# Coefficient of Variation to Assess the Reproducibility of Meal-Induced Glycemic Responses: Development of a Clustering Algorithm

## TL;DR

This paper introduces a clustering algorithm using glycemic variability to identify reproducible meal-specific glucose patterns in people with type 1 diabetes.

## Contribution

A novel clustering method based on coefficient of variation to assess reproducibility of postprandial glycemic responses.

## Key findings

- PPGR clusters varied by meal type and individual, with an average of 2.4-3.1 clusters per meal.
- Carbohydrate intake alone did not influence cluster formation, indicating complex meal composition effects.
- The algorithm shows potential for personalized diabetes management by predicting glycemic responses.

## Abstract

Managing type 1 diabetes (T1D) requires maintaining target blood glucose levels through precise diet and insulin dosing. Predicting postprandial glycemic responses (PPGRs) based solely on carbohydrate content is limited by factors such as meal composition, individual physiology, and lifestyle. Continuous glucose monitors provide insights into these responses, revealing significant individual variability. The statistical clustering method proposed here balances the number of clusters formed and the glycemic variability of the PPGRs within each cluster to offer a clustering technique on which treatment decisions could be based.

This study aims to develop and evaluate a PPGR clustering method that identifies reproducible meal-specific glucose patterns in people with type 1 diabetes.

Blood glucose data from the OhioT1DM dataset were used to assess clustering of PPGR based on the coefficient of variability (CV) of glucose. Clustering was performed using statistical clustering, with each PPGR isolated into 48 data points per event. A CV threshold of <36% was used to define clinically similar clusters. This aimed to cluster PPGRs with minimal glycemic variability. The approach aims to enhance precision in analyzing postprandial glycemic dynamics, assessing cluster cohesion via standard deviation and CV within meal categories.

The analysis revealed a reproducible set of PPGR clusters specific to meal types and individuals (mean [SD], 2.4 [1.8] for breakfast, 2.7 [0.9] for lunch, and 3.1 [1.0] for dinner), with the number of clusters varying across participants and meals in the dataset. Carbohydrate intake alone did not affect cluster formation, suggesting a complex relationship between meal composition and PPGR variability. However, certain individuals showed significant associations between carbohydrate intake and cluster formation for specific meals.

The meal-based glycemic clustering algorithm provides a promising framework for predicting PPGRs in people with type 1 diabetes, independent of carbohydrate intake. It emphasizes the need for personalized prediction models to optimize time in range and enhance diabetes management. Efforts to refine treatment strategies are crucial in reducing T1D-related complications.