Time Series Modeling for Blood Glucose Prediction in Individuals with Type 1 Diabetes
Mohajeri, Najmeh (2025)
2025
All rights reserved. This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2025051913321
https://urn.fi/URN:NBN:fi:amk-2025051913321
Tiivistelmä
Managing blood glucose levels remains a critical challenge for individuals with Type 1 diabetes, where timely intervention can significantly impact overall health outcomes. With the increasing adoption of Continuous Glucose Monitoring (CGM) devices that continuously record glucose values, time series forecasting for proactive glucose regulation has become both feasible and increasingly accessible.
This study aims to develop and evaluate deep learning models to predict future blood glucose levels at two clinically meaningful time horizons: 60 and 120 minutes. Accurate prediction of glucose levels can facilitate earlier interventions, reduce the risk of hypoglycemic and hyperglycemic events, and ultimately enhance the quality of life for individuals living with diabetes.
To address this challenge, a hybrid deep learning architecture was designed that integrates Long Short-Term Memory (LSTM) networks with Convolutional Neural Networks (CNNs) to enhance predictive performance across both time horizons. Configured as a multi-output network, the model simultaneously generates forecasts for both target horizons using preprocessed CGM time series data.
The scope of this study is confined to the HUPA-UCM dataset, which includes detailed contextual data such as timestamps, blood glucose readings, calories burned during intervals, heart rate, step count, dosage of slow-release and long-acting insulin administered via injection, and carbohydrate intake.
Model performance on the test dataset demonstrated strong predictive accuracy, particularly for the 60-minute horizon. Clinical validation using Clarke Error Grid Analysis showed that 99.50% of predictions fell within Zone A and 0.50% within Zone B—indicating 100% clinical acceptability. While prediction performance decreased slightly for the 120-minute horizon, the model remained effective at capturing overall trends. In this case, 74.60% of predictions were in Zone A and 17.80% in Zone B, yielding a clinical acceptability rate of 92.40%.
These findings highlight the potential of deep learning-based approaches for short-term blood glucose forecasting and offer a foundation for future integration into personalized diabetes management systems.
Keywords: Type 1 Diabetes, Blood Glucose Prediction, CGM, Time Series Forecasting, Deep Learning, LSTM, CNN, Personalized Diabetes Management
This study aims to develop and evaluate deep learning models to predict future blood glucose levels at two clinically meaningful time horizons: 60 and 120 minutes. Accurate prediction of glucose levels can facilitate earlier interventions, reduce the risk of hypoglycemic and hyperglycemic events, and ultimately enhance the quality of life for individuals living with diabetes.
To address this challenge, a hybrid deep learning architecture was designed that integrates Long Short-Term Memory (LSTM) networks with Convolutional Neural Networks (CNNs) to enhance predictive performance across both time horizons. Configured as a multi-output network, the model simultaneously generates forecasts for both target horizons using preprocessed CGM time series data.
The scope of this study is confined to the HUPA-UCM dataset, which includes detailed contextual data such as timestamps, blood glucose readings, calories burned during intervals, heart rate, step count, dosage of slow-release and long-acting insulin administered via injection, and carbohydrate intake.
Model performance on the test dataset demonstrated strong predictive accuracy, particularly for the 60-minute horizon. Clinical validation using Clarke Error Grid Analysis showed that 99.50% of predictions fell within Zone A and 0.50% within Zone B—indicating 100% clinical acceptability. While prediction performance decreased slightly for the 120-minute horizon, the model remained effective at capturing overall trends. In this case, 74.60% of predictions were in Zone A and 17.80% in Zone B, yielding a clinical acceptability rate of 92.40%.
These findings highlight the potential of deep learning-based approaches for short-term blood glucose forecasting and offer a foundation for future integration into personalized diabetes management systems.
Keywords: Type 1 Diabetes, Blood Glucose Prediction, CGM, Time Series Forecasting, Deep Learning, LSTM, CNN, Personalized Diabetes Management