Project–Endocrine

A Research-Grade Artificial Pancreas System (APS) Simulator

🌐 Overview

Project–Endocrine is a complete, modular, research-grade Artificial Pancreas System (APS) simulator combining:

• Physiological glucose–insulin modeling
• Adaptive PID-based insulin control
• Safety-critical logic (IOB, suspension, max-rate limits)
• Reinforcement Learning (RL) environment
• Cohort-level virtual patient testing (10–200 subjects)
• Publication-level visualization and evaluation tools

This simulator aims to provide an open, interpretable, and experiment-friendly APS research platform—ideal for:

• Biomedical engineering students
• Control systems researchers
• RL researchers
• Medical device prototyping
• Teaching and simulation environments

🧠 Background: Type-1 Diabetes and Closed-Loop Control

Type-1 Diabetes (T1D) results in absence of endogenous insulin, requiring lifelong external insulin delivery. Maintaining glucose in the target range (70–180 mg/dL) is challenging due to:

• Meal disturbances
• Sensor noise, delays, drift
• Large inter-patient variability
• Non-linear insulin action
• Uncertainty in digestion and absorption
• Noisy CGM readings

An Artificial Pancreas System (APS) automates insulin delivery using:

• A continuous glucose sensor (CGM)
• An insulin pump
• A control algorithm

Project–Endocrine models this entire loop end-to-end.

⚙️ System Architecture

       ┌───────────────────────────────┐
       │         Meal Input            │
       └──────────────┬────────────────┘
                      ↓
           ┌────────────────────┐
           │   Gut Absorption   │
           └──────────┬─────────┘
                      ↓
          ┌──────────────────────┐
          │  Glucose Dynamics    │───► CGM Sensor (Noise)
          └──────────┬───────────┘
                     ↑
     ┌──────────────────────────────────────┐
     │        Insulin Action (X)            │
     │  PK/PD + Delays + Nonlinear Effects  │
     └──────────────────┬───────────────────┘
                        ↑
                ┌───────────────┐
                │  Controller   │
                │PID / RL / APID│
                └──────┬────────┘
                       ↑
           ┌───────────────────────┐
           │  Safety Layer (IOB,   │
           │ Suspend, Rate Limits) │
           └──────────┬────────────┘
                      ↑
                Insulin Delivery

Fully modular — each block can be replaced, extended, or studied independently.

🧬 Physiological Model

The simulator includes a nonlinear glucose–insulin model:

Glucose Model

• Plasma glucose compartment
• Endogenous glucose production (EGP)
• Meal appearance via two-stage gut absorption pools
• Nonlinear insulin-dependent & insulin-independent clearance

Insulin Model

• Subcutaneous insulin absorption
• Plasma insulin with clearance
• Insulin effect (X) compartment
• Full PK/PD chain

Sensor Model (CGM)

• Gaussian noise
• Filtering
• Measurement bias (optional)

Example subject:

🎛️ Control Architecture

Adaptive PID Controller

• Proportional: responds to immediate deviation
• Integral: corrects long-term bias
• Derivative: anticipates rate-of-change
• Uses filtered CGM input
• Includes anti-windup and clamp logic

Adaptive Basal Learning

Learns patient-specific basal patterns:

Safety Layer

• Predictive low-glucose suspend
• Negative-trend veto
• Max insulin rate limiter
• Insulin-On-Board (IOB) constraints
• Hard safety clamps

🤖 Reinforcement Learning

Project–Endocrine exposes a Gym-style RL environment:

• reset(), step() API
• Full state vector extraction
• Reward functions (TIR-weighted, risk index, smoothness)
• Multi-patient training
• Deterministic and stochastic modes

Supports algorithms such as:

• PPO
• SAC
• TD3
• Model-Based RL

👥 Cohort Simulation (10–200 Subjects)

Each virtual subject varies in:

• insulin sensitivity
• carb absorption rate
• glucose clearance
• insulin clearance
• basal metabolic drift
• noise profile

Cohort Mean ± STD

Sample traces (first 6 subjects)

Full cohort (200 subjects)

📊 Cohort Performance Metrics

TIR distribution (200 subjects)

Subject-wise TIR

Glycemic Variability

Mean Glucose Distribution

Mean Glucose vs CV%

TIR / TAR / TBR Boxplot

🎥 Simulation Demo

download.mp4

🚀 Installation

git clone https://github.com/HeroicKrishna160905/Project-Endocrine.git
cd Project-Endocrine
pip install -r requirements.txt
Or [launch in Google Colab](https://colab.research.google.com/) (recommended).

## 🧪 Quickstart

### Single Patient Simulation
```python
from endocrine.simulator import APS_Simulator

sim = APS_Simulator()
results = sim.run()
sim.plot()

Cohort Evaluation

from endocrine.evaluation import run_cohort

stats, df = run_cohort(n_subjects=200)
print(stats)
df.head()

🧩 Project Structure

Project-Endocrine/
│
├── endocrine/
│   ├── physiology/      # PK/PD model, glucose, insulin
│   ├── controller/      # PID, adaptive basal, safety
│   ├── rl_env/          # Gym-style RL environment
│   ├── evaluation/      # Cohort simulation + metrics
│   └── utils/
│
├── notebooks/
│   ├── v6_3_realism.ipynb
│   └── v6_3_RL.ipynb
│
├── assets/              # Figures, demo video
├── results/             # Saved outputs
├── README.md
└── LICENSE

🧾 Key Contributions

Project–Endocrine provides:

• ✔ A fully custom glucose–insulin PK/PD model (Not dependent on UVA/Padova or simglucose)
• ✔ A modular adaptive PID controller Designed for interpretability & robustness.
• ✔ A complete safety layer implementation (LOWS, IOB, rate limits, rapid-drop veto)
• ✔ A Gym-compatible RL environment Allowing direct integration with PPO/SAC/TD3.
• ✔ Realistic cohort simulation 200-subject evaluation with variability.
• ✔ Publication-grade visualizations All included as reproducible notebooks.
• ✔ Academic grounding Supported by nonlinear control literature — e.g., observer-based LMI designs from our professor's paper (2019).

📚 References

Key Literature

• UVA/Padova T1D Simulator
• GluCoEnv: High-performance APS RL environment
• Control-IQ / DiAs APS systems
• PID/MPC control for APS
• Reinforcement Learning for glucose control

Our Professor's Related Work

Observer-based nonlinear control for glycemic regulation, using LMI-stabilized designs:

• 2019 – "Observer based nonlinear control: An LMI approach." (Full text provided internally)

📝 License

MIT License Open for academic, research, and personal use.

🤝 Contributions

Pull requests, ideas, and discussions are welcome.

❤️ Acknowledgments

• Inspired by global diabetes research.
• Built with guidance from professors and real control theory literature.
• Developed with attention to openness, reproducibility, and scientific rigor.