MedSync

MedSync - About the Project (Devpost Submission)

Inspiration

1.5 million Americans are harmed by medication errors every year. Over 7,000 die.

I learned this statistic during my research, and it stopped me in my tracks. Most of these errors happen during care transitions—when patients move from the hospital to home, or from a specialist back to their primary care doctor.

The problem isn't lack of caring. It's time pressure.

Pharmacists and nurses manually reconcile medication lists, checking for duplicates, interactions, and discrepancies across different systems. But with 10 patients waiting and only 5 minutes per chart, critical issues slip through.

I thought: What if AI could pre-screen medication lists and flag the dangerous stuff before human review?

Not to replace pharmacists—but to be a safety net. To catch the errors humans miss when they're overwhelmed.

That's why I built MedSync.

What It Does

MedSync is an AI-powered medication reconciliation assistant that analyzes patient medication lists during care transitions and flags safety issues in order of severity:

🔴 CRITICAL Issues (Immediate Action Required)

Contraindicated drug combinations: ACE inhibitor + ARNi (angioedema risk), duplicate anticoagulants, MAOIs + SSRIs
Dangerous interactions: Triple antithrombotic therapy, QT prolongation risks
Extreme dose discrepancies: >2x dose difference that could indicate transcription errors

🟡 WARNINGS (Review Required)

Duplicate medications: Brand vs generic confusion (Lipitor + Atorvastatin)
Dose discrepancies: Could be intentional titration or error—MedSync flags for review
Moderate drug interactions: Requiring monitoring or dose adjustment

🔵 INFORMATIONAL (For Awareness)

Orphan medications: Active meds without corresponding diagnoses
Missing medications: Heart failure without diuretic—intentional or gap in care?
Medications stopped during transition: Highlights changes for clinician awareness

How It Works

Gathers patient data via FHIRBridge MCP server (medications, conditions, encounters, allergies, labs)
Analyzes for safety issues using evidence-based guidelines and NLM RxNav API for drug interactions
Generates structured report with severity classification, confidence scores, and specific recommendations
Provides context-aware reasoning: Distinguishes errors from intentional changes based on encounter history

How I Built It

Architecture

┌─────────────────┐
│    MedSync      │ (Gemini AI + System Prompt)
│  Reconciliation │
│     Agent       │
└────────┬────────┘
         │ Calls FHIRBridge tools
┌────────▼────────┐
│  FHIRBridge     │ (MCP Server - 11 FHIR Tools)
│  - Medications  │
│  - Conditions   │
│  - Encounters   │
│  - Allergies    │
│  - Labs         │
└────────┬────────┘
         │ FHIR R4 API + NLM RxNav
┌────────▼────────┐
│   FHIR Server   │ (HAPI FHIR public test server)
│  + Drug DB      │
└─────────────────┘

Technology Stack

Agent Layer:

Gemini - LLM reasoning engine
Prompt Opinion Platform - A2A-compliant agent hosting
Comprehensive system prompt - 450+ lines defining reconciliation logic, severity classification, confidence scoring

Data Access Layer:

FHIRBridge MCP Server - My other hackathon project (11 FHIR R4 tools)
SHARP Extension Specs - Healthcare context propagation (patient_id, fhir_server_url, access_token)
FHIR R4 Client - Async HTTP calls to EHR systems

External APIs:

NLM RxNav - Real-time drug interaction checking
HAPI FHIR Test Server - Public FHIR endpoint for demo

Key Components

1. System Prompt (reconciliation_a2a_enhanced.md)

This is the brain of MedSync. 450+ lines that define:

Workflow: Gather Data → Analyze Issues → Generate Report
Severity classification logic: What makes something CRITICAL vs WARNING vs INFO
Confidence scoring: HIGH (90-100%), MEDIUM (60-89%), LOW (<60%)
Known contraindications: ACE-I + ARNi, duplicate anticoagulants, MAOIs + SSRIs/SNRIs
A2A collaboration protocol: How to request help from other agents
Evidence-based reasoning: Reference ACC/AHA, CHEST, and other clinical guidelines

Example logic:

#### CRITICAL (🔴) - Immediate Action Required

1. **Contraindicated Combinations**:
   - ACE inhibitor + ARNi (e.g., Lisinopril + Sacubitril/Valsartan)
   - **Confidence threshold**: Flag as CRITICAL only if >90% certain

2. FHIRBridge Integration

MedSync calls 7 primary FHIRBridge tools:

# Step 1: Gather patient context
get_patient_demographics(patient_id="maria-santos")
get_medication_list(patient_id="maria-santos", status="active")
get_medication_list(patient_id="maria-santos")  # All history
get_encounters(patient_id="maria-santos", max_results=5)
get_patient_conditions(patient_id="maria-santos", clinical_status="active")
get_allergies(patient_id="maria-santos")
get_lab_results(patient_id="maria-santos")  # If needed for monitoring

# Step 2: Check interactions
check_drug_interactions(medications=["Lisinopril", "Sacubitril/Valsartan", ...])

3. Demo Patient: Maria Santos

I created a realistic test case with planted safety issues:

68-year-old female with heart failure
Recent hospital admission (2024-06-10 to 2024-06-13)
Planted issues:
- 🔴 CRITICAL: Lisinopril (ACE-I) + Sacubitril/Valsartan (ARNi) - contraindicated
- 🟡 WARNING: Duplicate Atorvastatin entries
- 🔵 INFO: Furosemide dose increased 20mg → 40mg during admission (intentional titration)

4. A2A Capabilities

MedSync can collaborate with other agents:

Structured JSON output for programmatic consumption
Confidence scores on every finding
Agent metadata for discovery
5 invokable skills: reconcile_medications, check_medication_safety, analyze_medication_appropriateness, review_care_transition, flag_high_risk_medications

Example A2A interaction:

@PharmacyAgent: I've detected QT prolongation risk with Azithromycin +
Ondansetron in patient with eGFR 35. Can you provide risk stratification
and safer alternatives?

Challenges I Faced

1. Distinguishing Errors from Intentional Changes

The hardest part of medication reconciliation: is this discrepancy an error or on purpose?

Example: Furosemide dose changed from 20mg → 40mg.

Could be error (wrong dose prescribed)
Could be intentional (patient admitted for heart failure exacerbation, diuretic titrated up)

Solution: Context-aware reasoning

Check get_encounters for recent hospital admission
If discharge encounter + heart failure diagnosis + dose increase = likely intentional
Flag as INFO with explanation: "Dose increased during heart failure admission—likely intentional titration, but confirm with patient/clinician"

This required carefully crafting the system prompt to teach the AI how to reason about clinical context.

2. Fuzzy Medication Name Matching

Medications have multiple names:

"Lisinopril" vs "Lisinopril 10mg" vs "Lisinopril 10 MG Oral Tablet"
"Lipitor" vs "Atorvastatin" (brand vs generic)
"Aspirin" vs "ASA" vs "Acetylsalicylic Acid"

If you use exact string matching, you'll miss duplicates.

Solution: Teach the AI flexible matching in the system prompt

Be flexible with medication name matching:
- "Lisinopril" = "Lisinopril 10mg" = "Lisinopril 10 MG Oral Tablet"
- "Lipitor" = "Atorvastatin" (brand vs generic)
- Use fuzzy matching: consider stems, suffixes, brand names

Gemini's language understanding handles this surprisingly well—better than rule-based regex matching.

3. Confidence Scoring for Healthcare

In healthcare, uncertainty is dangerous. If the AI isn't sure, it needs to say so.

I added explicit confidence scoring:

HIGH (90-100%): Well-documented contraindication with guideline support
MEDIUM (60-89%): Likely issue but some data ambiguity
LOW (<60%): Insufficient data, conflicting info, edge case

Challenge: How do you teach an LLM to assess its own confidence?

Solution: Explicit scoring rubric in the prompt

Assign confidence based on:
- Complete patient data available → HIGH
- Limited encounter history → MEDIUM
- Conflicting information → LOW

**Always state confidence explicitly** for CRITICAL and WARNING items.

This makes MedSync's recommendations actionable while being humble about uncertainty.

4. Real-Time Drug Interaction Checking

I wanted real drug interaction data, not mock responses.

NLM provides a free RxNav API, but it's tricky:

Accepts medication names OR RxNorm codes (not both)
Returns different JSON structures depending on input
Rate limits
Requires fuzzy name matching

Solution: Two-phase lookup

Resolve medication names to RxNorm RxCUIs using /approximateTerm
Call /interaction/list with RxCUIs
Parse severity and descriptions

This gives MedSync real, evidence-based interaction data.

5. Creating Realistic Demo Data

Test data is usually too clean. Real EHR data is messy.

For Maria Santos, I needed:

Realistic ICD-10 codes (I11.0 for hypertensive heart disease)
Proper RxNorm codes for medications
Valid FHIR encounter structure
Planted safety issues that look plausible

Solution: Hand-crafted FHIR transaction bundle (400+ lines)

{
  "resourceType": "Bundle",
  "type": "transaction",
  "entry": [
    {
      "resource": {
        "resourceType": "Patient",
        "id": "maria-santos",
        ...
      }
    },
    // + 8 MedicationRequests, 2 Conditions, 1 Encounter
  ]
}

Uploaded to HAPI FHIR public server for live demo.

6. A2A Protocol Implementation

Prompt Opinion's A2A specs were new to me. I had to:

Define invokable skills with clear names/descriptions
Structure JSON output for agent consumption
Implement agent metadata for discovery
Design collaboration protocol (@AgentName: requests)

Solution: Study the spec, create examples, test iteratively.

The A2A-enhanced prompt grew to 450+ lines, but it's production-grade.

What I Learned

Clinical Lessons

Context is everything: The same data point (dose change) means different things in different contexts (titration vs error)
Medication reconciliation is perfect for AI: It requires cross-referencing multiple data sources (meds, conditions, encounters, allergies, labs) under time pressure—exactly what LLMs excel at
Confidence scoring matters in healthcare: "I'm 95% sure this is dangerous" is very different from "I'm 60% sure"—transparency about uncertainty builds trust

Technical Lessons

System prompts are powerful: 450 lines of carefully crafted instructions turned Gemini into a specialized medication safety expert
FHIR is complex but worth it: Standard compliance means MedSync works with any FHIR R4 server (HAPI, Epic, Cerner)—no vendor lock-in
A2A enables ecosystems: MedSync doesn't have to do everything—it can collaborate with PharmacyAgent, GuidelinesAgent, PriorAuthAgent
Demo data quality matters: Realistic test cases (like Maria Santos) make the difference between "cool demo" and "I'd actually use this"

Healthcare AI Lessons

Human-in-the-loop is essential: MedSync is an assistant, not a decision-maker. All recommendations require clinician review.
Explainability is non-negotiable: Every finding includes "Why This Matters" and "Evidence"—clinicians need to understand the reasoning
Standards compliance is table stakes: FHIR R4, SHARP, A2A—using standards makes integration possible

What's Next for MedSync

Short-term (Next 3 months)

Integrate with more guidelines databases: UpToDate, Lexicomp, ClinicalKey
Add allergy-medication checking: Cross-reference patient allergies with current meds (penicillin allergy + amoxicillin)
Renal/hepatic dose adjustments: Flag medications requiring dose adjustment based on kidney/liver function
Multi-language support: Spanish, Mandarin for patient-facing reports

Medium-term (6-12 months)

Longitudinal tracking: Monitor medication changes over time, detect gradual "medication creep"
Predictive analytics: Machine learning on historical data to predict which patients are high-risk for errors
Integration with CPOE systems: Real-time alerts when clinician prescribes contraindicated medication
Mobile app for patients: "Is this medication safe with my other drugs?"

Long-term Vision

Deploy across hospital systems: Partner with Epic/Cerner to integrate MedSync into EHR workflows
Expand to other reconciliation types: Diagnosis reconciliation, allergy reconciliation, procedure reconciliation
Multi-agent collaboration: MedSync + PharmacyAgent + GuidelinesAgent working together for comprehensive safety review
Continuous learning: Feedback loop where clinician corrections improve MedSync's accuracy

Impact Potential

If deployed across all US hospitals (35 million admissions/year):

Even a 10% reduction in medication errors would prevent:

$$ \text{Adverse Events Prevented} = 1.5M \times 0.10 = 150,000 \text{ per year} $$

$$ \text{Deaths Prevented} = 7,000 \times 0.10 = 700 \text{ per year} $$

$$ \text{Costs Saved} = \$21B \times 0.10 = \$2.1B \text{ per year} $$

But the math doesn't capture the real impact:

Families that don't lose loved ones to preventable errors
Patients who avoid hospitalization for adverse drug events
Pharmacists who can focus on complex cases instead of routine screening
Clinicians who sleep better knowing there's a safety net

Real Results: Maria Santos Demo

Testing MedSync with the Maria Santos demo patient:

Input: "Reconcile medications for patient ID maria-santos"

Output:

🔴 CRITICAL ISSUE: Contraindicated Medication Combination
Confidence: HIGH (95%)
Medications: Lisinopril 10 MG + Sacubitril/Valsartan 24-26mg
Problem: Concurrent ACE inhibitor + ARNi significantly increases angioedema risk
Why This Matters: This combination is strictly contraindicated per ACC/AHA guidelines.
                  Angioedema can cause dangerous airway swelling.
Recommendation: Discontinue Lisinopril immediately
Evidence: Both medications active in current med list. Recent hospital discharge
          for heart failure. ACE-I should have been stopped before ARNi initiation.

This error could have killed Maria Santos. MedSync caught it in 15 seconds.

Technical Innovation: Confidence Scoring + A2A

MedSync isn't just another chatbot. It's built for production healthcare with:

Confidence Scoring

Every finding includes a confidence level (HIGH/MEDIUM/LOW) based on:

Data quality (complete vs incomplete patient records)
Clinical certainty (well-documented contraindication vs edge case)
Evidence strength (guideline support vs clinical judgment)

Agent-to-Agent (A2A) Compliance

MedSync can:

Accept invocations from other agents: @MedSync.reconcile_medications(patient_id="...")
Request help from specialized agents: @PharmacyAgent: Can you provide renal dosing for...
Return structured JSON: For programmatic consumption by other agents
Expose capabilities via metadata: For agent discovery

This positions MedSync as infrastructure for an agent ecosystem, not a siloed application.

Why MedSync Matters

Healthcare is drowning in alert fatigue. EHR systems already throw hundreds of low-quality alerts that clinicians ignore.

MedSync is different:

✅ High signal-to-noise ratio: Severity classification (CRITICAL/WARNING/INFO) helps clinicians prioritize
✅ Context-aware: Distinguishes errors from intentional changes
✅ Explainable: Every finding includes "Why This Matters" and evidence
✅ Confident but humble: Explicit uncertainty when data is ambiguous
✅ Action-oriented: Specific recommendations, not just "review medications"

It's not about more alerts. It's about smarter alerts.

Try It Yourself

Prompt Opinion Marketplace: Search for "MedSync"
GitHub: github.com/jamador47/FhirBridge (includes MedSync demo)
Demo Patient: maria-santos on HAPI FHIR test server
Test Query: "Reconcile medications for patient ID maria-santos"

Acknowledgments

Prompt Opinion: For organizing the Agents Assemble Hackathon and building the A2A platform
Anthropic: For the MCP SDK
HAPI FHIR: For providing the public test server
NLM RxNav Team: For free drug interaction API
HL7 FHIR Community: For creating the interoperability standard
Pharmacists and nurses everywhere: Who inspired this by showing me what they deal with every day

Built With

Gemini - LLM reasoning engine
Prompt Opinion - A2A-compliant agent platform
FHIRBridge MCP - FHIR R4 data access (my other hackathon project)
FHIR R4 - Healthcare interoperability standard
NLM RxNav API - Drug interaction database
Python - FHIRBridge server implementation
SHARP Extension Specs - Healthcare context propagation
HAPI FHIR - Public test server

MedSync: Catching the errors that could kill your patients. Before it's too late.

Built for Agents Assemble Hackathon 2026

Built With

claude-mcp
gemini
promptopinion

MedSync - About the Project (Devpost Submission)

Inspiration

What It Does

🔴 CRITICAL Issues (Immediate Action Required)

🟡 WARNINGS (Review Required)

🔵 INFORMATIONAL (For Awareness)

How It Works

How I Built It

Architecture

Technology Stack

Key Components

Challenges I Faced

1. Distinguishing Errors from Intentional Changes

2. Fuzzy Medication Name Matching

3. Confidence Scoring for Healthcare

4. Real-Time Drug Interaction Checking

5. Creating Realistic Demo Data

6. A2A Protocol Implementation

What I Learned

Clinical Lessons

Technical Lessons

Healthcare AI Lessons

What's Next for MedSync

Short-term (Next 3 months)

Medium-term (6-12 months)

Long-term Vision

Impact Potential

Even a 10% reduction in medication errors would prevent:

Real Results: Maria Santos Demo

Technical Innovation: Confidence Scoring + A2A

Confidence Scoring

Agent-to-Agent (A2A) Compliance

Why MedSync Matters

Try It Yourself

Acknowledgments

Built With

Built With

Updates