clinical-note-summarization

Clinical Note Summarization

ID: biomedical.clinical.note_summarization Version: 1.1.0 Status: Production Category: Clinical AI / Documentation

---

Overview

The Clinical Note Summarization Skill converts unstructured clinical text (dictations, progress notes, discharge summaries) into structured SOAP format with explicit guardrails for accuracy, missing data, and contradictions. It is designed for documentation support workflows that require consistent structure, traceability, and safe handling of PHI.

---

Inputs

Field	Type	Notes
patient_context	dict	Optional metadata (age, sex, encounter type, PMH). Avoid PHI unless required.
note_text	str	Raw clinical note text (dictation, OCR, or EHR export).
output_format	str	markdown or json (recommended for downstream validation).

Recommended minimum input: note_text plus encounter type and age.

---

Outputs

SOAP Summary (Markdown)

• Subjective - HPI, ROS, symptoms, and patient-reported history
• Objective - vitals, exam findings, labs/diagnostics
• Assessment - problems with ICD-10 suggestions when supported
• Plan - actions per problem with timing and responsible role

Optional JSON Schema

{
  "soap": {
    "subjective": ["..."] ,
    "objective": ["..."] ,
    "assessment": [
      {"problem": "...", "icd10": "...", "evidence": "..."}
    ],
    "plan": [
      {"action": "...", "responsible_role": "...", "timing": "..."}
    ]
  },
  "alerts": ["..."],
  "missing_information": ["..."]
}

---

Workflow

1. Normalize input - identify vitals, labs, meds, and timelines; standardize units.
2. Extract structure - map content into SOAP sections.
3. Validate - check for contradictions, missing fields, and medication duplication.
4. Output - return structured SOAP and a data gaps checklist.

---

Prompt Template

Use prompt.md as the canonical system prompt. It enforces:

• No invention of data
• Explicit contradictions and missing info
• Problem-linked assessment and plan

---

Integration Examples

LangChain

from langchain.prompts import PromptTemplate
from langchain.chat_models import ChatOpenAI

with open("prompt.md") as f:
    system_prompt = f.read()

template = PromptTemplate.from_template(
    system_prompt + "\n\nClinical Note:\n{note}\n\nStructured SOAP:"
)

chain = template | ChatOpenAI(model="gpt-4")
result = chain.invoke({"note": clinical_note_text})

Direct API (OpenAI)

import openai

with open("prompt.md") as f:
    system_prompt = f.read()

response = openai.ChatCompletion.create(
    model="gpt-4",
    messages=[
        {"role": "system", "content": system_prompt},
        {"role": "user", "content": f"Structure this note:\n{note}"}
    ]
)

---

Guardrails

• No hallucinations: If not stated, say "not provided".
• No clinical decisions: This is documentation assistance only.
• PHI safety: Use HIPAA-compliant environments; avoid logging raw notes.
• Auditability: Store prompt and model metadata for traceability.

---

Validation Hooks

• Compare extracted vitals/labs against source text.
• Flag missing allergies, meds, or discharge plan.
• Require clinician review before EHR insertion.

---

Files

File	Description
prompt.md	System prompt template for SOAP structuring
usage.py	Example integration script

---

Supported Models

Provider	Model	Performance
OpenAI	gpt-4, gpt-4-turbo	Excellent
Anthropic	claude-3-opus, claude-3-sonnet	Excellent
Google	gemini-1.5-pro	Good
Open Source	Llama 3 70B, Mixtral 8x22B	Good (fine-tuning recommended)

---

Related Skills

• Trial Eligibility Screening: Use structured notes for matching.
• Medical Coding: ICD-10/CPT assignment from structured output.

---

Author

MD BABU MIA Artificial Intelligence Group Icahn School of Medicine at Mount Sinai md.babu.mia@mssm.edu