Substudy 1 is a retrospective cohort analysis of all Omani children with T1DM who received CGM through the national rollout between July 2024 and February 2025 across the 11 governorates. The minimum analytic exposure window is 3 months of CGM use from baseline.

T1DM in children is heterogeneous, and trajectories of glycemic control depend on age at onset, family history, comorbidities, and self-management capacity [11-13]. CGM has been shown to improve glycemic control in children when wear is consistent [14-16,19]. Quantifying baseline characteristics and short-term post-CGM outcomes in the Omani population will inform both clinical follow-up pathways and the variable set used to train OMNIdiasense.

Primary outcomes of substudy 1 are (1) the population-level prevalence of CGM optimizer status at ≥3 months and (2) the pre-post change in HbA_1c. Secondary outcomes are pre-post changes in weight, BMI, BMI z score, systolic and diastolic blood pressure, lipid panel, renal function test (RFT), and liver function test (LFT); regional variation across governorates; and identification of routinely captured clinical predictors of optimal use that will be carried forward to the AI model in substudy 3.

Regional pharmacists, supported by their RMs, will extract the following data from Al Shifa for each eligible child into a structured Google Form (Multimedia Appendix 1):

Descriptive statistics will be presented as frequencies, percentages, means, and SDs or medians and IQRs, as appropriate. Pre-post comparisons will use paired t tests or Wilcoxon signed-rank tests for continuous variables and McNemar tests for categorical variables. Differences across governorates will be evaluated with chi-square tests and ANOVA. Logistic regression will model predictors of optimal CGM use. Analyses will be performed in IBM SPSS Statistics (version 28) and R software (version 4.4; R Foundation for Statistical Computing). Reporting will follow STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) guidelines (Multimedia Appendix 2).

Substudy 2 is a cross-sectional mixed methods study using supervised face-to-face interviews with adolescents aged 10-18 years drawn from substudy 1. RMs will randomly sample participants from the Al Shifa–derived cohort. The dependent variable, pattern of CGM use, will be dichotomized as CGM optimizers (≥6 days/week) and CGM subusers (<6 days/week or discontinued), consistent with prior pediatric work [25].

Eligible participants will include children aged 10-18 years with T1DM who received CGM under the national initiative and whose primary caregiver consents. Children will be excluded if they are unable to complete the structured interview owing to severe cognitive or sensory impairment.

Despite CGM’s technical maturity, the majority of adolescents with T1DM do not meet glycemic targets [20-24]. Self-management decisions are shaped by perceived benefit and burden of CGM, self-efficacy, coping responses, diabetes distress, dietary patterns, and physical activity [25,27]. Substudy 2 measures these constructs and their open-ended elaborations to inform feature selection for OMNIdiasense.

Primary outcomes of substudy 2 are the standardized mean differences and odds ratios for each psychosocial and behavioral instrument (Benefits of Continuous Glucose Monitoring Scale [BenCGM], Burden of Continuous Glucose Monitoring Scale [BurCGM], Response to Stress Questionnaire-Pediatric Diabetes [RSQ-PD] coping factors, Self-Efficacy for Diabetes Management Scale [SEDM], Problem Areas in Diabetes-Pediatric version [PAID-Peds], Physical Activity Questionnaire for Adolescents [PAQ-A], and dietary screener) between optimizers and subusers. Secondary outcomes are the qualitative themes generated from open-ended responses, the convergent mixed methods joint display, and the ranked feature list passed to substudy 3 for model training (Multimedia Appendix 3).

Assuming an optimizer prevalence of approximately 50% [25] and aiming to detect medium-effect-size predictors in a multivariable logistic model with up to 8 predictors, a minimum of 240 participants will provide sufficient statistical power (≥10 events per variable). Recruitment will be stratified across the 11 governorates proportional to CGM dispensations.

The face-to-face interview will include the following components, administered in Arabic where validated translations are available (Multimedia Appendix 4). Questionnaires that require official approval from the original authors or involve payment or restricted use will be obtained accordingly. Components of the questionnaire include the following:

Quantitative analyses will compare optimizers and subusers using chi-square tests, t tests, and Mann-Whitney U tests, with multivariable logistic regression to identify independent psychosocial and behavioral predictors of optimal CGM use. Qualitative responses to open-ended items will be audio-recorded with consent, transcribed verbatim, and analyzed in NVivo (Lumivero) by 2 independent coders (1 bilingual Arabic-English nurse researcher and 1 public health researcher) using a hybrid inductive-deductive approach; discrepancies will be resolved by a third senior coder, and a 20% subsample will be member-checked for accuracy. Field notes will be written within 24 hours of each interview. Findings will be triangulated using a convergent mixed methods design with joint displays. Reporting will follow STROBE (Multimedia Appendix 2) for the quantitative component and COREQ (Consolidated Criteria for Reporting Qualitative Research; Multimedia Appendix 5) for the qualitative component.

Substudy 3 evaluates the AI-assisted assessment tool OMNIdiasense through 2 parallel components, both fed by an embedded AI model-development workstream described below.

All 150 participants are followed for 12 months, with assessments at baseline, 3, 6, and 12 months. The flow of participants is presented in Figure 1.

The AI tool will be codeveloped by the MoH digital-health team and computer scientists from Sultan Qaboos University. The development pipeline comprises 7 steps, expanded below to address missing data, class imbalance, overfitting, feature selection, calibration, fairness, and external validation.

Missingness will be quantified by variable and pattern. Variables with >40% missingness will be excluded unless they carry strong prior clinical value, in which case they will be retained with an explicit missing indicator. Item-level missing data on validated scales will be imputed using each scale’s published rules; remaining missingness will be addressed by multiple imputation by chained equations (m=20) under a missing-at-random assumption, with a tipping-point sensitivity analysis to test robustness against a missing-not-at-random scenario.

Class imbalance between optimizers and subusers will be addressed at 2 levels: stratified sampling during cross-validation to preserve outcome proportions in every fold, and class-weighted loss functions (inverse-frequency weights) for the gradient-boosted, SVM, and neural-network classifiers. Synthetic minority oversampling will be evaluated as a sensitivity analysis only and not as the primary approach, to avoid optimistic generalization.

Overfitting will be controlled through nested cross-validation (outer 5-fold for performance estimation, inner 5-fold for hyperparameter tuning), L1/L2 regularization for linear models, early stopping and shrinkage for boosted trees, and dropout and weight decay for the neural network. The full train-test partition (80/20) will be temporally separated where possible (children enrolled later become the held-out set) to give a more realistic estimate of prospective performance.

Feature selection will use a stability-selection wrapper around L1-penalised logistic regression with 100 subsampled fits; features selected in ≥60% of subsamples will be retained, in addition to a clinically prespecified core set (age, sex, diabetes duration, baseline HbA_1c, insulin modality, governorate, BenCGM, BurCGM, SEDM, and PAID-Peds). The final model will be sparse and interpretable; Shapley additive explanations (SHAP) values will be reported to support clinical interpretation of individual predictions.

Discrimination alone is insufficient when a model is used to triage clinical support. Calibration will be assessed using calibration plots, the Brier score, calibration-in-the-large, and calibration slope. If the raw model is miscalibrated, isotonic regression or Platt scaling will be applied to the inner cross-validation folds and reevaluated on the outer fold. Decision-curve analysis across plausible threshold probabilities will accompany the calibration report.

Fairness will be evaluated as a primary model-quality dimension, not as an afterthought. For each protected attribute (sex, age band 10-13 years vs 14-18 years, and governorate), we will report subgroup-specific area under the curve, sensitivity, specificity, calibration slope, and false-negative rate (the most consequential error, since a false negative could deny behavioral support to a child who needs it). Equality of opportunity and predictive parity will be reported. If a clinically meaningful disparity is detected (absolute difference in false-negative rate >10 percentage points), we will apply group-specific threshold adjustment and reevaluate before deployment.

Internal validation alone is insufficient before clinical use. The validation plan therefore includes three layers: (1) internal validation on the held-out 20% partition, (2) temporal validation on a prospectively collected cohort enrolled at least 3 months after model lock, and (3) a preplanned external multisite replication in 2 Gulf neighboring sites (institutions to be confirmed at the start of the pilot phase) using the same data-collection forms and the same outcome definitions. Until external validation is satisfactory, OMNIdiasense outputs will be presented to clinicians as decision support only and will not influence CGM allocation (see the Ethical Considerations section).

Reporting of the AI development will follow the TRIPOD+AI (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis + Artificial Intelligence) guideline [44,45], which provides a 27-item checklist for prediction-model studies that use traditional regression or AI/machine learning methods. The TRIPOD+AI checklist will be uploaded with the results paper.

OMNIdiasense will generate a per-patient adherence-risk score with personalized, MI-aligned recommendations covering meal timing, physical activity, stress management, and parental engagement. The system will be piloted in Muscat for 6 months prior to wider testing.

The pilot RCT sample size was based on the prevalence of optimal CGM use of approximately 50% [25] and a moderate effect (odds ratio 0.53) reported for adolescent CGM use. Using G*Power 3.1 (Heinrich Heine University Düsseldorf), at least 25 participants per arm are required to achieve 80% power at α=.05. The single-arm component (n=100) is powered for paired-sample within-person change in adherence and HbA_1c.

Quasi-experimental component: existing children with T1DM aged 10-18 years classified as CGM subusers in substudy 2; primary caregiver consents and child assents.

RCT component: newly diagnosed children with T1DM, aged 10-18 years, eligible for CGM under the national program, who have not yet received CGM. Exclusion: severe psychiatric or cognitive comorbidity precluding informed assent or interview participation.

Eligible families will receive a formal invitation letter, information sheet, and consent/assent forms (Multimedia Appendix 6) from RMs, who will explain study objectives and logistics in Arabic and English.

A 3-day bespoke training program will be delivered to RMs, diabetes nurses, and administrative staff, covering the OMNIdiasense interface, MI principles, outcome measurement procedures, and ethical considerations.

Participants in the quasi-experimental arm and the RCT intervention arm will receive 3 MI-guided consultations delivered by trained diabetes nurses at baseline, 3 months, and 6 months (Multimedia Appendix 7). MI sessions will (1) establish rapport, (2) explore pros and cons of CGM use, (3) provide tailored feedback on the OMNIdiasense risk score, (4) elicit imagery of life with and without CGM, (5) review options and develop an adherence plan, and (6) reinforce self-efficacy.

Permitted concomitant care includes routine pediatric diabetes follow-up, school health support, and standard insulin therapy. Discontinuation of OMNIdiasense recommendations will be considered if the family withdraws consent, if a clinical adverse event makes continued sensor wear unsafe, or if the treating clinician determines that a recommendation conflicts with safety. Adherence to the MI fidelity manual will be assessed by audio-recorded session sampling (approximately 10% of sessions per interventionist per quarter), scored using a brief MI fidelity rubric (Multimedia Appendix 8).

The primary outcome is the between-group difference in CGM adherence, defined as mean days per week of sensor wear over the 12-month follow-up. The CGM adherence outcome will be measured from pharmacy dispensing records, self-report, and clinical records. Differences between device types will be recorded, and significant findings across different devices will be analyzed.

Key secondary outcomes are change in HbA_1c, time in range, anthropometric measures (weight and BMI), blood pressure, lipid panel, RFT, LFT, drug/treatment changes, OMNIdiasense risk score change, barriers and facilitators of CGM use, and patient satisfaction (Multimedia Appendix 9). Table 2 shows the full schedule of assessments.

In the RCT component, computer-generated 1:1 simple randomization will be performed centrally by an independent statistician. Allocation concealment will use sequentially numbered, opaque, sealed envelopes opened only after enrollment confirmation. Outcome assessors and the data analyst will be blinded to allocation; participants and intervention deliverers will not be blinded owing to the nature of the intervention.

Analyses will follow the intention-to-treat principle. Descriptive statistics will summarize baseline characteristics. Within-person change in the quasi-experimental arm will be analyzed with paired t tests, Wilcoxon signed-rank tests, or generalized estimating equations as appropriate. Between-group differences in the RCT will use t tests, chi-square tests, and mixed effects models with random intercepts for repeated measures. Multivariable linear and logistic regressions will explore predictors of adherence change. Missing data will be handled by multiple imputation under a missing-at-random assumption with sensitivity analyses. Prespecified subgroup analyses will examine age band (10-13 vs 14-18 years), sex, governorate (Muscat vs others), and insulin modality (multiple daily injections vs pump); any post hoc subgroup analyses will be flagged as exploratory. The pilot RCT report will follow the CONSORT (Consolidated Standards of Reporting Trials) 2025 extension for pilot and feasibility trials, and the interventional protocol adheres to SPIRIT (Standard Protocol Items: Recommendations for Interventional Trials)-AI 2025 (Multimedia Appendix 10).

All adverse device effects (eg, persistent skin reactions and false hypoglycemia alarms with clinical sequelae) and serious adverse events will be solicited at every study visit and through a dedicated 7-day reporting channel. Events will be graded using the United States Food and Drug Administration Medical Device Reporting framework, reported to the DMC, and forwarded to the regional patient safety and quality departments. Independent audits of consent, source data, and protocol adherence will be performed at 6 and 18 months by the National Center of Studies and Research.

Substantive protocol amendments will be submitted to the MoH ethics committee for evaluation and approval. Prespecified stopping rules include (1) a device-related serious adverse event rate of 10% or greater, (2) a DMC recommendation on safety or futility grounds, or (3) interruption of national CGM supply.

OMNIdiasense generates a per-patient probability of becoming a CGM subuser. Throughout the studies described in this protocol, this score is used only as a decision-support adjunct to trigger MI-guided behavioral support and to personalize the content of that support. The score will not be used to deny, delay, or withdraw CGM access. CGM allocation will continue to follow the existing national clinical pathway determined by the treating pediatric diabetes team; the score is recorded in the research database, not in the patient’s clinical record, during the pilot phase.

Risk predictions will be communicated using verbal categorical language (“higher,” “moderate,” and “lower”) and not as raw probabilities. The communication template, developed with input from the parent advisory group, includes (1) an explanation that the score is a research-stage estimate, not a diagnosis, (2) an explanation that the score does not affect CGM access, (3) the actionable behavioral recommendations linked to that score, and (4) the family’s right to decline the recommendations without consequence for routine care. Counseling is delivered by the trained diabetes nurse, with the treating clinician available for clinical questions.

Identifiable data are stored only within the MoH Al Shifa environment behind the national firewall. The analytic dataset shared with the AI team is pseudonymized: the link key is held by the principal investigator in a separate encrypted store. Access is role-based and audit-logged. Model training and inference occur in a MoH-controlled environment; no patient data leaves the national network for model training or inference. Data sharing for external validation in Gulf neighboring sites will use federated analysis (aggregate parameter exchange) rather than transfer of individual-level records, where technically feasible, and otherwise a data transfer agreement compliant with Omani Personal Data Protection Law (Royal Decree 6/2022) and the host country’s equivalent will govern any transfer [46].

Subgroup performance (sex, age band, governorate, and insulin modality) will be reviewed by the DMC at 6-month intervals during the pilot. If a meaningful subgroup disparity is detected, deployment of model outputs for that subgroup will be paused while the model is recalibrated, and the DMC will advise on whether to resume. A bias monitoring log will be kept and included in the results paper.

Beyond the pilot, any future implementation will be conditional on (1) satisfactory external validation, (2) ethics reapproval, and (3) an algorithmovigilance plan covering version control, drift monitoring (population shift in predictors and outcomes), threshold review, complaint handling, and a clearly documented decommissioning pathway. The decommissioning pathway includes the conditions under which the tool would be withdrawn (eg, calibration drift, disparity, and supersession by a better tool) and the steps to remove it without disrupting routine care.