Deep Multitask Learning by Stacked Long Short-Term Memory for Predicting Personalized Blood Glucose Concentration

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Endocrinology

Background:

• Inaccurate insulin infusion in Type I diabetes (T1D) poses risks of fatal glycemic complications.
• Accurate blood glucose concentration (BGC) prediction is vital for artificial pancreas (AP) control and clinical decision support.

Purpose of the Study:

• To develop and evaluate a novel deep learning (DL) model with multitask learning (MTL) for personalized BGC prediction in T1D.
• To improve the accuracy and reliability of BGC forecasting for enhanced diabetes management.

Main Methods:

• A DL model with shared stacked long short-term memory (LSTM) layers and gender-specific clustered dense layers was designed.
• Subject-specific dense layers were incorporated for fine-tuning personalized glucose dynamics.
• The model was trained and validated using the OhioT1DM clinical dataset.

Main Results:

• The proposed model demonstrated robust and reliable BGC prediction performance across various prediction horizons (PH).
• Achieved low error rates: 30-min (RMSE=16.06, MAE=10.64), 60-min (RMSE=30.89, MAE=22.07), 90-min (RMSE=40.51, MAE=30.16), 120-min (RMSE=47.39, MAE=36.36).
• Clarke Error Grid Analysis confirmed clinical feasibility with >94% of predictions in the safe zone up to 120-min PH.

Conclusions:

• The novel DL-MTL model offers accurate and personalized BGC prediction for T1D management.
• The model's performance surpasses existing state-of-the-art statistical, ML, and DL methods.
• This approach holds significant potential for advancing AP technology and clinical decision support systems.