Machine Learning Glucose Prediction for Longer Forecast Horizons
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional glucose prediction techniques suffer from inaccuracies and delayed predictive horizons, rendering them unsuitable for timely intervention in diabetes management, particularly for predicting glucose levels further into the future.
Innovation Solution
A non-linear machine learning model, such as a recurrent neural network, is trained using historical time series glucose measurements to accurately predict upcoming glucose levels, leveraging data from wearable glucose monitoring devices, enabling more precise and timely predictions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional glucose prediction techniques are used, then the system can provide glucose level predictions, but the predictions are inaccurate and have delayed predictive horizons that do not match actual glucose levels
Solution Approach 1:
The patent transforms glucose prediction from conventional static correlation methods to dynamic machine learning models that learn temporal patterns and state transitions. The system changes the predictive parameters by using hidden Markov models and recurrent neural networks to capture the dynamic nature of glucose metabolism, allowing accurate prediction of both glucose levels and their timing without fixed delay assumptions
Solution Approach 2:
The patent replaces conventional mechanical correlation-based prediction systems with intelligent machine learning systems. Instead of using fixed mathematical correlations between observed and predicted glucose levels, the system employs trained models (HMMs, RNNs, LSTMs) that automatically learn and adapt to individual patient patterns, substituting rigid mechanical prediction mechanisms with adaptive intelligent systems
2Measurement precision
If machine learning models are trained on large amounts of historical glucose data, then prediction accuracy improves, but the complexity of processing and analyzing the data increases
Solution Approach 1:
The patent enables the machine learning models to automatically learn and optimize themselves from historical glucose data without requiring complex manual feature engineering or intervention. The models self-train on patient-specific data patterns, automatically adapting to individual metabolic characteristics. This self-learning capability reduces the need for complex external processing systems while maintaining high prediction accuracy
Solution Approach 2:
The patent segments the complex prediction task into distinct computational stages: data collection and preprocessing, model training phase, and prediction execution phase. By dividing the overall system into these manageable segments, the complexity is distributed across different time periods and computational resources, making the system more tractable and implementable
Data Source
AI summary
Glucose prediction using machine learning (ML) and time series glucose measurements is described. Given the number of people that wear glucose monitoring devices and because some wearable glucose monitoring devices can produce measurements continuously, a platform providing such devices may have an enormous amount of data. This amount of data is practically, if not actually, impossible for humans to process and covers a robust number of state spaces unlikely to be covered without the enormous amount of data. In implementations, a glucose monitoring platform includes an ML model trained using historical time series glucose measurements of a user population. The ML model predicts upcoming glucose measurements for a particular user by receiving a time series of glucose measurements up to a time and determining the upcoming glucose measurements of the particular user for an interval subsequent to the time based on patterns learned from the historical time series glucose measurements.