Development and Validation of a Meal Detection-Based Tool for User Behavior and Meal Management Analysis

Management of the timing and frequency of meal announcements remains a key determinant of postprandial glycemic control in hybrid automated insulin delivery (AID) systems. Tools that translate routine device data into clinically interpretable insights on meal management behavior for individual patients and broader cohorts remain limited. These tools can potentially provide clinicians and users objective data to formulate personalized meal-related recommendations to safely improve outcomes.

We developed a framework to quantify meal management behavior from routine AID data. The framework combines a meal detection model and a behavioral classification module that categorizes events as on-time, late, or missed based on temporal relationship between detected meals and carbohydrate entries. The meal detection model was trained on data from adults with type 1 diabetes using MiniMed 780G system (NCT04479826) and evaluated in a validation cohort of 157 individuals with type 1 diabetes using advanced hybrid closed-loop (AHCL) (NCT03959423) system.

Meal events detection accuracy reached 96.9% in the validation cohort, with precision ≥89.6%, recall ≥93.9%, and area under the curve = 0.996. Application of the tool to the validation cohort showed that a higher proportion of on-time announcements correlated with higher time in range (TIR) (ρ = 0.375, q < 0.001) and lower hyperglycemia. Each 10% increase in on-time announcements corresponded to a 2.66% increase in TIR. Increasing delay of late announcements was associated with worse postprandial glycemic outcomes.

This framework enables insights into meal announcement patterns that may support patient counseling, clinical decision support, and large-scale analyses in AHCL systems.

NCT04479826, NCT03959423.