu Database. Index date was second diagnosis. Patients were required to have insurance enrollment for 60 days before and 30 days after index. Exclusion criteria included hospitalizations within 30 days of index; a diagnosis of bipolar disorder, schizophrenia, or schizoaffective disorder prior to day 30; or missing sex, race, marital status, or clinical global impression scale (CGIS). Hospitalization was modeled by multivariable Cox proportional hazard models with elastic-net regularization. Diagnosis of bipolar disorder, schizophrenia, or schizoaffective disorder was entered as a time-varying covariate. Hazard ratio (HR) and 95% confidence intervals (CI) are reported. Model performance was estimated by C-index using 10-fold cross-validation (CV). IRB approval was obtained and included a waiver of HIPAA authorization.
Results: 7286 MDD patients were identified and randomized to train (N=5116) and test (N=2170) sets. Both cohorts had similar demographic and clinical characteristics. Train set patients were mostly unmarried (66.1%), white (81.2%), females (68.2%), with non-severe MDD (85%), and median CGIS 4. Median time to hospitalization was 12.6 months (43.5% patients hospitalized) and 12.9 months (41.9% patients hospitalized) in train and test sets, respectively. Important predictors of hospitalization included marital status (widowed HR=1.47, p< 0.001; divorced/separated HR=1.23, p=0.001; single HR=1.32, p< 0.001; ref=married), CGIS (HR=1.20, p< 0.001), substance use (HR=1.70, p< 0.001), and stressors related to family (HR=1.35, p< 0.001, finances (HR=1.15, p=0.011), legal issues (HR=1.25, p< 0.001), and occupation (HR=1.16, p=0.01). The model performed similarly in the train (CV C-index=0.687, 95% CI 0.668-0.688) and test sets (C-index=0.672, 95% CI 0.653-0.691).
Conclusions: Substance use, psychosocial stressors, and illness severity are strong predictors of hospitalization among newly-diagnoses MDD patients. However, stressors and CGIS are not routinely collected in EHR or claims data. Further development of comprehensive risk assessments for behavioral health outcomes is needed.
