A deep learning approach to adherence detection for type 2 diabetics.

Area of Science:

• Artificial Intelligence in Healthcare
• Biomedical Signal Processing
• Diabetes Management Technology

Background:

• Insulin treatment adherence is critical for managing type 2 diabetes (T2D) and preventing severe complications.
• Accurate monitoring of treatment adherence is a significant challenge in T2D patient care.
• Novel technological solutions are needed to improve adherence detection in T2D.

Purpose of the Study:

• To develop and evaluate a novel adherence detection algorithm for type 2 diabetes (T2D) patients.
• To leverage Deep Learning (DL) approaches for analyzing simulated Continuous Glucose Monitoring (CGM) signals.
• To compare the performance of different classification algorithms for adherence detection.

Main Methods:

• Simulated a large and diverse dataset of CGM signals for T2D patients using an adapted Medtronic Virtual Patient (MVP) model.
• Employed a comprehensive grid search to compare classification algorithms, including logistic regression, Multi-Layer Perceptrons (MLPs), and Convolutional Neural Networks (CNNs).
• Assessed algorithm performance based on classification accuracy.

Main Results:

• Convolutional Neural Networks (CNNs) demonstrated the highest classification performance.
• The best performing CNN model achieved an average accuracy of 77.5% in detecting adherence.
• Deep Learning approaches showed significant potential for adherence detection in T2D.

Conclusions:

• Deep Learning, particularly CNNs, shows considerable promise for developing effective adherence detection systems for T2D patients.
• The developed algorithm using simulated CGM data provides a foundation for future real-world adherence monitoring tools.
• This pilot study highlights the potential of AI in enhancing diabetes self-management and treatment adherence.