Scholarlink

Type 1 diabetes is a chronic disease that requires meticulous monitoring of blood glucose and insulin management. Effective diabetes care needs glucose prediction, yet it’s a complex task. This study aimed to develop a predictive model for blood glucose levels one hour into the future using a newly collected dataset “BrisT1D” from young adults in the UK with type 1 diabetes. Ensemble models like ExtraTrees, Bagging, Voting, and Stacking Regressor, and Non-Ensemble models like LSTM and BiLSTM were implemented. Ensemble models are trained using K-fold cross-validation, and evaluated on metrics like RMSE, MAE, and \(R^2\) . Log1p transformation is applied to stabilize the data and address skewness. BiLSTM demonstrated robust performance by achieving an average RMSE of 0.0814, MAE of 0.0480, and \(R^2\) of 0.9304. Loss Curve revealed that the BiLSTM effectively captured patterns in the data while mitigating overfitting. Generalization was validated by testing on unseen participants, ensuring the model’s applicability to various real-world scenarios. The proposed BiLSTM with attention mechanism effectively predicts blood glucose levels. This approach contributes to the advancement of personalized diabetes management and highlights the potential of machine learning to address complex medical challenges.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Machine Learning for Blood Glucose Prediction: Insights from Real-World Continuous Monitoring Data

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Muhammad Abdullah Sarwar,
  • Egle Belousoviene,
  • Robertas Damaševičius,
  • Irfan Abbas,
  • Rytis Maskeliunas

摘要

Type 1 diabetes is a chronic disease that requires meticulous monitoring of blood glucose and insulin management. Effective diabetes care needs glucose prediction, yet it’s a complex task. This study aimed to develop a predictive model for blood glucose levels one hour into the future using a newly collected dataset “BrisT1D” from young adults in the UK with type 1 diabetes. Ensemble models like ExtraTrees, Bagging, Voting, and Stacking Regressor, and Non-Ensemble models like LSTM and BiLSTM were implemented. Ensemble models are trained using K-fold cross-validation, and evaluated on metrics like RMSE, MAE, and \(R^2\) . Log1p transformation is applied to stabilize the data and address skewness. BiLSTM demonstrated robust performance by achieving an average RMSE of 0.0814, MAE of 0.0480, and \(R^2\) of 0.9304. Loss Curve revealed that the BiLSTM effectively captured patterns in the data while mitigating overfitting. Generalization was validated by testing on unseen participants, ensuring the model’s applicability to various real-world scenarios. The proposed BiLSTM with attention mechanism effectively predicts blood glucose levels. This approach contributes to the advancement of personalized diabetes management and highlights the potential of machine learning to address complex medical challenges.