summarizing-clinical-notes

Summarizing Clinical Notes with Span Citations

A clinical summary is only useful if it is faithful: every statement must trace back to something the chart actually says. The failure mode for note summarization is the confident hallucination — an invented dose, a fabricated allergy, a discharge diagnosis that was never made. This skill produces summaries where each line cites the source span that supports it, so a clinician can verify in one glance and catch any fabrication.

Not a medical device. OpenMed and this skill assist documentation; they do not diagnose, triage, or make autonomous clinical decisions. Every summary is a draft for clinician review and editing. Surface that disclaimer in any UI that renders these summaries.

When to use

• Drafting a discharge summary, transfer note, or SBAR/handoff from a long encounter.
• Building a problem-oriented view (problem list with supporting evidence).
• Generating a "one-liner" (the single-sentence patient summary) for rounds.
• Chart abstraction where reviewers need quick, verifiable evidence pointers.

Quick start

De-identify before anything else, extract entities to anchor against, then compose the summary with citations:

import openmed

note = """\
HPI: 68M with HTN, T2DM presents with 3 days of productive cough and fever to
38.9C. CXR shows RLL infiltrate. Started on ceftriaxone and azithromycin.
Hospital course: improved on IV antibiotics, transitioned to PO. Discharged on
amoxicillin-clavulanate. Follow up with PCP in 1 week.
"""

# 1) ALWAYS de-identify before summarizing or sending text anywhere.
deid = openmed.deidentify(note, method="replace", policy="hipaa_safe_harbor")

# 2) Extract entities; their offsets become your citation anchors.
ner = openmed.analyze_text(deid.text, output_format="dict")
spans = {
    (e["start"], e["end"]): e["text"]
    for e in ner["entities"]
}

# 3) Compose the summary. Every bullet references a (start, end) span so a
#    reviewer can click back to the exact evidence.
def cite(start, end):
    return f"[{start}:{end}] {deid.text[start:end]!r}"

# Example problem-oriented line, grounded in detected spans:
# "Community-acquired pneumonia (RLL infiltrate) — treated with ceftriaxone +
#  azithromycin." with cite(...) anchors for each entity.

analyze_text returns entities as {"text", "label", "confidence", "start", "end", "metadata"}; the start/end offsets index the de-identified text, giving you exact, verifiable citation anchors.

Workflow

1. De-identify with openmed.deidentify. Summaries are often shared or logged; PHI must be gone before this stage. Keep the mapping (keep_mapping=True) only if a downstream clinician must re-identify in a controlled context — never persist the mapping with the summary.
2. Extract grounding spans with openmed.analyze_text (problems, meds, labs, procedures). These define the allowed evidence set: a summary claim that cannot point at a span is unsupported.
3. Resolve context with openmed.clinical (negation, temporality, subject) so "no chest pain" and "father had MI" are not summarized as active patient problems. See resolving-clinical-context.
4. Compose by view:
   • One-liner: age/sex + key chronic problems + reason for encounter.
   • Hospital course: ordered problems → intervention → response, each line citing the spans it summarizes.
   • Problem-oriented: group entities into problems; attach supporting med/lab/procedure spans under each.
5. Enforce citation coverage. Reject or flag any output sentence with zero span citations. This is the anti-hallucination gate — keep it strict.
6. Mark it a draft. Render the medical-device disclaimer and require human sign-off before the summary enters the record.

Hand-off to / from OpenMed

• From OpenMed: consumes openmed.deidentify(...) output (de-identified text + entity spans) and openmed.analyze_text(...) (PredictionResult dict). Entity start/end offsets are the citation anchors.
• To OpenMed: the summary text itself can be re-run through openmed.analyze_text for a coded problem list, or through openmed.eval leakage gates to confirm no PHI leaked into the generated summary.
• Citation rendering: analyze_text(..., output_format="html") produces a span-highlighted view of the source — handy for a click-to-evidence UI.

Edge cases & gotchas

• Hallucination is the failure mode. If your summary backbone is an LLM, constrain it to the entity/span set and require a citation per sentence; do not let it introduce facts (doses, diagnoses, dates) absent from the spans.
• Negation & family history. Always run context resolution first; "denies", "ruled out", "FH of" must not become patient problems.
• Copy-forward / note bloat. EHR notes carry stale copy-pasted blocks. Cite the most recent supporting span and prefer the current encounter's text.
• Conflicting statements. When the chart contradicts itself (two different discharge diagnoses), surface both with citations rather than silently picking one.
• No autonomous action. Never auto-finalize, auto-sign, or auto-route a summary; it is decision support, not a clinical decision.
• PHI in the summary. A summary can re-introduce identifiers the model missed in the source. Run the output through openmed.extract_pii or an openmed.eval leakage gate before display or storage.

Standards & references

• HL7 C-CDA Discharge Summary / Continuity of Care Document section structure: https://www.hl7.org/ccdasearch/
• Joint Commission discharge summary required elements (CAMH / record of care): https://www.jointcommission.org/
• SBAR handoff communication (IHI): https://www.ihi.org/resources/tools/sbar-tool-situation-background-assessment-recommendation
• Weed LL, problem-oriented medical record (POMR) — the origin of problem-oriented summaries: N Engl J Med, 1968.
• FDA Clinical Decision Support Software guidance (device vs. non-device CDS): https://www.fda.gov/regulatory-information/search-fda-guidance-documents/clinical-decision-support-software