AWD-stacking: An enhanced ensemble learning model for predicting glucose levels

Area of Science:

• Biomedical Engineering
• Machine Learning
• Endocrinology

Background:

• Accurate blood glucose level prediction is crucial for optimizing insulin therapy and preventing complications in type 1 diabetes.
• Ensemble learning algorithms offer a promising avenue for enhancing predictive accuracy in complex physiological systems.

Purpose of the Study:

• To propose and evaluate an improved stacking ensemble learning algorithm for blood glucose level prediction.
• To assess the model's performance using standard metrics and compare it against existing non-ensemble methods.

Main Methods:

• Developed an ensemble model using three improved Long Short-Term Memory (LSTM) networks as base models.
• Integrated an improved nearest neighbor propagation clustering algorithm with adaptive weighting into the ensemble.
• Trained and validated the model on the OhioT1DM dataset.

Main Results:

• Achieved low Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) across various prediction horizons (30, 45, 60 minutes).
• Demonstrated significant improvements in RMSE (up to 27.92%) and MAE (up to 65.32%) compared to the StackLSTM model.
• Clarke Error Grid Analysis confirmed model errors were within 10%, indicating clinical relevance.

Conclusions:

• The proposed improved stacking ensemble model exhibits state-of-the-art predictive performance for blood glucose levels.
• The model's accuracy supports its suitability for clinical decision-making in type 1 diabetes management.
• This advancement holds significant importance for the effective treatment and management of diabetes.