What We Measure Determines What We Can Fix:

Research Infrastructure for Closed-Loop SDOH Care from the Emergency Department

Shuhan He, MD
Amal Mohamed MD
Emergency Medicine Informatics Section

In a companion piece in this section, Goss and Lupi reviewed the technology landscape for SDOH screening in the ED.¹ They closed with a line that I have not been able to put down. The data published on referral completion rates after ED-based screening they wrote, remain limited.

I think that is too kind. In most programs I have looked at, including my own, we cannot answer the question at all. We can tell you how many encounters a screen had documented. Sometimes we can tell you how many were positive. Beyond that, we have free-text notes, fax cover sheets, and the strong impression that something happened. The implementation literature has been saying as much for years. A 2021 evaluation of an integrated ED screening and referral program found that even with structured workflow design, completion of the downstream connection was the hardest step to verify.²

The instinct is to call this a measurement problem. I would call it the same problem twice. A workflow that does not produce structured, exchangeable data about what happened on the receiving end is also a workflow in which what happened on the receiving end is invisible to the next clinician. The research can’t see the loop because the loop did not actually close.

That is fixable. It needs the EM informatics community to do two unglamorous things: build the data infrastructure that makes closed-loop referrals studyable at population scale, and apply real evaluation methods to the screening, referral, and intervention pieces we already deploy.

What gets measured today, and what does not

EDs that have implemented systematic screening can usually produce a denominator of encounters with at least one screening item completed. The instrument lives in an EHR module that captures it as a discrete observation. They can usually produce a count of positive screens by domain. The value sets are standardized.

After that, the structured trail tends to go cold. Most programs cannot tell you, from structured data alone, whether a positive screen produced a placed referral. If a referral was placed, whether the receiving service acknowledged it. If acknowledged, whether anything actually happened. If something happened, what was the outcome? And if the outcome was negative, what the next move was.

Each of those is a Service Request, Task, Encounter, or Observation in FHIR terms, and each is profiled in the SDOH Clinical Care Implementation Guide that the Gravity Project maintains.³ In production, the data lives in free-text notes, fax queues, and case-management software that does not exchange with the ordering EHR. So even a serious ED-based screening program ends up producing a research dataset where the bottom of the funnel is dark. A multi-site review of EHR-based SDOH integration in 2025 reached the same conclusion: the standards layer is in place; the workflow layer is not.⁴

The fix isn’t new technology. The Gravity Project has published the FHIR profiles. The Office of the National Coordinator’s USCDI now requires SDOH assessments, goals, and interventions as data classes for certified health IT.⁵ HL7 eReferral and Da Vinci already define structured exchange of Service Requests and Tasks between ordering and receiving systems. The fix is operational adoption, and adoption is hardest where the receiving service is a community-based organization or a home-based care team whose information systems were never designed to participate in this exchange.

For programs trying to figure out where their loop breaks, decomposing the workflow into six structured stages (identification, order, acceptance, encounter, outcome, reconciliation) maps cleanly to the FHIR resources and exposes the failure points. A working walk-through of those stages, with the FHIR artifacts at each step, makes the audit easier to run. In my experience the receiving-side breaks at acceptance and reconciliation are the hardest to close, and they are exactly where the research data is leaking.

Evaluating the screening instruments

Even at the top of the funnel, we have measurement work to do. The literature on ED SDOH screening uses a heterogeneous set of instruments: AHC-HRSN, PRAPARE, CMS-aligned short forms, locally adapted checklists.⁶ The AHC-HRSN traces back to the original Accountable Health Communities tool, designed for use across clinical and community settings.⁷ The instruments differ in domain coverage, item count, reading level, and validation population. They are rarely benchmarked against each other in head-to-head studies, and when they are, agreement is imperfect.

A research-grade analysis of an SDOH screening implementation should report sensitivity, specificity, positive predictive value, and negative predictive value for each domain, with confidence intervals appropriate to the sample size. These are the same psychometric properties we expect for any laboratory or imaging test we use in the ED. A free browser-based calculator for those metrics is enough for the kind of back-of-the-envelope reporting most screening papers should be doing and frequently are not.

Inter-instrument agreement is its own problem. When a single program runs both AHC-HRSN and PRAPARE on the same patients, the percent-positive rates differ, sometimes in ways that are hard to predict. Variation in EHR-documented SDOH across institutions is also large at the network level, which makes single-site validation studies harder to generalize than they look.⁸ For two-instrument comparisons, Passing-Bablok or Deming method comparison analysis is the right tool, depending on the data structure. For program leaders deciding between instruments before running them in parallel, a side-by-side reference of the major SDOH screening tools covers the domain coverage, item counts, and validation context. What we should be aiming for is a published evidence base that lets program leaders pick a screening instrument with eyes open about how it performs in an ED population.

There is also a non-response problem. Patients with the most pronounced social risk are sometimes the least likely to complete a self-administered screen, especially on a tablet or kiosk.⁹ Patient and clinician acceptability of social risk screening varies widely across populations and care settings, and the variation is not random.¹⁰ Studies of ED screening programs need to be powered to detect differences in completion rates across demographic subgroups, and the relevant power calculation is straightforward to run before the study starts. A lot of published implementation studies are underpowered for the subgroup question, and we still do not know whether digital-only screening is widening or narrowing the social risk detection gap.

Evaluating the interventions

Once a screen is positive and a referral is placed, the research question shifts from psychometrics to efficacy. Does the referral, conditional on completion, change the outcome that motivated the screen?

For a long time the published answer was a shrug. The seminal systematic reviews of social and economic needs interventions concluded that the evidence base was small, methodologically heterogeneous, and dominated by short-term outcomes.¹¹ More recent work is better. A 2025 systematic review of structured SDOH data collection paired with referral action found reductions in health service utilization and cost across a range of program designs.¹² Reporting these intervention studies should follow the same evaluative discipline as any clinical trial: number needed to treat, relative risk reduction, absolute risk reduction. Effect-size calculations belong in every intervention paper that wants to be useful to a program leader trying to make a budget case.

The harder methodological problem is that closed-loop SDOH interventions are almost never single interventions. They are bundled care pathways, and the bundle varies by site. A community health worker home visit after a positive food insecurity screen looks different in Boston than in rural Ohio. Historically the home-based response side of the bundle has been the under-instrumented half. If we want to study the response side rigorously, the dispatch and routing infrastructure has to produce structured data, end to end: which clinician went where, what was performed, what the outcome was, all linkable back to the index ED encounter.

The operations research literature on multi-resource home health dispatch is mature; recent reviews summarize the routing and scheduling formulations production systems use.¹³ For ED-facing applications, purpose-built dispatch deployments translate that operations research into the actual layer an ED-to-home referral rides on. The point for research infrastructure is that whatever layer a program uses has to emit structured data that lets the index ED encounter, the dispatched visit, and the outcome be linked.

What closed loop looks like as research data

If a program wants to be able to answer the questions Goss and Lupi flagged, the data infrastructure has to produce, at minimum: a Service Request representing the ED-placed referral, a Task tracking acceptance and status transitions, an Encounter or Procedure for the actual service event on the receiving side, an Observation or Goal capturing the outcome, and a Communication or Document Reference linking the outcome back to the original order.

The other half of the data infrastructure is the documentation we do at the front end: the ICD-10 Z55 to Z65 codes that turn a free-text social history into a structured, searchable, billable element. Coding the Z is what makes the encounter visible to the population health dashboard, the quality reporting submission, and any future research enterprise. A working reference of the most common Z-codes ED teams encounter, with a documentation template for each, is one way to lower the friction of getting them into the chart.

Recent work using LLMs to extract SDOH from clinical notes has shown that NLP can recover much of what is lost when documentation lives only in free text, with sensitivity an order of magnitude higher than diagnostic-code capture alone.¹⁴ A systematic review of this literature in 2021 concluded that automated extraction of SDOH was feasible and improving rapidly, while flagging substantial heterogeneity in how SDOH concepts are defined across studies.¹⁵ NLP is a useful tool. It is also a workaround. The research case is much stronger when the structured data exists at the point of care, rather than being recovered after the fact by a model whose performance varies across institutions and demographic subgroups.

What EM informatics researchers can do now

A few practical moves are within reach for any ED informatics group, regardless of institutional resources.

The first is to audit the loop. For a defined cohort, say all positive food insecurity screens over a 90-day period, trace every referral end to end. Where does the data live as structured FHIR resources, where does it live as free text, and where does it not live at all? The audit itself is a publishable contribution to literature that needs more program-level transparency.

The second is to report the psychometrics. Programs that have switched between screening instruments, or that run more than one in parallel, are sitting on inter-instrument agreement data that the field needs. A short methods paper with a Passing-Bablok analysis and a confusion matrix is more useful than another implementation case report.

The third is to push for structured data capture from the receiving services. That includes community-based organizations and home-care teams. Where the receiving system cannot emit FHIR resources, push for a structured case-management export that can be linked back to the index ED encounter. Without it, every other research question downstream is a guess.

There is a policy lever too. Section members who participate in ACEP advocacy, in USCDI public comment, and in HL7 standards development are in a position to shape the next round of SDOH data classes that certified health IT must support. The standards are being written now.

The bottom line

The ED-to-community SDOH workflow is a coordination problem on the receiving side and a data problem across the whole loop. Those two are connected. If you cannot route the response, you cannot study it. If you cannot study it, you do not actually know whether it worked.

EM informatics has the standards, the methods, and the tools to close this. The work is to apply them, in real programs, until the data we have on what happens to our patients after they leave the department starts to look like the data we have on what happens to them while they are still in the department.

References

1. Goss F, Lupi A. From intake to intervention: how technology is transforming SDOH screening in the emergency department. ACEP Emergency Medicine Informatics Section Newsroom; 2026.
2. Wallace AS, Luther BL, Sisler SM, Wong B, Guo JW. Integrating social determinants of health screening and referral during routine emergency department care: evaluation of reach and implementation challenges. Implement Sci Commun. 2021;2(1):114.
3. Gravity Project. SDOH Clinical Care Implementation Guide. Accessed April 30, 2026.
4. Chishtie JA, Tea J, Ester M, et al. Integration of screening and referral tools for social determinants of health and modifiable lifestyle factors in the Epic electronic health record system: scoping review. J Med Internet Res. 2025;27:e73615.
5. Office of the National Coordinator for Health Information Technology. United States Core Data for Interoperability (USCDI). Accessed April 30, 2026.
6. Loo S, Molina M, Ahmad NJ, et al. Implementing social determinants of health screening in US emergency departments. JAMA Netw Open. 2025;8(3):e250137.
7. Billioux A, Verlander K, Anthony S, Alley D. Standardized screening for health-related social needs in clinical settings: the Accountable Health Communities screening tool. NAM Perspectives. 2017.
8. Cottrell EK, Dambrun K, Cowburn S, et al. Variation in electronic health record documentation of social determinants of health across a national network of community health centers. Am J Prev Med. 2019;57(6 Suppl 1):S65-S73.
9. Vest JR, Mazurenko O. Non-response bias in social risk factor screening among adult emergency department patients. J Med Syst. 2023;47(1):78.
10. de Marchis EH, Hessler D, Fichtenberg C, et al. Part I: A quantitative study of social risk screening acceptability in patients and caregivers. Am J Prev Med. 2019;57(6 Suppl 1):S25-S37.
11. Gottlieb LM, Wing H, Adler NE. A systematic review of interventions on patients’ social and economic needs. Am J Prev Med. 2017;53(5):719-729.
12. Yan AF. Collecting and using social needs data in health settings: a systematic review of health service utilization and costs. PMC/Academic Press; 2025.
13. Atta MN, Goyal SK, Jung H, et al. A concise review of the home health care routing and scheduling problem. Operations Research Perspectives. 2025;15:100347.
14. Guevara M, Chen S, Thomas S, et al. Large language models to identify social determinants of health in electronic health records. NPJ Digit Med. 2024;7(1):6.
15. Patra BG, Sharma MM, Vekaria V, et al. Extracting social determinants of health from electronic health records using natural language processing: a systematic review. J Am Med Inform Assoc. 2021;28(12):2716