Adaptive MPC Controller for Artificial Pancreas Using Trust Index
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current Model Predictive Control (MPC) algorithms in artificial pancreas systems face inaccuracies in glucose predictions, leading to uncertainty in insulin dosing, which can result in both hyperglycemia and hypoglycemia, necessitating a personalized and adaptive control approach to maintain safe glucose levels.
Innovation Solution
An adaptive MPC controller is developed that calculates a trust index based on prediction residuals, allowing real-time tuning of its aggressiveness by increasing or decreasing insulin delivery, using a forgetting function to penalize residuals and determine confidence intervals, thereby optimizing insulin dosing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a standard MPC controller is used with fixed tuning parameters, then the device complexity is low and ease of operation is high, but the reliability is poor due to inaccurate glucose predictions and inability to adapt to individual patients
Solution Approach 1:
The patent implements dynamic tuning of MPC controller parameters by continuously calculating prediction residuals and adjusting the aggressiveness parameter in real-time. The controller transitions from static to dynamic behavior by adapting its control strategy based on the accuracy of its predictions, using a forgetting function to weigh recent residuals more heavily. This dynamic adaptation improves reliability without requiring a completely new controller architecture.
Solution Approach 2:
The patent introduces a feedback mechanism where prediction residuals (the difference between predicted and actual glucose values) are continuously calculated and fed back into the controller. This feedback loop allows the system to learn from its prediction errors and adjust its aggressiveness parameter accordingly, improving glucose prediction accuracy over time while maintaining a relatively simple controller structure.
2Productivity
If the MPC controller increases aggressiveness to correct hyperglycemia quickly, then the productivity of glucose correction is improved, but the object-generated harmful factors increase due to risk of hypoglycemia
Solution Approach 1:
The patent dynamically adjusts the aggressiveness parameter based on the calculated trust index from prediction residuals. When predictions are accurate (high trust index), the controller can be more aggressive in correcting hyperglycemia. When predictions are less accurate (low trust index), the controller becomes more conservative. This dynamic adjustment optimizes correction speed while minimizing hypoglycemia risk.
Solution Approach 2:
The patent changes the aggressiveness parameter (a key control parameter) based on the trust index derived from prediction residuals. By modifying this parameter in real-time, the controller can adapt its response intensity to match the reliability of its predictions, achieving fast correction when safe and conservative control when uncertain, thus balancing productivity and safety.
3Reliability
If the MPC controller uses conservative tuning to avoid hypoglycemia, then the reliability is improved, but the productivity of glucose control decreases leading to prolonged hyperglycemia
Solution Approach 1:
The patent implements a dynamic tuning strategy where the controller aggressiveness is adjusted based on the trust index. This allows the system to switch between conservative and aggressive modes as needed, rather than being locked into a fixed conservative setting. The dynamic nature enables fast correction when conditions permit while maintaining safety when uncertainty is high.
Solution Approach 2:
The patent modifies the aggressiveness parameter based on the trust index calculated from prediction residuals. When the trust index is high (indicating reliable predictions), the parameter allows more aggressive control for faster glucose normalization. When the trust index is low, the parameter becomes more conservative for safety. This parameter adaptation resolves the trade-off between reliability and productivity.
4Adaptability or versatility
If the controller continuously adapts tuning parameters based on prediction residuals, then the adaptability is improved for personalized control, but the device complexity increases due to additional calculations and memory requirements
Solution Approach 1:
The patent implements a lightweight dynamic adaptation mechanism that calculates prediction residuals and updates the aggressiveness parameter in real-time. The adaptation logic is kept simple by using a forgetting function with a single decay parameter and a straightforward trust index calculation. This dynamic personalization achieves high adaptability while maintaining relatively low computational complexity suitable for embedded systems.
Solution Approach 2:
The patent focuses adaptation efforts on tuning a single key parameter (aggressiveness) based on prediction residuals, rather than optimizing multiple parameters simultaneously. This single-parameter adaptation strategy achieves personalized control with minimal computational overhead, balancing adaptability and device complexity effectively.
Data Source
Figure 1
Figure 2~3
Figure 4
AI summary
Presented herein are methods, and devices of adaptively tuning a zone based Model Predictive Control (MPC) controller, using at least one processor, which include determining, residuals based on prediction models storing, in a memory, the determined residuals, calculating a trust index by quantifying uncertainty of the prediction models using the stored residuals and tuning the MPC controller, in real time based on the calculated value of the trust index.