Open Medicine

Evidence-Based Clinical Reasoning for AI Agents

LLMs hallucinate medical math and guidelines. Open Medicine stops them.

Open Medicine is an open-source Python library and MCP Server that provides deterministic, DOI-traceable clinical reasoning for AI agents. Every calculator, score, and guideline returns its scientific source—forcing agents to rely on verified clinical standards rather than latent knowledge.

Why Open Medicine?

If you ask an LLM to evaluate a clinical plan, it might casually agree with "aggressive fluid resuscitation" for a variceal bleed. This is a common, deadly hallucination.

By plugging the open-medicine-mcp server into your agent (via LangChain, AutoGPT, Claude Desktop, etc.), the agent can query the actual NICE CG141 Guidance and correct the plan: "Modify. Guidelines mandate a cautious, restrictive transfusion strategy (target Hgb 7-8 g/dL). Aggressive fluids will increase portal pressure."

Quick Start

1. Install the Library

Open Medicine requires Python >= 3.10. Install the library via pip. This will automatically add the open-medicine-mcp executable to your system PATH.

pip install open-medicine

2. Configure Your MCP Client

Add the open-medicine-mcp server to your MCP client's configuration file (e.g., claude_desktop_config.json for Claude Desktop):

{
  "mcpServers": {
    "open-medicine": {
      "command": "uvx",
      "args": ["--from", "open-medicine", "open-medicine-mcp"]
    }
  }
}

(This uses uvx to automatically manage the virtual environment and fetch the latest version.)

3. Test with MCP Inspector

Alternatively, you can test the toolkit using the standard MCP testing tool:

npx @modelcontextprotocol/inspector open-medicine-mcp

4. Standalone Python Library

pip install open-medicine

Deterministic Clinical Calculators

from open_medicine.mcp.calculators.chadsvasc import calculate_chadsvasc, CHADSVAScParams

result = calculate_chadsvasc(CHADSVAScParams(
    age=72,
    hypertension=True,
    diabetes=False,
    congestive_heart_failure=False,
    stroke_tia_thromboembolism=True,
    vascular_disease=False,
    female_sex=False
))

print(result.value)             # 4
print(result.interpretation)    # "CHA2DS2-VASc score is 4. High risk..."
print(result.evidence.source_doi)  # "10.1161/CIR.0000000000001193"

Renal Dose Adjustment (Multi-Drug Lookup)

from open_medicine.mcp.calculators.renal_dose_adjustment import (
    calculate_renal_dose_adjustment, RenalDoseAdjustmentParams, RenalMetric
)

result = calculate_renal_dose_adjustment(RenalDoseAdjustmentParams(
    drug_name="vancomycin",
    renal_value=25.0,
    renal_metric=RenalMetric.CRCL,
))

print(result.value["adjusted_dose"])     # "15-20 mg/kg IV q24-48h; dose by levels"
print(result.value["adjustment_type"])   # "interval_extension"
print(result.evidence.source_doi)        # "10.1093/cid/ciz895"
print(result.value["warnings"])          # ["Extended interval dosing", ...]

Guideline Retrieval

from open_medicine.mcp.guideline_engine import search_guidelines, retrieve_guideline

# Search by topic
matches = search_guidelines("atrial fibrillation anticoagulation")

# Retrieve specific section
result = retrieve_guideline("acc_aha_af_2023", "anticoagulation")
print(result.evidence.source_doi)  # "10.1161/CIR.0000000000001193"

Available Tools (MCP)

Tool	Purpose
search_clinical_calculators	Find calculators by keyword (e.g., "GI bleed")
execute_clinical_calculator	Run a calculator with JSON schema validation
search_guidelines	Find guideline sections by topic
retrieve_guideline	Retrieve curated, DOI-backed guideline content

Current Coverage

Calculators (93): AA Gradient, ABCD2, AIMS65, Anion Gap, APACHE II, Apgar, Apixaban Dosing, ASCVD, AUDIT-C, BISAP, Bishop, BMI, BSA (Mosteller), CAGE, CAM-ICU, Canadian C-Spine, Caprini, Centor/McIsaac, CHA₂DS₂-VASc, Charlson Comorbidity, Child-Pugh, CIWA-Ar, CKD-EPI, Clinical Frailty Scale, Cockcroft-Gault, Corrected Calcium, Corrected QT, Corrected Sodium, COWS, CRB-65, CURB-65, Dabigatran Dosing, DAS28, Duke Criteria, ECOG, Edoxaban Dosing, Enoxaparin Dosing, EPDS, FIB-4, Fisher Grade, 4Ts HIT, Framingham, FRAX, GAD-7, GCS, Glasgow-Blatchford, GOLD COPD, GRACE, HAS-BLED, HEART Score, Heparin Dosing, Hunt & Hess, Insulin Basal Dosing, IPSS, Karnofsky, LRINEC, Maintenance IV Fluids, MASCC, MELD-Na, MEWS, NAFLD Fibrosis, Naranjo ADR, NEWS2, NIHSS, Osmolar Gap, Ottawa Ankle, Padua, Parkland, Pediatric GCS, PERC, PEWS, PHQ-9, PSI/PORT, qSOFA, Ranson's, RASS, RCRI, Renal Dose Adjustment (20 drugs), Rivaroxaban Dosing, Rockall, Revised Trauma Score (RTS), Rumack-Matthew, Serum Osmolality, SOFA, STOP-BANG, TBSA, TIMI STEMI, TIMI UA/NSTEMI, Warfarin Initiation, Wells' DVT, Wells' PE, Winter's Formula.

Guidelines (43): Including

• ACC/AHA AF 2023 (acc_aha_af_2023)
• KDIGO CKD 2024 (kdigo_ckd_2024)
• BTS CAP 2009 (bts_cap_2009)
• TIMI UA/NSTEMI 2000 (timi_ua_nstemi_2000)
• ACC/AHA ASCVD 2013 (acc_aha_ascvd_2013)
• Sepsis-3 2016 (sepsis3_2016)
• Wells PE 2000 (wells_pe_2000)
• GOLD COPD 2024 (gold_copd_2024)
• AHA/ACC Chest Pain 2021 (aha_acc_chest_pain_2021)
• AHA/ASA Ischemic Stroke 2019 (aha_asa_stroke_2019)
• AASLD Cirrhosis 2023 (aasld_cirrhosis_2023)
• ESC ACS 2023 (esc_acs_2023)
• NICE UGIB 2012 (nice_ugib_2012)
• RCP NEWS2 2017 (rcp_news2_2017)

Design Principles

• Deterministic: Same input → same output. No LLM calls, no randomness.
• Evidence-Backed: Every ClinicalResult includes a source_doi and evidence level.
• FHIR-Compatible: Outputs include LOINC/SNOMED codes for direct integration with EHR systems.
• Strictly Typed: Pydantic models validate all clinical inputs at the boundary.

License

MIT