Identification of the Optimal Meal Detection Strategy for Adults, Adolescents, and Children with Type 1 Diabetes: An in Silico Validation

Current management of Type 1 Diabetes mellitus (T1D) resorts to manual meal announcements from the patient to manage postprandial glycemia; nevertheless, suboptimal glycemic control is observed in real data, with the presence of many hypoglycemic and hyperglycemic events. The utilization of Continuous Glucose Monitoring (CGM) sensors and Artificial Intelligence (AI) is paving the way for improved and automated glycemic control. A step in this direction is represented by the automation of meal detection, which would not require patients to perform tasks such as carbohydrate estimation and meal announcement that are error-prone, especially for children and elderly patients. In this work, we investigate several AI models for meal detection from in silico data of 10 adults, 10 adolescents, and 10 children with T1D using only CGM data, and compare them to the standard detection method based on the glycemic threshold. We generate 30 days of data per patient that include 5 meals per day and introduce human error on carbohydrate estimation to make data more similar to the real ones. The AI models can detect more than 81% of meals from any cohort of patients while producing a relatively small amount of false positives. The feedforward neural network, the support vector machine, and the threshold method are the most promising meal detection strategies for adult, adolescent, and child populations, respectively, and may improve patients' health and disease management.