Prediction of Blood Glucose Concentration Based on OptiScanner and XGBoost in ICU

Abstract

Hyperglycemia, a stress-induced physiological condition, is associated with severe complications, including sepsis, multiple organ failure, and higher mortality rates. The seminal 2001 Leuven study highlighted the potential for strict blood glucose control (80-110 mg/dL) to lower mortality rates by 34% among critically ill surgical patients. Consequently, monitoring blood glucose levels in ICU patients has become imperative. This study aims to use recent medical technology advancements to streamline the monitoring of blood glucose levels, traditionally requiring trained personnel to operate a blood glucose monitor. We used the OptiScanner to collect patient blood data, separate plasma, and acquire mid-IR-related data. XGBoost was used to improve the prediction of blood glucose concentration based on patient classification types and its performance was compared with two other machine learning algorithms. We also used the LASSO model to predict plasma blood glucose concentrations. Additionally, we applied SHAP (SHapley Additive exPlanations) to identify critical wavelengths in the classifier and compared these with the functional groups corresponding to the actual IR spectrum. Our experimental findings demonstrate that XGBoost exhibits promising performance. Furthermore, the interpretation of the model is in alignment with domain knowledge. Through this study, we emphasize the potential of advanced medical technology, particularly machine learning algorithms such as XGBoost, to improve the efficacy and precision of blood glucose monitoring in ICU settings.

Keywords

• Hyperglycemia
• blood glucose monitoring
• OptiScanner
• XGBoost
• machine learning
• LASSO model