Publication: Predicting Blood Glucose Levels with Machine Learning Techniques
4
0
Issued Date
2023-01-01
Resource Type
Scopus ID
2-s2.0-85180763089
Journal Title
Proceedings of the 2023 IEEE 6th International Conference on Knowledge Innovation and Invention, ICKII 2023
Start Page
515
End Page
519
Rights Holder(s)
SCOPUS
Bibliographic Citation
Proceedings of the 2023 IEEE 6th International Conference on Knowledge Innovation and Invention, ICKII 2023 (2023) , 515-519
Suggested Citation
Kitchainukoon N., Sae-Bae N. Predicting Blood Glucose Levels with Machine Learning Techniques. Proceedings of the 2023 IEEE 6th International Conference on Knowledge Innovation and Invention, ICKII 2023 (2023) , 515-519. 519. doi:10.1109/ICKII58656.2023.10332673 Retrieved from: https://hdl.handle.net/20.500.14740/20175
Author(s)
Author's Affiliation
Corresponding Author(s)
Other Contributor(s)
Abstract
We explored the effectiveness of machine learning algorithms in predicting blood glucose levels based on various input features, including demographic data, lifestyle factors, and physiological measurements using a dataset obtained from the OhioT1DM database. In particular, the dataset comprises data from 12 individuals with type 1 diabetes, spanning the years 2018 and 2020. The data for each individual is a comprehensive set of 22 features collected from devices, wearable sensors, and self-reports, capturing various lifestyle, and physiological aspects such as eating dairy, steps, heart rate, skin temperature, air temperature, and galvanic skin response (GSR). Machine learning algorithms, including Long short-term memory (LSTM) and Support Vector Regression (SVR) models, with historical data from 5, 10, 15, and 20 steps were employed to develop models to predict blood glucose levels at 30 and 60 minutes ahead. The accuracy of the predicted horizons was evaluated using three metrics, namely, MAE, MAPE, and RMSE to compare the performance of LSTM and SVR for each patient ID. To train and test the predictive models, the data for each individual in the dataset was split into training and testing sets. This approach ensured that the model learned from one individual's data and was evaluated on unseen data from the same individual. This individual-based training and testing procedure allowed for personalized predictions and accounts for inter-individual variations in diabetes management. The experimental results indicated that SVR performed better for the 30-min prediction, while LSTM performed better for the 60-min prediction. Among the patient IDs, ID596 achieved the best prediction results in terms of MAE, and RMSE, with values of 10.41 mg/dl MAE and 15.15 mg/dl RMSE at the 30-min prediction, and 18.26 mg/dl MAE, and 24.99 mg/dl RMSE at the 60-min prediction. Additionally, ID570 achieved the best prediction results in terms of MAPE, with values of 7.2% at the 30-min, and 13.3% at the 60-min prediction.