Extracting SDOH and Mapping to ICD-10-CM Z-Codes

Social determinants of health (SDOH) — the conditions in which people live, work, and age — drive an estimated 80% of health outcomes, yet they live almost entirely in free-text narrative. Multiple chart-review studies find SDOH documented in notes but coded with a Z-code under ~2% of the time. The information is there; the structured signal is not. This skill recovers it: run OpenMed NER over de-identified notes, then map the resulting spans to the ICD-10-CM Z55–Z65 family.

When to use

• A note clearly describes a social risk ("lives in her car", "skips meals to afford insulin", "no ride to dialysis") and you want a coded, queryable signal.
• You are building health-equity dashboards, risk stratification, or closed-loop referral feeds and need SDOH as discrete data.
• You want to reconcile what the chart says against what was coded, and flag Z-code gaps for a coder or care team to confirm.

This is a decision-support step. It proposes Z-codes; a human assigns them. SDOH coding is sensitive — never expose individual SDOH inferences outside the care/coding workflow, and never feed them to coverage or pricing decisions.

Quick start

De-identify first, run NER, then map spans to Z-codes:

import openmed
from sdoh_zcode_map import SDOH_ZCODES  # see references/sdoh_zcode_map.md

note = (
    "62F with CHF. Reports she lost her apartment last month and is "
    "staying in a shelter. Often runs out of food before month-end. "
    "No car; misses appointments because the bus does not run to clinic."
)

# 1) Strip PHI before any downstream processing or storage.
deid = openmed.deidentify(note, method="replace", policy="hipaa_safe_harbor")

# 2) Run clinical NER. Use an SDOH/clinical model from the registry; discover
#    available keys with openmed.get_models_by_category(...).
result = openmed.analyze_text(deid.text, output_format="dict")

# 3) Map each entity span to a candidate Z-code.
for ent in result["entities"]:
    code = SDOH_ZCODES.get(ent["label"].lower())
    if code:
        print(f"{ent['text']!r:40}  {ent['label']:18}  -> {code}")

analyze_text returns entities shaped as {"text", "label", "confidence", "start", "end", "metadata"}. The start/end offsets index into the text you passed in, so you can anchor every suggested Z-code back to its exact source span for human review.

Workflow

1. De-identify the note with openmed.deidentify (HIPAA Safe Harbor or a stricter policy). SDOH text is dense with PHI (addresses, employer names).
2. Extract entities with openmed.analyze_text. Pick a model whose label set covers social concepts; if your model only emits clinical findings, run a second pass with a zero-shot model (openmed zero) using SDOH labels such as housing_instability, food_insecurity, unemployment, transportation_barrier, social_isolation, financial_strain.
3. Map spans to Z-codes using a curated lookup keyed by label (references/sdoh_zcode_map.md). Keep the span offsets and the model confidence on every suggestion.
4. Stage for confirmation. Emit (span, label, suggested_code, confidence) tuples for a coder or the Gravity Project pipeline to accept or reject. Do not auto-bill a Z-code from an inference alone.
5. Normalize to value sets. Align labels to the Gravity Project SDOH domains so codes are interoperable with FHIR (Condition, Observation, Goal) and USCDI v3 SDOH elements.

Z-code families you will hit most (ICD-10-CM Z55–Z65)

The full curated label→code table lives in references/sdoh_zcode_map.md.

Hand-off to / from OpenMed

• From OpenMed: this skill consumes openmed.analyze_text(...) output (PredictionResult dict). Each entity["start"]/["end"] anchors a Z-code suggestion to source text.
• To OpenMed: always run openmed.deidentify upstream so no raw PHI reaches the SDOH store, logs, or coder queue.
• Onward: emit suggestions into a FHIR Condition/Observation with the Z-code as code.coding (system http://hl7.org/fhir/sid/icd-10-cm). OpenMed's openmed.clinical.exporters.fhir helpers (to_bundle, to_operation_outcome) assemble the envelope; ICD-10-CM itself is public-domain in the US release.

Edge cases & gotchas

• Negation and history. "Denies food insecurity" or "previously homeless, now housed" must not produce an active Z-code. Run negation/temporality resolution (openmed.clinical, resolving-clinical-context) before mapping.
• Hypotheticals and screening prompts. Template text ("Do you have stable housing?") and family-member SDOH ("his mother is unhoused") are common false positives — check the subject and modality.
• One span, one domain. Do not stack multiple Z-codes onto one phrase; map to the most specific single code and let the coder add others.
• Granularity drift. ICD-10-CM adds SDOH codes most fiscal years (e.g. Z59.4x food, Z59.82 transportation). Pin your code set to a release year and re-validate annually.
• Do not infer protected attributes. Surface only what the note states; never derive race, immigration status, or income bracket as an SDOH "finding".
• Restricted terminology. SNOMED CT SDOH refsets and LOINC SDOH panels are licensed separately — OpenMed does not bundle them; load the user's own copy out-of-process if you cross-map beyond ICD-10-CM.

Standards & references

• ICD-10-CM official guidelines, Z55–Z65 SDOH codes (CDC/CMS, public domain): https://www.cdc.gov/nchs/icd/icd-10-cm.htm
• Gravity Project (HL7 SDOH Clinical Care value sets & FHIR IG): https://www.hl7.org/gravity/ and https://confluence.hl7.org/display/GRAV
• n2c2 2022 Track 2 — SDOH extraction shared task (Social History Annotation Corpus): https://n2c2.dbmi.hms.harvard.edu/
• CMS ICD-10-CM Z-code SDOH resources: https://www.cms.gov/files/document/zcodes-infographic.pdf
• USCDI SDOH data classes: https://www.healthit.gov/isa/uscdi-data-class/sdoh