A hybrid Transformer-LSTM model apply to glucose prediction.

Area of Science:

• Endocrinology
• Biomedical Engineering
• Data Science

Background:

• Global diabetes prevalence is rising, affecting over 536.6 million people worldwide.
• Accurate blood glucose monitoring is crucial for preventing severe complications like hypoglycemia and hyperglycemia.
• Existing methods face challenges in precise real-time glucose level prediction.

Purpose of the Study:

• To develop and evaluate a hybrid Transformer-LSTM model for enhanced blood glucose level prediction.
• To improve the accuracy of forecasting future glucose readings from Continuous Glucose Monitoring (CGM) data.
• To reduce the risk of diabetic complications through more reliable glucose level predictions.

Main Methods:

• Utilized a dataset of over 32,000 CGM data points from eight patients, including historical glucose readings and calibration values.
• Developed a hybrid Transformer-Long Short-Term Memory (LSTM) model integrating Transformer and LSTM architectures.
• Compared the performance of the hybrid model against a standard LSTM model using Mean Square Error (MSE).

Main Results:

• The hybrid Transformer-LSTM model demonstrated superior performance compared to the standard LSTM model.
• Achieved Mean Square Error (MSE) values of 1.18, 1.70, and 2.00 for 15, 30, and 45-minute prediction intervals, respectively.
• The model's accuracy in predicting future glucose levels was significantly enhanced.

Conclusions:

• Advanced machine learning, specifically the hybrid Transformer-LSTM model, shows significant potential for proactive diabetes management.
• Improved glucose level prediction accuracy can lead to better patient outcomes and reduced diabetic complications.
• This approach offers a critical step towards mitigating the global impact of diabetes.