Neural Networks With Gated Recurrent Units Reduce Glucose Forecasting Error Due to Changes in Sensor Location

Area of Science:

• Biomedical Engineering
• Data Science
• Endocrinology

Background:

• Continuous glucose monitors (CGMs) are vital for diabetes management, providing glucose level estimates.
• Neural networks (NNs) are increasingly used for forecasting glucose values from CGM data.
• Feedforward (FF) NNs can experience increased forecast errors when CGM sensor location changes.

Purpose of the Study:

• To evaluate gated recurrent unit (GRU) neural networks (NNs) for reducing forecast errors associated with CGM sensor location changes.
• To assess the efficacy of GRU NNs in mitigating glucose prediction inaccuracies in diabetes patients.

Main Methods:

• A neural network (NN) model utilizing gated recurrent units (GRUs) was investigated.
• The study involved 13 participants with type 2 diabetes using blinded CGMs on both arms for 12 weeks.
• Glucose prediction errors were analyzed concerning sensor location variations.

Main Results:

• GRU NNs did not yield significantly different glucose prediction errors when CGM sensor locations changed.
• Statistical analysis confirmed no significant impact of sensor location changes on GRU NN prediction accuracy (P < .05).

Conclusions:

• Gated recurrent unit (GRU) neural networks effectively mitigate prediction errors caused by variations in continuous glucose monitor (CGM) sensor placement.
• GRU NNs offer a promising approach to improve the reliability of glucose forecasting in diabetes management despite sensor location shifts.