Direct Versus Iterated Multi-Step Forecasting of Glycaemia in Type 1 Diabetics Using Autoregressive Models

Area of Science:

• Biomedical Engineering
• Computational Medicine
• Time Series Analysis

Background:

• Accurate glycaemia forecasting is crucial for diabetes management.
• Existing forecasting methods include direct and iterative approaches.
• The iterative approach's error accumulation is a known limitation, but its performance relative to the direct method remains unclear.

Purpose of the Study:

• To compare the performance of direct and iterative multi-step ahead glycaemia forecasting methods.
• To evaluate different Autoregressive with Exogenous Inputs (ARX) models for glycaemia forecasting.
• To determine the optimal forecasting strategy for short-term glycaemia prediction.

Main Methods:

• Implemented and compared direct and iterative forecasting approaches using various ARX models.
• Evaluated model performance for multi-step ahead predictions, specifically focusing on one-hour ahead (12-steps ahead) forecasting.
• Utilized a one-month period of training data for model development and comparison.

Main Results:

• The iterative forecasting approach demonstrated superior performance compared to the direct method for one-hour ahead glycaemia predictions.
• Classical linear ARX models outperformed more complex non-linear ARX models in this forecasting task.
• The findings suggest that iterative methods can be effective despite potential error accumulation.

Conclusions:

• The iterative approach is a viable and potentially superior strategy for short-term multi-step ahead glycaemia forecasting.
• Linear ARX models provide a robust and efficient solution for glycaemia forecasting with limited training data.
• Further research could explore hybrid models or longer-term forecasting horizons.