2. Background
● More than 100 million adults in U.S. alone living with diabetes
● This condition results from high blood glucose level due to inadequate natural insulin
(Type 2) or obsolete (Type 1) in the body
● Treatment involves maintaining healthy blood glucose level at all time by injecting
appropriate amount of synthetic insulin at appropriate time
● Two types of insulin are used
○ basal for continuous blood glucose control
○ bolus as a short acting insulin with meal disturbances
3. CGM and Insulin Pump for Diabetic Control
Key steps in CGM and Insulin Pump based blood
glucose control
● Continuous glucose level monitoring with glucose
sensor
● Algorithm to determine amount and type of insulin to
be delivered at specific time
● Delivering insulin into body through insulin pump
It is a complex problem to keep a healthy level of
blood glucose at all the time!!!
4. Maintaining Appropriate Blood Glucose Level
● Effective insulin control for diabetic Type-1 patient requires that a healthy-level of
glucose is maintained throughout the day with minimal fluctuations in either direction
○ Inadequate insulin causes high blood sugar (Hyperglycemia) resulting in longer term
complication
○ Excessive insulin leads to low blood sugar (Hypoglycemia) which is often fatal if is too low
● Goal of insulin dependent diabetic care is to administer appropriate amount of insulin
at appropriate time such that glucose level is maintained at near target level without
reaching hypo or hyper level
● Maintenance of “right” amount of blood glucose is a very complex problem because of
lot of day-to-day variability in patient’s condition
○ diet/nutrition changes, exercise amount, exposure to sun, daily life-style etc.
5. Closed Loop Control System for Optimal Insulin
● Interaction between Glucose and administered Insulin is a closed loop control system
with feedback
○ Typically insulin control modeled as a PID controller to determine optimal insulin that
needs to be administered
● Such model driven approaches have limitations due to complex nature induced by time
varying, non-linear conditions
● Instead we approach this as an AI problem - specifically we consider Reinforcement
Learning model for insulin controller and evaluate its efficacy
6. Insulin Control as a Reinforcement Learning Model
● We propose Reinforcement Learning based
insulin control where
○ an agent repeatedly interacts with the environment,
each time receiving feedback (reward) for its actions
○ goal of the agent is to learn an optimal policy that
maximizes this feedback (reward) in the long run
● Specifically we propose and evaluate DDPG - a
deep RL framework - for insulin controller
○ suitable for both continuous action and continuous
space
○ Allows us to explicitly account for both basal and
bolus glucose regulation
7. DDPG Background and Why DDPG
DDPG Background
● Is an off-policy algorithm
● Works with environments with continuous
action spaces
● Is similar to deep Q-learning for continuous
action spaces
● Employs Actor-Critic model
● Learns directly from the observation spaces
through policy gradient method
DDPG as insulin Controller
● Environment: patient-glucose insulin
interaction
● Sate Space: glucose level and meal amount
● Action Space: insulin amount at each time step
which is in a continuous space
8. DDPG Formulation
- It learns a Q-function and policy:
Q-function:
Policy:
-It uses off-policy data and the Bellman equation to learn the Q-function, and uses the Q-function to learn the policy.
Bellman equation:
Deterministic Gradient Policy (Off-policy):
10. Implementation and Environment
● We evaluate performance of DDPG based insulin controller using synthetic data
generated with well known UVA/Padova model that models human glucose level
● It includes models for 30 virtual patients, 10 adults, 10 adolescents, 10 children
● DDPG insulin controller implemented in Python using OpenAI Gym framework by
extending SimGlucose simulator
● Simulator parameters consists of the following features
○ Meal frequency and amount
○ Patient age, weight, height
11. Evaluation Framework
● DDPG based Insulin controller vs baseline model based controller (BBController)
to study how well DDPG controller performs under varying/dynamic conditions
● We evaluate under three different scenarios by inducing disturbances in blood
glucose through
○ single meal
○ multiple meals taken frequently in short intervals
○ multiple meals spread across longer intervals
● Rationale behind these choices is to understand if DDPG based controller has any
advantage over typical model based insulin controller under such induced
disturbances in glucose level
12. Experiment I: Single Meal Disturbance
Disturbance with a single meal of size 30 (CHO value 10) introduced at 7:00am.
● Glucose level rose much higher with BBController before stabilizing to normal range whereas blood glucose fluctuation is
relatively lower with DDPG Controller.
● Achieves tighter control with DDPG due to continuous adjustment based on the environment with purely data driven
approach instead of preset model as with BBController.
BBController DDPG Controller
13. Experiment II: Multi-Meal Frequent Disturbances
Frequent Multi-meal Disturbance where three meals were taken with shorter gaps between meals
● Blood glucose rose to higher level (>250) for longer duration of interval with BBController and did not react fast
enough after each disturbance as compared to DDPG based controller
● DDPG controller maintained glucose with relatively less fluctuations and closer to target level
● DDPG based controller more suitable for handling rapid and dynamic blood glucose disturbances even alone with just
basal insulin due to continuous reactive nature of the controller
BBController DDPG Controller
14. Experiment III: Multi-Meal Disturbances Spread
Out
Spread Out Multi-meal Disturbance: This scenario evaluates performance with multiple meal disturbances with longer gaps
between meals
● BBController is able to handle this spread out meal disturbances relatively better compared to when disturbances are
more rapid and dynamic.
● We find that both model based and DDPG controller with basal insulin have similar performance under this scenario. We
need to further investigate if DDPG controller could achieve superior performance when bolus insulin is also combined.
BBController
15. Conclusion
● Our study is first to propose and evaluate deep RL based insulin controller for Type-1
diabetic blood glucose management
● We implement DDPG based controller using SimGlucose that supports Open AI Gym
framework
● We compare with model based baseline insulin controller (Padova) under three
scenarios
○ Single meal
○ Frequent Multiple meal disturbance
○ Spread out multiple meal disturbance
● Our evaluation indicates that DDPG based controllers are more suitable to handling
rapid and dynamic blood sugar disturbance conditions compared to model based
controller
● DDPG controller is able to achieve more controlled blood glucose level even alone with
basal insulin, possibly due to the continuous reactive nature of such algorithms
16. Future work
● Results from our initial evaluations are promising that deep RL based insulin
controllers could be more effective in handling rapid fluctuations in blood sugar
● Further evaluate the effectiveness of DDPG based insulin controller by varying
environment variables and additional physical factors such as exercise, life-style
changes
● Incorporate bolus insulin into the DDPG controller action to further improve its
efficacy in reducing blood sugar fluctuations
17. Acknowledgement
Jinyu Xie, SimGlucose creator
Yuan Zhao, Mobiquity Networks
Dr. Chong Li, Columbia University
InsightZen for infrastructure support.