Personalized nutrition and machine-learning: Exploring the scope of continuous glucose monitoring in healthy individuals in uncontrolled settings

Abstract

Machine-learning models, when combined with continuous glucose monitoring (CGM), can help effectively analyze extensive datasets of glycemic responses to food. A total of 3,296 90-minute-long CGM meal responses from 927 healthy individuals were used to train/test a machine-learning model (XGBRegressor). The model input were the individual's anthropometric characteristics, macronutrients, and features related to the 24-hour CGM trace preceding the meal. The model output consisted in the parameters of a bell-shaped equation, used to analytically describe the glycemic response to food. To interpret the machine-learning model, the Shapley's values method was employed. A multi-linear regression model was used to study the impact of food macronutrients on the magnitude of the glycemic response. The machine-learning model was able to predict the magnitude of the glycemic response with a root mean square error of 13.2 ± 9.5 mg/dL, and a correlation coefficient of r = 0.48 (p < 0.001) but suffered from a systematic bias (r = 0.83, p < 0.001). The Shapley's values revealed that age brings a positive contribution to the magnitude of the glycemic response beyond 40 years. The multi-linear model (R2 = 0.14, p < 0.001) highlighted the positive and the negative impact of the carbohydrates (β = 0.263, p < 0.001) and fat (β = −0.108, p < 0.001) on the magnitude of the glycemic response, respectively. The maximum attainable accuracy in predicting the glycemic response to food using this machine-learning model may be inherently limited by the uncontrolled nature of the adopted dataset. Nevertheless, this model holds significant educational value for CGM users, as it facilitates comprehension of the intricate relationships between individual characteristics, meal composition, and glucose levels.