(601a) Machine Learning for Automated Meal Detection in Glucose Control Systems for People with Diabetes

Feedforward disturbance announcements enable attenuation of the effects of disturbances on the controlled variable in closed-loop control. Meals are major disturbances to the glucose homeostasis and automated detection of meal consumption and carbohydrate estimation of the consumed meal are critical for fully automated artificial pancreas control systems [1]. In this study, a machine learning approach that integrates trend detection based on qualitative analysis, and meal estimation based on a fuzzy logic system is used to develop a meal detection and estimation module.

A qualitative analysis technique translates the variations in the time-series glucose concentration measurements to a set of seven qualitative variables (QV) [2]. The QVs are associated with different patterns of glucose change inferred by signs of the first and second derivatives. The meal is detected when the pattern of glucose variation among the seven QVs best matches the QV associated with an accelerating increase (i.e., both derivatives are positive). A fuzzy logic approach is employed to quantify the similarity between the glucose variations and the seven QVs. Positive, zero and negative fuzzy sets are defined for both the first and second glucose derivatives where the degrees of membership for the glucose derivatives belonging to these fuzzy sets are directly used for computing the similarities with the QVs [3].

Following meal detection, a fuzzy logic-based system quantifies the meal size, or more specifically the absorbed carbohydrates in the bloodstream. The development of the algorithm using fuzzy logic is motivated by the approximate reasoning ability and low computation requirements of fuzzy systems since the module will be used in computational environments such as smart phones that have limited computational power. The two inputs of fuzzy estimator are the ratio of measured glucose rise and the estimated effective insulin present in the bloodstream. The estimation algorithm determines a distributed approximation of the carbohydrates (meal size), followed by the calculation of the amount of insulin infusion (manipulated variable) to compensate for the unannounced meal. The learning component of the meal detection and carbohydrate estimation module captures the intra-and inter patient variability. Two parameters of the algorithm are adaptively modified depending on the risk function defined based on glucose deviation from the control target range.

The algorithm is employed for both in silico subjects and clinical experiments. For in silico study, the performance of glucose control using the proposed automated meal detection module is compared with a meal announced approach [3]. Using clinical experiments, we evaluate the accuracy of the meal detection and meal size estimation module over 117 consumed meals/snacks. Sensitivity, the percentage of correctly detected meals, is 93.5% for meals and 68.0% for snacks. Detection time defined as the time elapsed from the start of the meal to the detection is 34.8 22.8 min which is a reasonable due to physiological delay and gradual dynamic of CHO absorption into bloodstream. The estimation errors, the actual minus estimated CHO, approximately follow a Gaussian distribution centered around zero (mean: -1.7, standard deviation 28.1 g).

[1] A. Cinar, "Multivariable adaptive artificial pancreas system in type 1 diabetes". Curr Diab Rep 17:88, 2017.

[2] B. R. Bakshi and G. Stephanopoulos, "Representation of process trends-III. multiscale extraction of trends from process data," Comput. Chem. Eng., vol. 18, no. 4, pp. 267-302, 1994.

[3] S. Samadi, K. Turksoy, I. Hajizadeh, et al.: "Meal detection and carbohydrate estimation using continuous glucose sensor data". IEEE J Biomed Heal Informatics vol. 21, no. 3, pp. 619-627, 2017.