Optimizing blood glucose predictions in type 1 diabetes patients using a stacking ensemble approach

The diabetes pandemic, including 828 million adults worldwide in 2022, would benefit from continued development of novel, effective and accurate blood glucose prediction systems. Using the DiaTrend dataset, this study used stacking machine learning optimized by Grey Wolf Optimizer to construct and assess prediction models for blood glucose levels in type 1 diabetes patients. The DiaTrend dataset includes 27,561 days of continuous glucose monitoring and 8220 days of insulin pump data for 54 patients with type 1 diabetes. Grey Wolf optimization was used to tune and evaluate three machine learning algorithms – Random Forest, LSTM, GRU – for blood glucose predictions, whose predictions were then combined into an XGBoost stacking ensemble meta-learner. This study looked at three baseline algorithms for predicting blood glucose levels. Machine learning models Random Forest, LSTM, and GRU served as baselines, with MAE, RMSE, and MARD values. GRU had the best predictive accuracy of the initial models. Grey Wolf optimization contributed to achieving optimal baseline model results. Stacking ensemble learning via XGBoost meta-learner (MAE = 10.65, RMSE = 14.59, MARD = 6.98) achieved higher performance than the baseline models. The GRU method with Grey Wolf optimization outperformed the other models with the lowest MAE, RMSE, and MARD, but the Stacked XGBoost model fared best. These findings emphasize the need to improve parameter selection with approaches such as Grey Wolf or stacking ensemble methods to achieve accurate blood glucose predictions. These prediction models can aid in the continued development of monitoring devices, and algorithms for these devices, which contain alert systems for impending abnormal blood glucose levels, allowing for timely diabetes self-management. • Machine learning models (Random Forest, GRU, and LSTM) can achieve better results when optimized by Grey Wolf Optimizer in predicting blood glucose levels than their standalone models. • Gated Recurrent Unit with Grey Wolf optimization achieved the lowest Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Relative Deviation (MARD). • Stacking Ensemble Meta-Learner of XGBoost combined the predictions of the base models and achieved superior predictive accuracy • Including bolus data may improve predictive accuracy than merely using historical continuous glucose monitoring data • The blood glucose predictions may be used to improve continuous glucose monitoring devices that use automated user-alert systems.