A Study of Trained Clinicians’ Blood Glucose Predictions Based on Diaries of People with Type 1 Diabetes

Abstract

Summary Objectives: How accurate can trained clinicians predict blood glucose concentrations? Good clinical treatment is, among other things, related to understanding the factors influencing blood glucose level. We analyze trained clinician’s prediction accuracy in comparison with selected computer-implemented prediction algorithms and models. Methods: We have in this study included diaries of 12 people with type 1 diabetes. This test group consists of seven males and five females, ages 24 to 60, HbA1c 6.0 to 8.9 and a BMI between 20 and 28 kg/m2. Eight experienced clinicians tried to predict the blood glucose measurements based on minimum three days of diary history. Selected prediction algorithms and models were used for comparison. The reason we focus on type 1 diabetes is that it has the most critical insulin requirement, so accurate prediction can be more critical than for type 2. Results: An accuracy of 28.5% and an error of 26.7% were found from predictions made by the clinicians. A physiological model and an artificial intelligence model showed higher accuracy of 32.2% and 34.2% in comparison with the clinicians (p < 0.05). A simple predictor algorithm based on the mean blood glucose history showed significant (p < 0.05) lower total root mean square error compared to predictions made by the clinicians. Conclusion: To predict blood glucose level from diaries has shown to be profoundly difficult even for experienced clinicians in comparison with predictions from computer algorithms and models. This suggests that computer-based systems incorporating predicting algorithms and models are likely to contribute positively to the day-to-day treatment of people with diabetes.