A glucose-specific metric to assess predictors and identify models

Area of Science:

• Biomedical Engineering
• Endocrinology
• Medical Informatics

Background:

• Mean Square Error (MSE) is standard for assessing glucose prediction models in diabetes.
• MSE equally weights errors, failing to account for the differing clinical impact of hypo-, eu-, and hyperglycemia.
• This limitation can lead to suboptimal model selection and inaccurate predictions in critical glycemic states.

Purpose of the Study:

• Introduce a novel cost function, glucose-specific MSE (gMSE), to address MSE's limitations.
• Enhance the assessment of glucose prediction quality and the identification of glucose models.
• Incorporate clinical relevance into glucose prediction error metrics.

Main Methods:

• Developed gMSE by modifying MSE with a penalty function inspired by the Clark error grid.
• gMSE penalizes overestimation in hypoglycemia and underestimation in hyperglycemia.
• Evaluated gMSE through ad hoc experiments and a prediction assessment problem comparing real glucose profiles.

Main Results:

• gMSE demonstrated sensitivity to accuracy, precision, and distortion in glucose predictions.
• In a comparison of two prediction profiles, MSE selected the clinically riskier scenario, while gMSE identified the safer one.
• Models optimized with gMSE showed improved accuracy in hypo- and hyperglycemia compared to those optimized with MSE.

Conclusions:

• gMSE offers a clinically relevant alternative to MSE for evaluating glucose prediction models in diabetes.
• The metric's mathematical properties, including no local minima, facilitate model identification.
• gMSE-identified models provide more accurate predictions in critical glycemic conditions, reducing clinical risk.