Machine Learning–Based Calibration of Commercial Continuous Glucose Monitoring Sensor in Nonserum Solutions: An In Vitro Validation Study

Abstract Continuous glucose monitoring (CGM) systems, such as the FreeStyle Libre Pro (Abbott Diabetes Care), offer noninvasive glucose measurement. However, their accuracy in cerebrospinal fluid (CSF) glucose monitoring remains unvalidated. This study evaluates the performance of the FreeStyle Libre sensor against a standard laboratory analyzer and proposes a regression-based calibration model to enhance measurement accuracy in neurotrauma ICU. A FreeStyle Libre sensor was integrated into an experimental setup using an adapter. Sensor readings were recorded with glucose concentrations ranging from 50 to 275 mg/dL. A standard laboratory analyzer was used as the reference. A linear regression model was trained to correct sensor deviations, with interpolation (SciPy's interp1d) used for refined predictions. Real-time data acquisition was facilitated via Universal asynchronous receiver / transmitter (UART)-based serial communication, and adaptive learning enabled model retraining upon accumulating 10 sensor laboratory value pairs. Initial sensor readings exhibited significant deviations from laboratory values, particularly at lower glucose concentrations (mean absolute relative difference [MARD]: 30.45%). Postcalibration, the MARD was reduced to 8.92%, demonstrating improved accuracy. Interpolation further minimized deviations, correcting values such as 40 mg/dL (20% deviation) to 49.1 mg/dL (1.8% deviation) and 72 mg/dL (42.4% deviation) to 123.5 mg/dL (1.2% deviation). Adaptive learning progressively reduced the root mean square error (RMSE) from 23.7 to 9.8 mg/dL after 30 updates. The calibration model makes the FreeStyle Libre sensor more accurate for CSF glucose measurements. This method might be promising for monitoring CSF glucose continuously of patients with external ventricular drainage, improving patient care in the neurotrauma ICU.