Validation of an electronic health record algorithm for identifying housing-related needs in a safety-net health system

OBJECTIVE: Accurate, efficient identification of housing-related needs, including homelessness or housing instability, is crucial for health systems addressing health-related social needs (HRSN). We developed and validated a novel, pragmatic electronic health record (EHR)-based algorithm to identify patients with housing-related needs. 

STUDY DESIGN AND SETTING: We retrospectively evaluated sensitivity and specificity of the housing-related needs algorithm within our safety-net hospital, Boston Medical Center (BMC). 

DATA SOURCES AND ANALYTIC SAMPLE: The algorithm included six EHR structured data elements tailored to BMC operations, including HRSN screening results. We assessed each element's performance, alone and combined, using 12 months of BMC EHR data among two reference groups: (1) 433 patients with verified housing-related needs at housing program enrollment (2019-2023), and (2) a stratified random sample of 400 patients (200 adult, 200 pediatric) with ≥ 1 primary care medical visit (2022), whose charts we manually reviewed to verify housing status. We calculated algorithm sensitivity in both groups and specificity in the primary care group. 

PRINCIPAL FINDINGS: With all data elements included, algorithm sensitivity was 62% (95% CI: 57%-66%) among housing program enrollees and 81% (95% CI: 68%-91%) among primary care patients. Among primary care patients (13% with chart review-verified housing-related needs), specificity was 97% (95% CI: 95%-98%). HRSN screening yielded the highest single-element sensitivity, but screening alone remained limited: 57%-62% of those with verified housing-related needs were detected via screening. Patient address information and diagnostic codes had low single-element sensitivities. 

CONCLUSION: Pragmatic EHR algorithms leveraging structured data elements tailored to local context present an accessible, efficient method for health systems to identify patients with housing-related needs. This is the first study to validate such an algorithm in a safety-net setting; we found it had moderate sensitivity and high specificity. The algorithm identified more housing-related needs than diagnostic codes alone, demonstrating the value of integrated clinical and administrative data. Further algorithm improvements require changes to HRSN screening and EHR documentation.