README.md

metadata

title: Sundew Diabetes Commons
sdk: docker
colorFrom: green
colorTo: blue
pinned: true
emoji: 🌿
license: mit

🌿 Sundew Diabetes Watch — Advanced Edition

Mission: Deliver low-cost, energy-aware diabetes risk monitoring for everyone — with a special focus on communities across Africa.

This app demonstrates the full capabilities of Sundew’s bio-inspired adaptive algorithms, including:

• ✨ PipelineRuntime with a custom DiabetesSignificanceModel
• 📊 Real-time energy tracking with bio-inspired regeneration
• 🎯 PI-control threshold adaptation with live visualization
• 📈 Bootstrap confidence intervals for statistical validation
• 🔬 Six-factor diabetes risk computation (glycemic deviation, velocity, IOB, COB, activity, variability)
• 🤖 Ensemble model (LogReg + RandomForest + GBM)
• 💾 Telemetry export for hardware validation workflows
• 🌍 89.8% energy savings versus always-on inference (validated on real CGM data)

✅ Proven Results

Tested on 216 continuous glucose monitoring events (≈18 hours):

• Activation rate: 10.2% (22/216 events) — intelligently selective
• Energy savings: 89.8% — essential for battery-powered wearables
• Risk detection: Correctly identifies hypoglycemia (<70 mg/dL) and hyperglycemia (>180 mg/dL)
• Adaptive thresholds: PI controller dynamically adjusts from 0.10 → 0.95 based on glucose patterns

🚀 Quick Start

1. Try the demo: Sundew Diabetes Watch
2. Upload data: Use your CSV or the sample_diabetes_data.csv
3. Observe: Real-time significance scoring, threshold adaptation, and energy tracking
4. Experiment: Tweak Energy Pressure, Gate Temperature, and presets

🛠️ How It Works

1. Upload CGM data with columns: timestamp, glucose_mgdl, carbs_g, insulin_units, steps, hr
2. Custom significance model computes a multi-factor diabetes risk score
3. Sundew gating decides when to run the heavy ensemble model
4. PI control auto-adjusts thresholds to maintain target activation
5. Energy management uses bio-inspired regeneration and realistic costs
6. Statistical validation via bootstrap 95% CIs (F1, Precision, Recall)
7. Telemetry export (JSON) for power-measurement correlation

📺 Live Visualizations

• Glucose levels: Continuous CGM stream
• Significance vs. threshold: See the PI controller adapt in real time
• Energy level: Bio-inspired regeneration over time
• Risk components (×6): Interpretable breakdown of the score
• Performance dashboard: F1, Precision, Recall with confidence intervals
• Alerts: High-risk notifications

🔧 Configuration Presets

• custom_health_hd82: Healthcare-optimized (≈82% energy savings, ~0.196 recall)
• tuned_v2: Balanced general-purpose baseline
• auto_tuned: Dataset-adaptive configuration
• conservative: Maximum savings (lower activation)
• energy_saver: Battery-optimized for edge devices

Disclaimer: Research prototype. Not medical advice. Not FDA/CE approved.

💻 Developing Locally

python -m venv .venv && source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -r requirements.txt
streamlit run app_advanced.py
🧠 Technical Details
Algorithm: Sundew bio-inspired adaptive gating

Model: Ensemble (LogReg + RandomForest + GBM)

Risk factors: Six-component diabetes-specific significance model

Control: PI threshold adaptation with energy-pressure feedback

Energy model: Random regeneration (1.0–3.0 per tick) + realistic costs

Validation: Bootstrap resampling (1,000 iterations) for 95% CI

📚 References
Sundew Algorithms

Documentation

Paper (coming soon)

Built with ❤️ for underserved communities worldwide.