Blood Glucose Prediction with Variance Estimation Using Recurrent Neural Networks

Area of Science:

• Artificial Intelligence in Medicine
• Biomedical Data Science
• Endocrinology and Diabetes Management

Background:

• Blood glucose levels in type 1 diabetes are influenced by complex factors like insulin action and carbohydrate intake, with varying effects.
• Accurate short-term glucose level prediction is crucial for manual insulin dosing and automated closed-loop systems.
• Continuous glucose monitoring (CGM) systems offer rich data for developing predictive models.

Purpose of the Study:

• To develop and evaluate an approach for predicting blood glucose levels in type 1 diabetics up to 1 hour into the future.
• To provide a measure of certainty alongside glucose predictions to aid user interpretation.
• To create a computationally inexpensive, end-to-end trainable model requiring only historical glucose data.

Main Methods:

• Utilized recurrent neural networks (RNNs) trained in an end-to-end fashion using only patient glucose level history.
• The RNN was trained to parameterize a univariate Gaussian distribution for output, providing a certainty estimate.
• Evaluated performance using standard Root Mean Squared Error (RMSE) and the Surveillance Error Grid (SEG) metric.

Main Results:

• The proposed RNN approach achieved state-of-the-art performance on the Ohio T1DM dataset for blood glucose prediction.
• The model successfully provided a meaningful estimate of prediction certainty.
• The method requires no manual feature engineering or data preprocessing and is computationally efficient.

Conclusions:

• Recurrent neural networks offer a powerful and efficient tool for short-term blood glucose prediction in type 1 diabetes.
• The integrated certainty estimation enhances the clinical utility of glucose forecasting models.
• This data-driven, end-to-end approach simplifies model development and deployment for diabetes management.