1. Problem and User

Clinicians, medical residents, and clinical researchers increasingly use Large Language Models (LLMs) to draft discharge summaries, triage notes, and medication reconciliation plans. However, single-pass LLM outputs frequently suffer from silent clinical hallucinations—such as failing to adjust antibiotic or antidiabetic dosages for patients with impaired renal function (low eGFR), overlooking drug-drug interactions (e.g., Warfarin + Fluconazole, Sertraline + Linezolid), or missing pregnancy contraindications.

• Target Users: Junior hospital residents, clinical pharmacists, primary care physicians in resource-constrained settings, and health-AI researchers auditing clinical LLM pipelines.

---

2. Method and Comparison to a Simple Baseline

We built MedEvidence-Checker (powered by our open-source SpecGuard-Crew verification skeleton), a spec-driven multi-agent clinical verification pipeline that decomposes unstructured clinical text into atomic claims and verifies each claim against structured clinical guidelines via the Model Context Protocol (MCP):

1. Formal EARS Specification Compiler (compiler.py): Translates clinical safety invariants written in EARS syntax (WHEN eGFR < 30 mL/min/1.73m2 AND medication is Metformin THE SYSTEM SHALL flag a critical contraindication) into deterministic verification nodes.
2. Three-Agent Verification Crew (crew_orchestrator.py):
   • Agent 1 (ClaimExtractor): Extracts patient context (labs, comorbidities, pregnancy status, allergies) and proposed interventions.
   • Agent 2 (CounterexampleHunter): Queries the local MCP Clinical Knowledge Base (tools_clinical_kb.py) covering renal dosing rules, teratogenic categories, allergy cross-reactivity, and drug-drug interaction matrices to actively search for safety violations.
   • Agent 3 (EvidenceAuditor): Synthesizes an evidence-graded audit report with guideline citations, severity levels (CRITICAL, WARNING, SAFE), and safe alternative recommendations.
3. Comparison to Simple Baseline: Unlike a Simple Single-Prompt Baseline (which checks unstructured notes in one shot without structured physiological parameter decomposition or MCP pharmacology lookups), MedEvidence-Checker enforces structured claim decomposition + deterministic MCP guideline verification + property-based invariant checks (Hypothesis).

---

3. Evaluation

Quantitative Evaluation Report: Single-Prompt Baseline vs. SpecGuard-Crew

Generated deterministically via python3 benchmarks/run_evaluation.py

1. App #1: RepoNecromancer (benchmarks/dataset_code_bugs.json, N=15)

Method	TP	FP	TN	FN	Precision	Recall (Sensitivity)	Specificity	F1-Score	False Negative Rate
Single-Prompt Baseline	2	0	5	8	100.00%	20.00%	100.00%	33.33%	80.00%
SpecGuard-Crew (Kiro + Hypothesis PBT)	10	0	5	0	100.00%	100.00%	100.00%	100.00%	0.00%

2. App #2: MedEvidence-Checker (benchmarks/dataset_clinical_cases.json, N=20, Zero-PHI)

Method	TP	FP	TN	FN	Precision	Recall (Sensitivity)	Specificity	F1-Score	False Negative Rate
Single-Prompt Baseline	1	0	8	11	100.00%	8.33%	100.00%	15.38%	91.67%
SpecGuard-Crew (Kiro + MCP Guideline KB)	12	0	8	0	100.00%	100.00%	100.00%	100.00%	0.00%

• Reproducibility: Run python3 benchmarks/run_evaluation.py and python3 -m pytest to reproduce all metrics and Property-Based Tests deterministically.

---

4. Who Benefits, and Who Could Be Harmed

• Who Benefits:
  • Medical Residents & Pharmacists: Reduces cognitive load during medication reconciliation and catches high-risk dosing errors before order sign-off.
  • Patients with Complex Polypharmacy / Renal Impairment / Pregnancy: Protects vulnerable subpopulations most frequently harmed by standard-dose LLM hallucinations.
• Who Could Be Harmed & Safety Guardrails:
  • Automation Bias Risk: Over-reliance on an automated verifier could cause a clinician to assume an unflagged note is 100% safe, even for rare hepatic/metabolic conditions outside the knowledge base.
  • Alert Fatigue Risk: Excessive low-severity warnings could desensitize clinicians.
  • Built-in Guardrails: (1) Explicit Coverage Scope Banner showing which clinical domains were checked vs. unchecked; (2) Mandatory Human-in-the-Loop Sign-off requirement; (3) Strict Zero-PHI local execution architecture with automatic SSN/identifier redaction.

---

5. Limitations and What We Would Do With One More Month

• Current Limitations: Our local MCP guideline database currently covers high-priority internal medicine rules (KDIGO renal dosing, ACOG pregnancy teratogenicity, AAAAI allergy cross-reactivity, and ISMP high-alert drug-drug interactions) on synthetic vignettes rather than full HL7 FHIR EHR integration.
• One-Month Roadmap:
  1. Integrate standard RxNorm / OpenFDA / DailyMed and SMART on FHIR connectors into the MCP server.
  2. Expand evaluation to 500+ multi-turn clinical cases validated by board-certified clinical pharmacists.
  3. Add localized multilingual patient-facing medication safety explanations.

---

6. Full Attribution & AI Tools Disclosure

• AI Tools Disclosure: We used Kiro (Specs, Steering Docs, Agent Hooks, Custom Agents, MCP, Powers) and AI coding assistants for specification drafting, scaffolding, property-based test generation (Hypothesis), and documentation. All clinical rules, evaluation logic, and safety invariants were verified via automated tests.
• Datasets & Libraries: FastAPI, React, Tailwind CSS, Hypothesis (Property-Based Testing), and 100% synthetic clinical vignettes grounded in public FDA package inserts, KDIGO, ACOG, and PubMedQA open-access medical literature.

Built With

• clinical-decision-support
• fastapi
• hypothesis
• kiro
• mcp
• multi-agent-systems
• pubmedqa
• python
• react
• tailwindcss
• typescript