/T1DMPC

Primary LanguageJupyter NotebookMIT LicenseMIT

T1DMPC

A study in blood glucose control for Type 1 Diabetes Mellitus patients

Control Examples

Ideal System

System with Noise

Average measurement Response

Literature review and Models

Literature

Reinforcement Learning in Diabetes Blood Glucose Control

  • General framework includes a Glucose sensor which measures every 5 minutes, a control algorithm with no information other than the measurements, and supplied dosage of insulin.
  • A broader scope could include information such as food intake, physical activity, infections, and stress level.
  • RL does not require labeled training data, but rather looks at cause and effect
    • have to set correct cost function for RL to accurately reflect desires
    • Performance Metrics included Acceptable ranges, , Control Variability grid Analysis, meal disturbance rejection
  • AC Learning and Q-Learning were the most popular RL algorithms

Engineering The Artificial Pancreas

  • Challenges for Feedback control
    • Sensor lag is about 5-15 minutes
    • CGM systems have an absolute error of 13-16%
    • Rapidly Responsind Insulin takes about 60-90 Minutes (use two types?)
    • Time Variation fo Human body:
      • Absorbtion Time varies 20-35% (by site?)
      • Insulin sensitivity changes as much as 50% through the day
    • Overall Error is 24% to 37% with a lag of 70-110 minutes

Models

  • Time Series Numerical model
  • Nonlinear Model with No delay
    $$\begin{array}{l} \dot{G}(t)=-P_{G}\left[G(t)-G_{0}+\varepsilon(t)\right]-\mathrm{SI}(t)\left[G(t) Q(t)-G_{0} Q_{0}\right]+K_{2} P_{S}(t) \ \dot{I}(t)=-\left(n_{T}+n_{I}\right) I(t)+n_{I} Q(t)+\frac{K_{X} U_{S}(t)}{V_{P}} \ \dot{Q}(t)=-\left(n_{I}+n_{C}\right) Q(t)+n_{I} I(t) \ \dot{P}{S}(t)=\frac{P{X}(t)+P_{C}(t)}{V_{G}}-K_{1} P_{S}(t) \ \dot{U}{S}(t)=K{X}\left[U_{X}(t)-U_{S}(t)\right] \end{array}$$

Project Scope

  • over different situations, and disturbances, how does the controller respond?
    • give arrays for disturbances of food intake, desired lebels, base metabolic rate, and observe objective function after control.
  • Are we varying the Model or the control method?
  • how do we simulate the acceptable glucaose range? An interpolated function?
  • How are results generalized?

Simple Project Ideas

  • Focus on safety
  • Add units of glucose
  • ML Recognition of disturbances
  • structuring the oobjective functio for optimal results
  • Two types of insulin
  • more accurate prediction and parameters
  • Stochacity Modeling