README.md

metadata

title: Sundew Health Demo
sdk: gradio
sdk_version: 6.0.2
app_file: app_space_wow.py
pinned: false

Sundew Health Monitor

Neurosymbolic, energy-aware AI for continuous ECG monitoring

Combines adaptive gating, ML inference, and symbolic reasoning to deliver trustworthy, energy-efficient medical biosignal monitoring.

Key Features

• 85% energy reduction via Sundew gating algorithm
• Explainable alerts through symbolic rule chains
• Real-time streaming with WebSocket support
• Multi-patient dashboard for clinical monitoring
• Production-ready API with comprehensive SDK

Quick Start

Docker (Recommended)

docker-compose up --build

API: http://localhost:8000 | Docs: http://localhost:8000/docs

Local Development

pip install -e . --extra-index-url https://download.pytorch.org/whl/cpu
alembic upgrade head
uvicorn app.main:app --reload

Live Demo

python app_space_live.py

Architecture

ECG Stream → Sundew Gating → ML Inference → Rule Engine → Alert
              (50-90% reduction)  (PyTorch CNN)   (Symbolic)

Three-stage neurosymbolic pipeline:

1. Gating - Adaptive windowing with significance scoring
2. Inference - Energy-efficient neural network
3. Rules - Medical knowledge base for explainability

API Endpoints

• POST /ecg/infer - Run full pipeline
• WS /ws/ecg/{patient_id} - Real-time streaming
• GET /dashboard/stats - System metrics
• GET /dashboard/patients - Patient summaries
• GET /dashboard/alerts - Alert queue

SDK Usage

from sundew_sdk import SundewClient

client = SundewClient(base_url="http://localhost:8000")
result = client.infer_ecg(
    patient_id="p001",
    signal=[...],
    age=72,
    has_prior_stroke=False
)

print(f"{result['label']} | Alert: {result['alert_level']}")

See sdk/example_client.py for complete examples.

Benchmarking

Evaluate against MIT-BIH Arrhythmia Database:

bash scripts/run_full_benchmark.sh

Metrics tracked:

• Accuracy, Precision, Recall, F1
• FLOPs reduction
• Energy savings
• Latency

Results saved to benchmarks/results/

Documentation

• DEPLOYMENT.md - Production deployment guide
• LAUNCH.md - Complete feature overview
• CLAUDE.md - Architecture documentation

Technology Stack

• Backend: FastAPI, SQLAlchemy, Alembic
• ML: PyTorch, Adaptive Sparse Training
• Algorithms: Sundew (gating), symbolic reasoning
• Database: PostgreSQL (production), SQLite (dev)
• Demo: Gradio