Constrained IoT-Based Machine Learning for Accurate Glycemia Forecasting in Type 1 Diabetes Patients

Area of Science:

• Biomedical Engineering
• Computer Science
• Endocrinology

Background:

• Diabetes mellitus type 1 (DM1) management requires frequent blood glucose monitoring and prediction.
• Continuous Glucose Monitoring (CGM) and Internet of Things (IoT) devices generate vast amounts of biomedical data.
• Limited computational power of constrained devices hinders complex machine learning (ML) model deployment.

Purpose of the Study:

• To investigate the feasibility of local glycemia prediction on constrained IoT devices.
• To evaluate the performance of lightweight ML algorithms for real-time glucose forecasting.
• To assess the potential of smartphone-based ML for diabetes self-management.

Main Methods:

• Utilized continuous glucose monitoring (CGM) data and other biomedical signals.
• Applied edge computing principles to process data on constrained devices.
• Implemented lightweight machine learning algorithms, specifically Random Forest, with preprocessing and feature extraction.
• Tested the computational burden on constrained IoT devices, including smartphones.

Main Results:

• Demonstrated the feasibility of running ML algorithms on constrained IoT devices for local prediction.
• Achieved accurate glycemic level forecasting up to 45 minutes in advance.
• Confirmed acceptable prediction accuracy using the Random Forest model on a smartphone.

Conclusions:

• Local, real-time glycemia prediction is achievable on resource-constrained devices like smartphones.
• This approach can empower individuals with DM1 to better manage their glucose levels proactively.
• Integration of ML and edge computing in IoT devices offers a promising avenue for advanced diabetes care.