Artificial Intelligence in Blood Glucose Level Prediction for Type 1 Diabetes Management

Effective management of type 1 diabetes mellitus (T1DM) reduces the associated complications. T1DM management aims to maintain blood glucose levels (BGLs) within a target range.
BGL prediction is an important tool to help maximise the time BGL is in the target range and thus minimise both acute and chronic diabetes-related complications.
Data-driven BGL prediction models estimate future BGL utilising current and past information and provide early warnings concerning inadequate glycaemic control.
Despite many works performed on BGL prediction, further improvements in prediction accuracy are still desired.
This thesis focuses on BGL prediction in T1DM using artificial intelligence.

As part of this thesis, advanced artificial intelligence techniques, including deep learning, ensemble learning, causal analysis, and data fusion are explored to enhance the performance of BGL prediction.
Leveraging deep learning and ensemble learning, three deep-ensemble models are proposed.
The superior performance of the proposed ensemble models over non-ensemble benchmark models is shown.
Also, the relations between BGL and affecting variables, including carbohydrate intake, injected bolus insulin, and physical activity, via the causality context are examined. Then, by proposing novel approaches, leveraging causality information as prior knowledge for BGL prediction is investigated.
The results show the effectiveness of using causality information in BGL prediction.
Moreover, new approaches for extracting information from physical activity, as a crucial factor in T1DM management, are developed, and the fusion of this information with BGL data at multiple levels is explored. Based on the results, incorporating physical activity into BGL prediction can improve prediction performance.
Furthermore, the performance of different data-driven time series forecasting approaches with different inputs for BGL prediction is examined and assessed to provide useful information regarding the primary choices of the model structure and input.