Lead/Presenter: Stephanie Chassman, COIN - Los Angeles
All Authors: Chassman SA (Desert Pacific Mental Illness, Research, Education, and Clinical Center (MIRECC)), Chapman AB (IDEAS Center, Veterans Affairs Salt Lake City Health Care System; Department of Internal Medicine, University of Utah School of Medicine) Cordasco KM (VA LA HSR&D Center of Innovation) Nelson RE (IDEAS Center, Veterans Affairs Salt Lake City Health Care System; Department of Internal Medicine, University of Utah School of Medicine) Montgomery AE (National Center on Homelessness among Veterans, University of Alabama School of Public Health; Birmingham VA Medical Center, Birmingham, AL) Jackson NJ (Department of Medicine Statistics Core at UCLA) Gabrielian SE (VA LA HSR&D Center of Innovation)
Objectives:
Measuring longitudinal housing stability is critical for evaluating evidence-based practices the VA has implemented to address Veteran homelessness. However, it is challenging to measure Veterans’ housing outcomes due to a lack of standardized documentation in the VA’s Electronic Health Record (EHR). The goal of this study was to compare available approaches to measure housing stability using VA data in a cohort of homeless-experienced Veterans (HEVs) engaged in the recently established VA Grant and Per Diem “Aftercare†program (which offers time-limited case management for HEVs undergoing housing transitions).
Methods:
In a cohort of 334 Veterans enrolled in Aftercare during FY2021 with any VA encounters, we gathered retrospective data from VA Corporate Data Warehouse (CDW) for the study period of July 2021 to June 2022. We captured structured data elements from CDW that indicate unstable housing: ICD-10 codes, use of outpatient or inpatient VA homeless services, and responses to the Homelessness Screening Clinical Reminder (HSCR). To capture information contained in unstructured clinical progress notes, we applied a previously validated natural language processing (NLP) system to identify notes documenting unstable housing. We compared counts of Veterans with one or more sources of documentation for unstable housing and tested the pairwise probability of Veterans having each source of documentation using McNemar’s test of marginal homogeneity.
Results:
Two-thirds (n = 225, 67%) of the Veterans in our cohort had at least one documented instance of housing instability during the study period. The source with the highest yield was NLP (n = 197, 59%). Of Veterans identified by NLP as experiencing unstable housing, 67 (20%) did not have any structured data elements of housing instability. Less than half the cohort (n = 158, 47%) had at least one structured data element; the most common sources were outpatient homeless service use (n = 127, 38%) and ICD-10 codes (n = 113, 34%). A small portion, (n = 28, 8%) of individuals with structured data elements did not have any evidence in clinical notes. All pairs of data sources showed statistically significant heterogeneity (p < 0.001) except for ICD-10 versus outpatient homeless service visits (p = 0.10) and inpatient services versus HSCR (p = 0.17).
Implications:
Multiple sources of documentation were necessary for maximizing ascertainment of longitudinal housing stability in VA’s EHR. In this cohort, some instances of homeless experiences were only documented in clinical notes and would not be captured using structured data elements alone. NLP may be more sensitive in identifying unstable housing than structured documentation (e.g., ICD-10 codes).
Impacts:
Multiple data elements, including unstructured information extracted from clinical notes using NLP and structured data elements, should be used to assess housing outcomes of VA interventions for HEVs.