Model-Guided Deep Learning for Blood Glucose Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Cyber-physical control systems, such as closed loop blood glucose control systems, face inaccuracies in predictive models used for decision-making, leading to suboptimal performance due to the complexity of physical systems and the need for personalization, which existing technologies struggle to address effectively.
Innovation Solution
A model-guided deep learning approach is introduced, where a non-linear physiological model serves as a guide for a deep neural network to improve prediction accuracy by iteratively refining the model through data-driven training, specifically using a modified Bergman Minimal Model as a guide for predicting blood glucose levels in Type 1 Diabetes patients.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a traditional predictive model (Bergman Minimal Model) is used for blood glucose prediction, then the model structure is simple and easy to implement, but the prediction accuracy is insufficient and cannot accurately represent the complex physical system
Solution Approach 1:
The patent introduces a physiological model as an intermediary component between the input data and the deep neural network. This intermediary physiological model processes the input data to generate physiologically meaningful features, which are then fed into the DNN. This approach improves prediction accuracy by incorporating domain knowledge while maintaining a structured and interpretable system architecture.
2Measurement precision
If a deep neural network without guidance is used for prediction, then the model can capture complex non-linear relationships, but the prediction error is reduced only by a factor of 50 compared to model-guided approach
Solution Approach 1:
The patent applies preliminary action by using the physiological model to pre-process the input data before it enters the deep neural network. The physiological model performs initial predictions and generates intermediate features that guide the DNN's learning process. This preliminary processing reduces the training complexity and improves convergence by providing the DNN with better-initialized inputs that already incorporate physiological constraints.
Solution Approach 2:
The patent implements feedback by using the physiological model's predictions as guidance signals during the DNN training process. The physiological model provides continuous feedback on the合理性 of the DNN's predictions, allowing the system to iteratively improve accuracy. This feedback mechanism ensures that the DNN learns within physiologically plausible boundaries, reducing training complexity while improving accuracy.
3Adaptability or versatility
If personalization is achieved through data-driven approaches, then the model can adapt to individual patients, but the data requirements and computational resources increase significantly
Solution Approach 1:
The patent applies parameter changes by allowing the physiological model parameters to be personalized for each patient based on their specific characteristics. Instead of requiring large amounts of patient-specific training data, the system adjusts physiological parameters (such as insulin sensitivity, glucose production rates) to match individual patient profiles. This approach achieves personalization with minimal data requirements while maintaining physiological interpretability.
Data Source
AI summary
Systems and methods are provided for controlling predictive medical monitoring. A non-linear-predictive-guide model estimates a patient parameter that is used as a guide in a deep neural network for improving accuracy of estimation by the deep neural network. The guide model generates a guiding first estimated patient parameter based on the guide model and patient input data. The deep neural network generates a second estimated patient parameter based on the deep neural network, the patient input data, and the guiding first estimated patient parameter. The deep neural network includes an input layer that receives the guiding first estimated patient parameter, and hidden layers including respective artificial neurons configured to perform a linear or nonlinear transformation on output of at least one artificial neuron from an adjacent layer in the deep neural network. An output layer receives at least one output from a hidden layer.