Hybrid Glucose Prediction Model Augmented by Physiological Simulation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing glucose prediction models, both purely data-driven and hybrid, face challenges in sensitivity to insulin and carbohydrate intakes, particularly when trained on real-world user data from narrow treatment regions, limiting their ability to generalize outside these regions, where predictions are most critical.
Innovation Solution
A method and system that generate a software-implemented module using a combination of physiological models and machine learning, where augmented data sets are created by simulating additional input parameters to enhance training data, allowing the model to capture a wider range of physiological responses and improve prediction accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If a purely data-driven machine learning model is trained on real-world user data, then the model can be implemented with a single model type, but the model lacks sensitivity to insulin and carbohydrate intakes because the training data reflects only a narrow treatment region
Solution Approach 1:
The patent merges a physiological model with a machine learning model into a hybrid system. The physiological model generates synthetic training data that covers a broader range of physiological responses to insulin and carbohydrate intakes, while the machine learning model learns from this augmented data. This combination allows the system to maintain implementation simplicity while improving sensitivity and reliability in predicting glucose responses to dietary and medication inputs.
2Reliability
If a hybrid approach combining physiological and machine learning models is used, then sensitivity to insulin and carbohydrate intakes is improved, but the system complexity increases due to maintaining two types of models
Solution Approach 1:
The physiological model is used in advance to generate synthetic training data that augments the real-world dataset. This preliminary action creates a more comprehensive training corpus that covers edge cases and broader physiological responses. The machine learning model then trains on this pre-augmented data, achieving improved sensitivity without requiring complex real-time interactions between multiple models during deployment.
Solution Approach 2:
The physiological model creates synthetic copies of training data by simulating various insulin and carbohydrate intake scenarios. These synthetic data copies expand the training dataset to include rare and extreme cases that would be difficult to capture in real-world data collection. The machine learning model learns from these copied scenarios, improving its sensitivity without requiring additional physical sensors or complex system architecture.
3Ease of manufacture
If training data is limited to real-world user data from normal treatment regions, then data collection is simple, but the model cannot generalize to critical situations outside the normal treatment region
Solution Approach 1:
The physiological model systematically varies key parameters such as insulin dosage, carbohydrate amount, and timing to generate synthetic training scenarios that cover a broader range of physiological conditions. This parameter exploration allows the machine learning model to learn responses to extreme and critical situations (e.g., severe hypoglycemia, large carbohydrate loads) without requiring actual collection of such rare real-world data, thereby improving generalization while keeping data collection simple.
Data Source
AI summary
A method for generating a software-implemented module for determining a glucose value in a body fluid. A first set of input data indicative of first values measured for first and a second input parameters is provided. A second set of input data indicative of second values for the first and second input parameters is also provided. The first and second sets of input data are processed by a physiological model to determine first and second sets of glucose values, respectively, in a body fluid. Training data is determined and a set of test data different from the training data is also determined. A software-implemented machine learning model configured to determine a glucose value in a body fluid of a patient is provided and is trained by the training data and is tested by the test data.