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.