Optimizing Glucose Sensor Calibration With Lightweight Neural Networks: A Comparative Study

Abstract

Diabetes presents a significant global health challenge, necessitating precise blood glucose monitoring for effective management. Continuous glucose monitoring (CGM) devices offer a minimally invasive approach, yet require accurate calibration models to improve reliability. This letter investigates various neural network architectures for predicting time errors in CGM sensor readings, aiming for high accuracy and minimal computational burden. Using simulated data, models were trained and evaluated, with Legendre memory unit and temporal convolutional network architectures emerging as promising candidates. With these architectures, it was possible to lower the sensor's reading error to, respectively, 24.22 and 25.34 mg/dL, decreasing the error by 40.6% and 37.9%. Furthermore, the letter explores the impact of sensor calibration frequency on prediction accuracy, revealing optimal performance with calibrations once every three or five days, obtaining an error in the reading of approximately 16 and 15 mg/dL. These findings underscore the potential for enhancing glucose monitoring systems and suggest avenues for future research in diabetes management.