Reinforcement Learning Insulin Pump Control

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current Artificial Pancreas systems face challenges in real-time glucose monitoring and insulin infusion due to subcutaneous measurement and absorption delays, high uncertainty from genetic, lifestyle, and environmental factors, and significant inter- and intra-patient variability, which existing control algorithms like PID, MPC, and fuzzy logic struggle to address effectively.

Innovation Solution

A model-free Actor-Critic reinforcement learning algorithm is developed with automatic and personalized tuning using insulin-to-glucose signal estimation, incorporating an exploratory policy and supervisory control to enhance adaptability and safety, allowing for real-time optimization of insulin infusion rates and carbohydrate ratios based on patient-specific characteristics.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If subcutaneous route is used for glucose measurement and insulin infusion, then the system can be implemented with current technology, but delays of up to 30 minutes for glucose measurement and 20 minutes for insulin absorption occur

Engineering Contradiction:
Improvefeasibility of subcutaneous implementationVSAvoiddelay in glucose measurement and insulin absorption
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent implements a closed-loop control system where glucose measurements and insulin delivery are continuously monitored and adjusted. The feedback mechanism compensates for the subcutaneous delays by using historical data and predictive algorithms to anticipate glucose levels and adjust insulin delivery accordingly, resolving the time delay issue while maintaining subcutaneous feasibility.

Inventive Principle:
Principle #23Feedback

2Measurement precision

If advanced sensor technology is used to provide more information to the control algorithm, then measurement accuracy improves, but the level of uncertainty remains very high due to multiple affecting factors

Engineering Contradiction:
Improveaccuracy of glucose readingsVSAvoiduncertainty from genetic, lifestyle and environmental factors
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The patent transforms the control approach by changing from traditional parameter-based control to reinforcement learning-based adaptive control. The system learns optimal control policies by interacting with the patient's glucose system, automatically adapting to genetic, lifestyle, and environmental variations without requiring manual parameter adjustment, thereby resolving the uncertainty issue while utilizing accurate sensor data.

Inventive Principle:
Principle #35Parameter changes

3Device complexity

If existing control algorithms like PID, MPC, or fuzzy logic are used, then the system structure is well-defined, but they struggle to handle uncertainties and personalisation effectively

Engineering Contradiction:
Improvestructure of control algorithmVSAvoidability to handle uncertainties and personalisation
Core Design Contradiction:
Device complexityVSAdaptability or versatility

Solution Approach 1:

The patent implements a self-learning control system using reinforcement learning that automatically adapts to individual patient characteristics and varying conditions. The algorithm learns optimal insulin delivery strategies through continuous interaction with the patient's glucose system, eliminating the need for manual personalization and enabling effective handling of uncertainties while maintaining computational efficiency.

Inventive Principle:
Principle #25Self-service

4Adaptability or versatility

If adaptive components are added to enhance algorithms for patient variability, then personalisation capability improves, but the algorithm complexity increases

Engineering Contradiction:
Improvepersonalised insulin treatment capabilityVSAvoidcomplexity of control algorithm
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent replaces complex manual adaptive adjustment mechanisms with a reinforcement learning-based automatic adaptation system. The algorithm learns patient-specific patterns and optimizes insulin delivery autonomously, achieving personalization without requiring complex manual tuning or multiple separate adaptive components, thereby reducing overall system complexity while maintaining adaptability.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Data Source

PatentEP3485409B1Estimation of insulin based on reinforcement learning
Publication Date: 2024.11.06 UNIVERSITY OF BERN
  • EP3485409B1 patent drawingFigure 1
  • EP3485409B1 patent drawingFigure 2
  • EP3485409B1 patent drawingFigure 3

AI summary

The optimal insulin to be delivered by an insulin infusion pump is determined by using a reinforcement learning algorithm aiming to the personalized glucose regulation. The algorithm optimizes the daily basal insulin rate and insulin: carbohydrate ratio for each patient, on the basis of his/her measured glucose profile. The proposed algorithm is able to learn in real-time patient-specific characteristics captured in the daily glucose profile and provide individualised insulin treatment. An automatic and personalised tuning method contributes in the optimisation of the algorithm's performance.