Carbohydrate Count Aided by a Simulation in People With Type 1 Diabetes Mellitus. A Protocol for a Clinical Trial

Overview

Insulin remains the only approved treatment for type 1 diabetes mellitus patients and is used by many with type 2 diabetes. Carbohydrate counting is the most recommended way to prescribe prandial insulin dose because it is safe and efficacious, and also it allows a more variate diet to patients. Methods to improve carbohydrate counting include automatization of the process, optimizing carbohydrate meal content estimation, and including other nutrients such as fat into the equation. Being an iterative process that patients perfect by practicing and repeating, we believe that using simulations can improve carbohydrate counting. Simulations allow individuals to practice in a safe environment and help build confidence in one's ability to perform a task. In this clinical trial, patients assigned to the intervention group will have installed the STUDIA app, an automatic carbohydrate counter coupled to a mathematical model that simulates glucose excursions at the individual level using the patients' parameters in their smartphone. Time in range will be measured using a continuous glucose monitor.

Study setting Patients with T1DM trained in carbohydrate counting whose follow-up is considered feasible will be recruited from university hospitals and clinics dedicated to treating people with DM in the city of Medellín, Colombia. Cointerventions Since only patients with T1DM will be recruited, patients in both groups are expected to be treated with insulin, either with a multiple injection scheme or an insulin pump. There are no other approved interventions for the treatment of T1DM. Researchers will be instructed to avoid the use of unapproved drugs for T1DM. Protocol discontinuation The protocol will be discontinued if any of the following conditions are met: The patient does not wish to continue participating in the study. The patient's treating physician decides that she should not continue in the study due to the need for additional therapies that could affect the patient's ability to interact with the application or substantially modify glucose behavior and require other interventions, such as treatment with high doses of glucocorticoids. The patient requires hospital treatment for more than 72 hours, and the use of the application at the researcher's discretion may interfere with her care or put her health at risk. The patient develops a condition that contraindicates continuous glucose monitoring, such as intractable allergies to adhesives, frequent radiological evaluations that could contraindicate the implantation of devices, or any generalized skin condition that makes it impossible to insert the device in any area of the body. Sample size The sample size has been calculated based on the primary outcome, i.e., the expected difference in the TIR. Twenty-eight patients, 14 per group, will be required to analyze time in range using a mixed-effects model assuming a standard deviation of 4%, the statistical power of 90%, a type I error rate of 0.05%, and intra-subject variability of 5%, and inter-subject variability of 9%. The sample size was adjusted to 15 patients per group to admit a withdrawal rate of 10% in the intervention group, assuming that no patients in the control group could access the intervention. To maintain statistical power, sample size was adjusted by a factor that represents the retirement ratio. Recruitment Participants will be recruited from University Hospitals and clinics specializing in the treatment of T1DM, to which patients are referred for treatment with an insulin pump or CGM support in Medellín, Colombia. Participants will be invited to the study by their treating physicians at each site. Random sequence generation Participants will be randomly assigned in a 1: 1 ratio to the intervention or control group by a sequence of computer-generated random numbers, using permuted blocks of 4 to 6 individuals of variable size. The size of the blocks will not be revealed to ensure the concealment of the sequence. The sequence will be generated using a web application. Allocation concealment mechanism Allocation concealment will be maintained using a web-based central randomization system. This system will not deliver the allocation code to a coordinator until after the last evaluation visit, preventing the researcher from knowing the group in which the patient will be included. Random sequence implementation During the study, randomization will be implemented by the Pablo Tobón Uribe hospital research unit. The study coordinator will access the sequence list and assign the generated code to each patient. Within the application, the codes and the assignment of each code to one of the arms of the study will be prerecorded. Doing so will ensure the independence of the researchers and those responsible for data analysis. Blinding Due to the nature of the intervention, patients and investigators will not be blinded. Instead, those responsible for data analysis will be blinded until the data analysis is complete. This will be done by assigning codes to the patients with whom those in charge of processing the data will identify each data set and build the databases. The assignment groups will be named Group A and Group B, and the intervention of each group will be known only to the coordinator in the research unit. The groups will be declassified for interpretation once the study is finished and the analysis plan is complete. Data collection methods The data will be collected using the code assigned to each patient during the generation of the randomization sequence, and the databases will be built weekly with the following information: Time in range and time out of range: The time in range and out of range will be calculated from the data obtained from the continuous glucose monitor at each weekly visit by one of the investigators. This person will be trained in retrieving the data from the CGM. Grade 3 hypoglycemic episodes: At each visit, one of the investigators will inquire and record the Grade 3 hypoglycemic episodes that the patient has presented. Glycemic variability: The glycemic variability will be calculated based on the data obtained from the CGM. Degree of agreement of the phenomenological-based model: The data to estimate the degree of agreement in the glucose levels of the phenomenological-based model will be obtained by comparing the simulation given by the model and the data obtained by the CGM. The glucose values estimated by the model will be stored on a server for the duration of the study and will later be entered into a database for analysis. Actions carried out by the participants: The decision entered by the patient is reported in the same application that performs the carbohydrate count and stored on a server for later analysis. Usability of the application: The participants will fill out the usability scale during the last visit. This data will be stored on a server and subsequently analyzed. Figure 2 shows the patient's flowchart. Retention Given the short follow-up time, no losses are expected. However, all reasonable efforts will be made to ensure follow-ups. The next visit will be scheduled at each follow-up visit, and the participants will be reimbursed for transportation costs if required to attend the visits. Participants may leave the study at any time and for any reason. In addition, researchers may withdraw participants if they consider any danger to the participant continuing in the study. Participants may be withdrawn from the study if the competent authority decides to do so. Data management All data will be recorded electronically in forms stored on a virtual server during the study. Access to these forms will be restricted and only authorized by the Pablo Tobón Uribe Hospital research unit. The data obtained from the CGM will be stored in its original format on a virtual server once the patient has been de-identified and can be accessed for data quality verification reasons. Pablo Tobón Uribe Hospital's research unit will store all data indefinitely, but any information that could identify the patients will be deleted. Statistical methods Table 1 shows the initial variables of the groups contained in the table for the operative definition of variables (Supplementary Table1) and will be presented as divided by the treatment assignment groups. The normality analysis will be performed using a Q-Q (quantile-quantile) graph and the Shapiro-Wilk test. According to data distribution, continuous variables will be presented as mean and standard deviation or medians and interquartile ranges. Discrete variables will be presented as medians and interquartile ranges, and qualitative variables will be presented as absolute frequencies and proportions. Analysis population Complete analysis set: The full analysis set will consist of all randomized individuals, including follow-up and protocol deviation losses. Per-protocol analysis set: The per-protocol analysis set will be made up of individuals who use the CHOC + Sim application at least 80% of the time. This proportion of the use of the simulation is proposed based on previous results that show that CGM sensor wear time is associated with greater reductions in glucose levels. Objectives analysis plan A complete analysis set will be used for the primary endpoint and intention-to-treat analysis. This will be compared with the analysis result as a sensitivity analysis, taking into account only the set of per-protocol analyses. A mixed model analysis will be used to evaluate the time in range difference. The TIR is included as the dependent variable in the mixed-effects model analysis. The type of treatment and the time since randomization are included as independent variables. The TIR before randomization, type of insulin treatment, and pre-randomization HbA1c will be included as covariates. The secondary counting outcomes will be analyzed in the per-protocol set. The number of episodes of Grade 3 hypoglycemia and number of DKA or HSS will be compared in both groups using a ratio of incidence rates, calculated using a Poisson regression if the variance is less than the mean or by negative binomial regression otherwise (60, 61). Agreement between the phenomenological-based model and CGM measurements will be analyzed using limits of agreement (LoA) in the complete data set. However, the clinical interpretation of LoA is not straightforward without a minimal clinically significant difference. More importantly, it does not consider the clinical risk derived from an error in the estimation. Therefore, the authors also decided to perform an error grid analysis.