This prospective parallel pilot trial took place over two weeks in September 2018. Study design is depicted in Fig. 1. We recruited participants by visiting the schools and holding an assembly in which we (E.B. and S.S.) informed the students about the procedures of the study and invited them to participate approximately one month prior to the beginning of the study. At this time, informed consent forms in both English and the native language of Kiribati were distributed to ensure both the participant and their guardian had time to read over and understand the procedures of the study. The only inclusion criterion was enrollment in one of the two high schools on Kiritimati. The only exclusion criterion was age > 18 years. All procedures and data collection occurred at school during school hours.
Informed consent was obtained from all participants and their guardian in either English or the native language of Kiribati. Participants were provided $20USD compensation and each school also received $1000USD. The Institutional Review Board at the University of Southern California approved the study which was conducted according to the protocol registered at ClinicalTrials.gov (NCT04319003). The study was overseen by the local representative of the Kiribati Ministry of Health and Medical Services and conformed to the principles embodied in the Declaration of Helsinki.
Randomization and masking
We used a random number generator in Microsoft Excel (function: =RANDBETWEEN(1,2)) to assign the two high schools on the island as the control school (1) and the intervention school (2). No one was blinded to the allocation, although schools/students were not specifically told of the allocation. Participants were, however, blinded to their CGM readings, as only the investigators had access to the CGM readers. J.P. and C.R. conducted the procedures of the study.
We conducted the intervention one-week after baseline measures were taken in the intervention school. The intervention consisted of: 1) installation of a no-power water filtration system at the intervention school (LifeStraw Community, Vestergaard Frandsen Inc., Washington, DC); 2) the provision of colorful insulated stainless steel water bottles to students (DHGate.com Vacuum Insulated Water Bottle KKA2155, China); and 3) a 30-minute presentation given at school assembly (with students and teachers present) by a registered dietitian on the role of sugar in the pathogenesis of T2DM and how sugar reduction can reduce risk.
Height and weight were measured in duplicate and then averaged using a Tanita BC-549 scale (Tanita Corporation, Arlington Heights, IL) and a Seca 213 stadiometer (Seca, Chino, CA). Participants wore light clothing and no shoes. Gender and birthdate were self-reported. BMI categories (underweight, normal weight, overweight, and obese) were determined based on the WHO definition (overweight: > +1 standard deviation (SD) and obese: > +2SD from the mean BMI for the age and sex of the child) using the AnthroPlus packagefor R . Height-for-age z-scores were also calculated using this package.
Continuous glucose monitors (CGMs; Abbott FreeStyle Libre Pro, Chicago, IL) were attached on the first day of the study and covered by an adhesive patch to prevent accidental loss. (SIMPATCH, New York, NY). In the event that a CGM fell off, students were asked to retain the CGM so data could be downloaded, and another CGM was attached. To assess accuracy of the CGM, we also conducted a random fingerstick measurement on a calibrated glucometer (Contour NEXT EZ, Ascensia Diabetes Care, Parsippany, NJ) during the second week of the study. We developed a beverage frequency questionnaire of common Kiribati drinks by asking the participants and teachers about the different types of drinks available on the island, and by visiting local stores during a previous visit to the island. For each drink, we verbally asked, “how many days in the last week did you drink each beverage listed below” (0, 1–2, 3–4, 5–6, or 7 days) and “how much of this drink did you drink each day” (0.5, 1, 1.5, 2, or 3 cups) and recorded responses on a worksheet that also included pictures of drink amounts which we pointed to. From this, an estimate of servings/week was calculated, where a serving was considered to be 1 cup (8 oz). In addition, 24-hour diet recalls were conducted in a subset of participants (every second participant to complete the beverage recall at baseline) by an experienced registered dietitian using pen/paper, and subsequently entered into Nutrition Data System for Research software version 2018, developed by the Nutrition Coordinating Center (NCC), University of Minnesota, Minneapolis, MN, upon return to the United States. We were only able to conduct 24-hour recalls on a subset due to the time-consuming nature of the procedure (~ 20 minutes per participant) within the context of the school day. Recipes for frequently consumed drinks on the island including “tioka” and “ice block” were obtained from study participants as well as teachers and other community members who often prepared and provided the drinks to the students. These recipes were then entered as custom recipes into NDSR.
One-week after the intervention, CGMs were removed and the data was downloaded to a portable hard-drive. The control school was provided with the intervention upon completion of the study.
Raw CGM data from each participant was separated into two files: pre-intervention and post-intervention, each approximately one-week in duration. In the event that more than one CGM was used for a student, the data was manually stitched together using Microsoft Excel. Statistics were then calculated for each file using GlyCulator . GlyCulator is an online tool for analyzing CGM data, and generates measures of glucose variability including mean glucose, estimated HbA1c, standard deviation (SD), J-index, area under the curve (AUC), MAGE (Mean amplitude of glycemic excursions) for the entire duration of measurement . Diabetes was defined as an estimated HbA1c ≥ 6.5% (≥ 48 mmol/mol) and prediabetes was defined as an estimated HbA1c in the range of 5.7–6.4% ( 39–47 mmol/mol) .
As this was a pilot trial, and there was no existing data on diet or blood glucose in Kiritimati adolescents, we did not conduct a power calculation. Instead, we attempted to enroll as many students as were interested in participating. We intend that the data gathered from this pilot will be used to inform a power calculation for future, larger studies.
Patient and public involvement in research
The protocol of this study was designed with input from members of the Kiribati public, including doctors, nurses, and teachers on the island, and members of the Ministry of Health. On previous trips to the island, we enquired about the following: 1) biggest concerns facing the population with regards to high T2DM rates; 2) barriers to a healthy diet; and 3) factors they believed would improve their metabolic health. The taste and smell of water on the island was determined to be a major barrier contributing to high SSB consumption. We designed the intervention with this in mind – specifically the water filter installed in the school (where many students not only learn, but actually live) and the provision of individual insulated water bottles. As stated previously, we also had input from students and teachers in creating the beverage frequency questionnaire.
All statistical analyses were performed using R Studio version 1.2.5001 , with significance assumed at p < 0.05. To account for any baseline differences between the two schools, we conducted a difference-in-differences analysis, in which only the time x group interaction term was considered [13, 14], according to the following formula:
Y = β0 + β1*[Time] + β2*[Group] + β3*[Time*Group] +ε
Where Y represents the dependent variable, β0 represents the baseline average, β1 represents the time trend in the control group, β2 represents the difference between the two groups pre-intervention, β3 represents the difference in changes over time, and ε is the error term.
For outcomes, outliers were assessed as any value outside 1.5 x inter-quartile range. These values were then capped – meaning observations outside the lower limit were replaced with the value of the 5th percentile, and those above the upper limit were replaced with the value of the 95th percentile. We used the gvlma function in R to test the assumptions of linear regression: skewness, kurtosis, link function, and heteroscedasticity. Where data violated these assumptions, a box-cox transformation was applied, and the p-value represents the transformed data. However, the data itself remains untransformed in tables for clarity. Post-intervention data are included regardless of whether the participant was present on the day of the educational presentation (intention-to-treat principle).