Insulin Omission Surrogate (iOS)

Our main hypotheses:

• that people living with Type 1 diabetes who report intentional insulin omission will have a higher risk of disordered eating behaviors and diabetes-related comorbidities;
• that it will be possible to establish different predictors of intentional insulin omissions behaviors by using machine learning techniques.

Statistical Analysis:

The normality of the data distribution will be checked for each value using a graphical analysis of the distribution and a Kolmogorov-Smirnov test before parametric tests are performed. A description of the characteristics of the participants included in this study will be performed. Continuous variables will be presented by the mean and standard deviation for normal distributions, median and range for others. Categorical variables will be presented by the number and percentage of each modality. The level of significance of the tests must be equal or lower than 0.05. The risk of alpha error is set at 0.05, with Bonferroni correction if necessary.

• Objective 1 - 24-hour dietary recall: Macronutrients composition and repartition (in the day), as well as the serving sizes of the food intake will be calculated using the Canadian Food Guide 2007 or the Canadian Nutrient File.
• Objective 2 - Prevalence and severity of diabetes and its comorbidities: Both groups, having or not intentionally omit insulin, will be compared using conventional statistical analyses.
• Objective 3 - Predictive modeling: Based on a previous published methodology (Iceta et al., 2021), the first step will be to reduce entropy in the dataset using a ranking procedure (the Fast Correlation-Based Filter, FCBF) to identify the most discriminating predictor between the two class-labeled datasets (with and without insulin omission). Only items with an FCBF score > 0.1 will be retained for the subsequent steps of the data-mining analysis. The second step of the data-mining analysis will aim at selecting the most relevant predictive algorithm for intentional insulin omission. Performances of different predictive algorithms will be tested and compared: logistic regression, artificial neural networks, naive Bayes classification, decision trees, AdaBoost meta-algorithm, CN2 rule inducer algorithm, SVM algorithm, k-nearest neighbors' algorithm and stochastic gradient algorithm. These artificial intelligence algorithms will be cross-validated (ten times in a row) with a randomized learning sample, renewed ten times and representing 66% of the study population. The validation sample will be represented by the other 33% of the population. The predictive algorithm with the best precision and F1 score will be considered as the best valuable algorithm.

Inclusion criteria

• Adolescents and adults living with type 1 diabetes
• Participated in the baseline phase of the BETTER registry
• Completed a 24-hours dietary recall within the BETTER registry (objective 1 only)
• Completed the two key questions on insulin omission in the BETTER registry (objectives 2 and 3 only)

Exclusion criteria

• Did not participated in the BETTER registry