AFTS: A patient-agnostic encoder-decoder architecture with directional attention for blood glucose forecasting

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Time Series Analysis

Background:

• Accurate blood glucose forecasting is crucial for diabetes management but remains challenging due to patient variability and complex glycemic dynamics.
• Existing deep learning models often struggle to generalize across different patients and capture intricate temporal patterns.

Purpose of the Study:

• To introduce AFTS (Adaptive Feature Time Series), a novel patient-agnostic deep learning architecture for improved blood glucose forecasting.
• To evaluate AFTS's performance against state-of-the-art models on real-world continuous glucose monitoring (CGM) datasets.

Main Methods:

• Developed AFTS, a deep learning architecture integrating a bidirectional LSTM encoder-decoder with cascaded Directional Representation (DR) modules.
• Implemented an axis-wise attention mechanism within DR modules to separately process temporal and feature dimensions.
• Evaluated AFTS on KDD18 and CDD23 CGM datasets using a patient-wise 80/20 split, comparing against twenty baseline models.

Main Results:

• AFTS achieved competitive Mean Absolute Errors (MAE) of 7.02 mg/dL (KDD18) and 7.39 mg/dL (CDD23) at a 30-minute prediction horizon.
• Ablation studies confirmed the significant contribution of the axis-wise attention mechanism to reducing prediction errors in complex glycemic scenarios.
• AFTS demonstrated robustness in patient-agnostic forecasting, balancing short-term and long-term trend capture.

Conclusions:

• AFTS presents a robust architectural approach for patient-agnostic blood glucose forecasting.
• The specialized axis-wise attention mechanism effectively refines features and minimizes prediction errors.
• AFTS offers a promising solution for improving diabetes management through accurate glycemic predictions.