Study design and participants

SMASH (simplified meal announcement study) was an open-label, two-centre, two-period (each 3 months), crossover randomised trial conducted in two tertiary care specialist diabetes centres in Switzerland (Bern and Zurich). Prior to study commencement, approval was received by the Ethics Committees in Bern and Zurich, Switzerland (BASEC-ID 2022-D0087) and the study was pre-registered with clinicaltrials.gov (NCT05481034). Study monitoring was performed by the Clinical Trials Unit (University of Bern). Participants or parents/guardians (for participants aged 12–13 years) signed informed consent before any study-related activities. Consecutive eligible patients were informed about the study and invited to participate, in an effort to obtain a study sample representative of the broader population of interest. The participants’ sex was self-reported.

Key inclusion criteria were diagnosis of type 1 diabetes for at least 6 months, age 12–20 years (inclusive), HbA1c levels ≤107.7 mmol/mol (12.0%) and any prior insulin treatment modality to generate a representative sample of youth and young adults with type 1 diabetes. Key exclusion criteria were pregnancy or planned pregnancy (detailed eligibility criteria are reported in electronic supplementary material [ESM] Table 1).

Procedures

Patients were screened for eligibility with locally measured HbA1c (DCA Vantage Analyzer, Siemens Healthcare Diagnostics, Tarrytown NY, USA) and a urine pregnancy test in women and girls of childbearing age. The study flow is outlined in ESM Fig. 1. Following enrolment, participants were randomised to CC-SMA or SMA-CC using permuted block randomisation. The randomisation sequences were generated in R and uploaded into the randomisation module in REDCap by a person not involved in the study, to ensure allocation concealment until the time of randomisation. There was no washout between periods.

Before starting the first period, participants attended training and received instruction on the allocated meal management intervention. Participants were provided with study equipment and underwent measurement of HbA1c, weight and height. Participants filled in questionnaires for psychosocial metrics and dietary habits (food frequency questionnaire) [16, 17] and completed an electronic food picture quiz to evaluate their CC skills (16 dishes, meal CHO content 7–140 g).

After completing the first study period, participants attended the research facility before crossing over to the second period. Study assessments (HbA1c, anthropometrics, questionnaires) were re-measured. Meal settings in the HCL app were changed according to allocated intervention sequence, with respective instructions for use, and body weight was updated, where applicable.

At the end of the study, participants underwent the study assessments (HbA1c, anthropometrics, questionnaires), returned the devices and transitioned back to usual care treatments as per arrangements with their treating diabetologists.

Throughout the study, participants or their treating diabetologists were free to adjust their diabetes therapy. No active treatment optimisation, dietary restrictions or remote monitoring were undertaken by the research team.

All participants were provided with a 24 h telephone helpline to contact the research team for study-related support.

HCL system and meal management intervention

All participants were treated with the mylife CamAPS FX system consisting of the mylife CamAPS FX app (CamDiab, Cambridge, UK) on an Android phone (study phone or compatible private phone of participants), the Dexcom G6 sensor (Dexcom, San Diego, CA, USA) and the YpsoPump (Ypsomed, Burgdorf, Switzerland) with Orbit soft or micro or YpsoPump Inset infusion sets. The HCL app used the Cambridge adaptive model predictive control algorithm to modulate pump insulin delivery every 8–12 min, according to an adjustable glucose target (4.4–11.0 mmol/l). HCL was initialised using the participant’s body weight, total daily insulin dose (TDD) and the default target of 5.8 mmol/l.

Participants were instructed to use the bolus calculator in the app for meal insulin bolus delivery (see Fig. 1), which was identical during both periods but navigated differently. In the CC period, participants were asked to enter the number of grams in CHO into the bolus calculator (lower line for CHO entry). For the SMA period, standard CHO meal sizes were set on the basis of a 3-day CHO intake record prior to starting the first study period. Participants quantified consumed CHO by means of their usual CC method and provided documentation according to their preference. The mean CHO intake per meal was rounded and classified as a medium meal size. A snack was defined as 25%, a small meal 50% and a large meal 150% thereof. In SMA, participants were asked to choose between one of the four meal size icons on the bolus calculator in the app (upper section of the bolus calculator). In the CC period, standard CHO meal sizes were set to minimum values (0 g, 2 g, 4 g, 6 g) to avoid using the meal size icons. During both periods, the bolus dose was calculated based on the programmed carbohydrate to insulin ratio (CIR).

Fig. 1

User interface of the bolus calculator on the mylife CamAPS app used for CC (left) and SMA (right). During CC, participants entered the counted number of grams of carbohydrates (red circle). To avoid using the meal size icons along the selection line at the top of the bolus calculator, standard CHO meal sizes were pre-set to minimum values (0 g, 2 g, 4 g, 6 g) in ‘Settings’ prior to starting the CC period. Before starting the SMA period, standard CHO meal sizes were personalised according to participants’ 3 day carbohydrate intake record, and participants were asked to tap on the meal size icons for prandial insulin delivery. During both periods, prandial insulin doses were calculated based on the programmed CIR

Additionally, participants were asked not to apply advanced meal functions in the HCL app except for hypoglycaemia correction. CIR and mandatory settings were programmed during the on-boarding visit. We recommended using the system with a single CIR over a 24 h period. Changes in the settings during the course of study were recorded. Given the systems’ compatibility with both ultra-fast and fast-acting insulin analogues, participants were free to choose, but were asked to keep the type of insulin identical throughout the study.

Endpoints

The primary endpoint assessment was a non-inferiority comparison between the percentage of time spent with glucose levels between 3.9 and 10.0 mmol/l during each study period. The non-inferiority margin was set to 5 percentage points (pp), in line with the international consensus on a clinically meaningful difference [18].

Secondary endpoints included time spent at glucose levels >10.0 mmol/l, >13.9 mmol/l, <3.9 mmol/l and <3.0 mmol/l; mean sensor glucose; glucose variability measured by CV and SD; extended hypoglycaemia and hyperglycaemia event rate (number of events with glucose <3.9 mmol/l or >13.9 mmol/l lasting for ≥120 min); number of hypoglycaemia events lasting ≥15 min (<3.0 mmol/l, <3.9 mmol/l); HbA1c; glucose management indicator (GMI); and peak postprandial glucose levels (assessed within 180 min following CHO entry >25 g).

Additional points of interest included insulin doses, CHO metrics, utility evaluations (per cent time of sensor glucose availability and HCL operation when sensor glucose was available) and questionnaires assessing psychosocial metrics and dietary habits. Safety evaluation included the frequency of severe hypoglycaemia, diabetic ketoacidosis and other adverse or serious adverse events (SAEs).

Statistical analysis

Assuming a mean difference of 3pp in time spent with glucose levels between 3.9 and 10.0 mmol/l and an SD of 6.5pp, a sample size of 42 was determined to achieve a power of 80% at an alpha level of 0.025 for the primary endpoint using a non-inferiority margin of 5pp. To account for dropouts (expected dropout rate was 5%) we aimed for a target sample size of 45 participants. The SD was estimated based on the results of a clinical trial using the CamAPS FX system [19].

Primary and secondary endpoints were compared using generalised linear mixed-effects models. For endpoints derived from count data, a negative binomial distribution and a log link function was used. For all other endpoints, a linear mixed-effect model was used. For highly skewed data, a logarithmic transformation was applied, appropriate for the characteristics of the data. Models were adjusted for the period effect and accounted for within-participant correlations arising from the crossover design and for the variability across centres (period was considered as a fixed effect and participants nested within centres as a random effect). Model assumptions were evaluated using graphical methods including Residuals vs Fitted Values and Scale-Location Plots to assess the homoscedasticity and linearity assumptions; Normal Q-Q Plot to examine normality distribution of model residuals; Cook's Distance Plot and Residuals vs Leverage Plots to identify influential data points and detect any observations with high leverage and large residuals. Missing data were not imputed. Analyses were performed on an intention-to-treat base. For the non-inferiority comparison of the primary endpoint, SMA was considered non-inferior to CC if the lower limit of the 95% CI for the between-group difference was above the non-inferiority margin (−5pp). This is equivalent to one-sided non-inferiority testing with an alpha level of 0.025. For secondary endpoints, analysed according to a superiority framework, significance tests were based on two-sided tests with an alpha level of 0.05. In addition to intention-to-treat, we conducted a per-protocol analysis. Definitions for corresponding protocol deviation with observed frequencies and results of the per-protocol analysis are reported in ESM Tables 2 and 3. Results reported in the main manuscript refer to the intention-to-treat analysis. Data are presented as mean±SD or median [25th percentile; 75th percentile], unless stated differently. Analyses were conducted with R version 4.3.0 (The R Foundation for Statistical Computing, Vienna, Austria).