Multi-specialty clinical decision support for complex chronic disease patients. Consilium reduces guideline conflict, medication-safety risk, and physician decision fatigue by combining specialist LLM agents with deterministic clinical ranking.
Originally developed for an agent interoperability challenge, Consilium is designed as a general-purpose clinical AI system that can run as a deployed A2A agent, consume FHIR context when available, and support standalone clinical demos.
Patients with heart failure, diabetes, and chronic kidney disease often receive recommendations from multiple specialists whose guidelines collide:
- Cardiology may prioritize guideline-directed HF therapy.
- Nephrology may prioritize renal dosing, potassium, and eGFR thresholds.
- Endocrinology may prioritize glycemic control and cardiometabolic benefit.
The primary care physician is left to reconcile tradeoffs manually. Consilium turns that reconciliation into an explainable, auditable workflow: specialist recommendations are generated separately, validated structurally, and ranked with deterministic multi-criteria scoring.
Consilium exposes one A2A endpoint that can be connected to Prompt Opinion. At runtime it:
- Reads patient context from Prompt Opinion FHIR metadata when available, or from clinician-entered text when not.
- Calls three ADK specialist agents: cardiology, nephrology, and endocrinology.
- Requires each specialist to return structured JSON:
specialty,recommendation,risks, andcitation. - Converts those outputs into deterministic TOPSIS inputs.
- Returns a ranked clinical decision, action plan, conflicts resolved, citations, and a safety disclaimer.
If the text input lacks enough patient-specific information, Consilium refuses to invent a plan and asks for more clinical context.
Live frontend demo: https://asm-healthcare-agent.vercel.app
Deployed A2A backend: https://consilium-1085209557278.us-central1.run.app
Public agent card: https://consilium-1085209557278.us-central1.run.app/.well-known/agent-card.json
| Layer | Current Implementation |
|---|---|
| A2A Backend | Cloud Run service serving orchestrator.app:a2a_app |
| Platform | Prompt Opinion BYO A2A agent connection |
| Patient Context | Prompt Opinion FHIR context extension + SMART scopes |
| Specialist Models | DeepSeek V4 Flash through LiteLLM |
| Frontend Demo | React/Vite clinical workspace deployed on Vercel |
| Local Fallback | Direct free-text patient summary when no FHIR context is present |
The agent card advertises the official Prompt Opinion FHIR extension:
https://app.promptopinion.ai/schemas/a2a/v1/fhir-context
It also advertises five concrete A2A skills so the marketplace entry reflects the real workflow:
Multi-Specialty OrchestrationFHIR Clinical Context SummarySpecialist Consult PanelMedication Safety ReconciliationTOPSIS Clinical Ranking
Prompt Opinion / React Demo
|
| A2A message/send
v
Consilium Orchestrator
|
| FHIR context if available
v
FHIR patient summary
|
| ADK Runner calls
+---------------------+----------------------+----------------------+
v v v
Cardiology Agent Nephrology Agent Endocrinology Agent
ACC/AHA HF KDIGO CKD ADA Diabetes
| | |
+---------- structured JSON recommendations -+
|
v
Deterministic TOPSIS scorer
|
v
Ranked decision + conflicts resolved
This version implements parallel specialist consult + deterministic reconciliation. The agents do not yet conduct a multi-round negotiation with each other. That is intentional for the current release: it keeps the decision path inspectable, reliable, and fast enough for live A2A demos. A future version should add a negotiation loop where specialists exchange rationale and revise recommendations before TOPSIS ranking.
The LLM specialists do not assign TOPSIS scores. They only produce recommendations, risks, and citations. Consilium computes scoring dimensions in code.
| Dimension | Default Weight | How It Is Computed |
|---|---|---|
| Evidence Level | 0.30 | Deterministic mapping from guideline labels such as ACC/AHA Class I, KDIGO Grade 1A, ADA Level A |
| Patient Match | 0.30 | Deterministic match to patient state: LVEF, eGFR, HbA1c, HF, CKD, diabetes |
| Medication Safety | 0.20 | Deterministic penalty from risk flags such as hyperkalemia, lactic acidosis, hypotension, volume depletion |
| Guideline Priority | 0.20 | Deterministic guideline-strength proxy using the same evidence hierarchy |
Dynamic patient-state weighting is implemented for safety-critical cases:
| Trigger | Weights |
|---|---|
| eGFR <30 and patient is on Metformin | Evidence 0.20, Patient Match 0.25, Medication Safety 0.35, Guideline Priority 0.20 |
| LVEF <35 without the renal-safety override | Evidence 0.25, Patient Match 0.25, Medication Safety 0.20, Guideline Priority 0.30 |
The renal-safety override takes precedence because eGFR <30 + Metformin is an immediate medication-safety conflict.
Input: 68M, HFrEF LVEF 32%, CKD stage 4 eGFR 28, T2DM HbA1c 8.2%, on Lisinopril, Metformin, Furosemide.
| Rank | Specialty | Score | Why |
|---|---|---|---|
| 1 | Nephrology | 0.900 | eGFR <30 makes Metformin safety the highest-priority conflict |
| 2 | Endocrinology | 0.625 | Diabetes therapy must change, also supports SGLT2i |
| 3 | Cardiology | 0.350 | HFrEF therapy matters, but unresolved renal safety is more urgent |
Conflict resolved: Stop Metformin; start SGLT2i if tolerated; continue ACEi/ARB with potassium and creatinine monitoring.
Input: 55F, HFpEF LVEF 58%, eGFR 82, no diabetes, no kidney disease, on Lisinopril and Carvedilol.
| Rank | Specialty | Score | Why |
|---|---|---|---|
| 1 | Cardiology | 0.900 | HF phenotype drives the only active specialty priority |
| 2 | Endocrinology | 0.625 | No glucose-lowering therapy is indicated |
| 3 | Nephrology | 0.350 | No CKD-specific medication conflict detected |
Conflict resolved: No Metformin or CKD conflict is invented; the output stays scoped to the available patient data.
Prompt Opinion demo prompt: Run the full multi-specialty orchestration for this patient: elderly male with diabetes.
Output: Consilium returns More Patient Information Needed and does not call specialist agents or TOPSIS. This prevents the model from hallucinating a clinical plan when neither FHIR context nor a meaningful patient summary is available.
- FHIR-aware: reads
fhirUrl,fhirToken, andpatientIdfrom A2A metadata when Prompt Opinion sends FHIR context. - No token leakage: production defaults do not log full JSON-RPC bodies, FHIR tokens, patient ids, or full FHIR payloads.
- Minimum-context gate: insufficient free-text requests are rejected before specialist LLM calls.
- Specialist contract validation: malformed JSON, missing required fields, wrong specialty, or empty recommendations are rejected.
- Per-specialist fallback: one failed or timed-out specialist falls back deterministically without failing the whole task.
- A2A compatibility: middleware accepts
SendMessage,SendStreamingMessage,ROLE_USER,ROLE_AGENT,messageId, andmessage_idvariants. - CORS-ready: browser demos can call the deployed A2A service.
Current scope: heart failure, chronic kidney disease, diabetes, medication conflict reconciliation, and advisory clinical decision support.
Out of scope: emergency diagnosis, autonomous prescribing, replacing physician judgment, and claims of validated clinical outcome improvement.
| Component | Technology |
|---|---|
| Agent Framework | Google ADK |
| A2A Transport | A2A JSON-RPC endpoint |
| Platform | Prompt Opinion |
| LLM Provider | DeepSeek V4 Flash via LiteLLM |
| Decision Engine | Custom TOPSIS implementation |
| Clinical Data | FHIR R4 + SMART scopes |
| Backend Runtime | FastAPI/Starlette through ADK A2A adapter |
| Frontend | React + Vite + lucide-react |
| Deployment | Google Cloud Run |
git clone https://github.com/eiddiedev/Consilium.git
cd Consilium
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env
# Set DEEPSEEK_API_KEY and API_KEYS.
# Defaults use deepseek/deepseek-v4-flash via LiteLLM.
uvicorn orchestrator.app:a2a_app --host 0.0.0.0 --port 8003
curl http://localhost:8003/.well-known/agent-card.jsonsource .env
curl -sS -X POST http://localhost:8003/ \
-H 'Content-Type: application/json' \
-H "X-API-Key: ${API_KEYS%%,*}" \
-d '{
"jsonrpc": "2.0",
"id": "demo",
"method": "message/send",
"params": {
"message": {
"role": "user",
"messageId": "demo-message",
"parts": [{
"kind": "text",
"text": "68-year-old male with HFrEF LVEF 32%, type 2 diabetes HbA1c 8.2%, CKD eGFR 28. Current medications: Lisinopril, Metformin, Furosemide, Aspirin, Glipizide."
}]
}
}
}'cd frontend
cp .env.example .env
# Set VITE_A2A_AGENT_URL and VITE_A2A_API_KEY.
npm install
npm run dev# Backend unit/regression tests
.venv/bin/python -m pytest -q
# Frontend build
cd frontend && npm run build
# Agent card
curl -sS http://localhost:8003/.well-known/agent-card.json
# A2A smoke
curl -sS -X POST http://localhost:8003/ \
-H 'Content-Type: application/json' \
-H "X-API-Key: ${API_KEYS%%,*}" \
-d '{"jsonrpc":"2.0","id":"smoke","method":"message/send","params":{"message":{"role":"user","messageId":"smoke-message","parts":[{"kind":"text","text":"55-year-old female with HFpEF LVEF 58%, eGFR 82, no diabetes, no kidney disease. Current medications: Lisinopril and Carvedilol."}]}}}'Expected:
- Tests pass.
- Agent card includes
securitySchemes,supportedInterfaces, Prompt Opinion FHIR extension URI, and SMART scopes. - A2A response returns a completed task artifact.
- Logs show only safe FHIR status summaries unless
LOG_FHIR_DEBUG=trueis explicitly set.
Consilium/
├── orchestrator/ # Main A2A endpoint and orchestration logic
├── cardiology_agent/ # ACC/AHA-oriented specialist agent
├── nephrology_agent/ # KDIGO-oriented specialist agent
├── endocrinology_agent/ # ADA-oriented specialist agent
├── shared/ # A2A app factory, middleware, FHIR hooks/tools
├── tools/ # TOPSIS scorer and helper tools
├── frontend/ # React demo UI
├── data/ # Example FHIR bundles and guideline weights
└── tests/ # TOPSIS, parser, safety, and regression tests
- Current specialist agents run in parallel and do not yet negotiate with each other in a multi-turn deliberation loop.
- The system has not been validated against clinical outcomes or physician time-motion studies.
- Guideline mappings are intentionally narrow in the current release: HF, CKD, and diabetes medication conflicts.
- This is advisory clinical decision support, not autonomous prescribing software.
- ACC/AHA 2022 Heart Failure Guideline: https://www.jacc.org/doi/10.1016/j.jacc.2021.12.012
- KDIGO 2024 CKD Guideline: https://kdigo.org/guidelines/ckd-evaluation-and-management/
- ADA 2025 Standards of Care in Diabetes: https://diabetesjournals.org/care/issue/48/Supplement_1
Consilium is a clinical decision support tool. It does not replace physician judgment, local policy, patient preference, or emergency clinical assessment. Final treatment decisions rest with the treating clinician.
MIT