AttenGluco: Multimodal Transformer-Based Blood Glucose Forecasting on AI-READI Dataset

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Endocrinology

Background:

• Diabetes mellitus is a chronic metabolic disorder with serious complications.
• Accurate blood glucose level (BGL) prediction is crucial for diabetes management.
• Existing methods struggle with multimodal, irregularly sampled data and long prediction horizons.

Purpose of the Study:

• To develop a robust framework for long-term blood glucose prediction.
• To effectively integrate multimodal data sources like CGM and activity data.
• To enhance forecasting accuracy by addressing data fusion and temporal dependency challenges.

Main Methods:

• Proposed AttenGluco, a multimodal Transformer-based framework.
• Utilized cross-attention for fusing CGM and activity data with different sampling rates.
• Employed multi-scale attention to capture long-term temporal dependencies.

Main Results:

• AttenGluco demonstrated improved performance across key error metrics (RMSE, MAE) and correlation.
• Outperformed the multimodal LSTM baseline by approximately 10% in RMSE and 15% in MAE.
• Evaluated on the AIREADI dataset across diverse subject cohorts (healthy, prediabetes, type 2 diabetes).

Conclusions:

• AttenGluco offers a promising approach for accurate long-term blood glucose forecasting.
• The framework effectively handles multimodal, irregularly sampled data for improved diabetes care.
• Further research can explore its adaptability and performance with evolving datasets.