ECG Agentic AI — CardioClinician-Agent Baseline

A multi-agent AI system for cardiological ECG analysis, evaluated on the ECG-Expert-QA benchmark.

Architecture

User Query ──► Chief Coordinator (Router)
                    │
        ┌───────────┼───────────┬──────────────┐
        ▼           ▼           ▼              ▼
   Inquiry     ECG Interp.  Diagnostic    Prognosis
   Agent       Agent        Reasoner      Agent
        └───────────┼───────────┴──────────────┘
                    ▼
            Ethics & Safety Guardian  ◄── parallel review
                    │
                    ▼
              Final Synthesis

Agents (LangGraph state-machine orchestration):

• Inquiry Agent – multi-round patient dialogue, memory correction
• ECG Interpreter Agent – cross-modal ECG ↔ text diagnosis
• Diagnostic Reasoner Agent – complex/long-text diagnosis
• Prognosis Agent – prognostic analysis, treatment planning
• Ethics & Safety Guardian – patient rights, risk assessment, ethical review
• Chief Coordinator – routes tasks, synthesises final answers, revision loop

Tools: ECG waveform parser, CHA₂DS₂-VASc / GRACE risk calculators, AHA/ESC guideline lookup.

Baseline Results (gpt-4o-mini, n=20 per task)

Task	Description	ROUGE-1	ROUGE-2	ROUGE-L	BERTScore F1
EBK	ECG Basic Knowledge Q&A	0.1254	0.0700	0.1080	0.0664
CK	Cardiology Knowledge Q&A	0.1561	0.0713	0.1045	0.0570
CMD	Cross-Modal Diagnosis	0.1929	0.0620	0.1124	-0.0186
CD	Complex Diagnosis	0.1017	0.0480	0.0776	0.0751
MROD	Multi-Round Dialogue	0.2100	0.0812	0.1292	0.0861
MC	Memory Correction	0.3919	0.1405	0.1917	-0.0225
LTD	Long-Text Diagnosis	0.3281	0.1308	0.1714	0.1011
MEE	Medical Entity Extraction	0.2060	0.0799	0.1273	-0.2972
PP	Patient Prognosis	0.1212	0.0331	0.0806	0.0434
PRTK	Patient's Right to Know	0.2226	0.1219	0.1842	0.2890
GRA	Generate Risk Assessment	0.4865	0.2033	0.2353	0.0015

Project Structure

src/
├── config.py                 # Central configuration
├── run_baseline.py           # Main entry point
├── dataio/
│   ├── __init__.py
│   └── loader.py             # Unified loader for 11 ECG-Expert-QA tasks
├── agents/
│   ├── __init__.py
│   ├── base.py               # LLM backend factory (OpenAI / HuggingFace)
│   ├── prompts.py            # System prompts per agent
│   └── coordinator.py        # LangGraph state-machine orchestrator
├── tools/
│   ├── __init__.py
│   ├── ecg_parser.py         # ECG parameter → clinical findings
│   ├── risk_calculators.py   # CHA₂DS₂-VASc, GRACE score
│   └── guideline_lookup.py   # AHA/ESC guideline retrieval
└── evaluation/
    ├── __init__.py
    ├── metrics.py            # ROUGE, BERTScore, Entity F1
    └── evaluator.py          # Task-level evaluation driver

Dependencies

• LangGraph / LangChain – agent orchestration
• OpenAI – LLM backend (default: gpt-4o-mini)
• Transformers / PyTorch – HuggingFace local model fallback
• rouge-score, bert-score, nltk – evaluation metrics
• NumPy, Pandas, SciPy – data processing
• WFDB – ECG signal processing

Installation

# Create and activate conda environment
conda create -n ecg-agentic-ai python=3.12
conda activate ecg-agentic-ai
poetry install

Usage

# Set your OpenAI API key (or add to .env file)
export OPENAI_API_KEY=sk-...

# Run evaluation on all tasks (default 20 samples per task)
MAX_SAMPLES=20 python -m src.run_baseline

# Run specific tasks
python -m src.run_baseline --tasks EBK CK CMD

# Quick smoke test
MAX_SAMPLES=3 python -m src.run_baseline --tasks EBK --verbose

# Use HuggingFace local model
LLM_BACKEND=huggingface python -m src.run_baseline

Results are written to results/ as JSON files and a summary table.