Glycemic Velocity and Early Retinal Microvascular Change With GLP-1RA Versus SGLT2i Initiation in Type 2 Diabetes (GLIDE) (GLIDE)

Background and rationale:

GLP-1 receptor agonists (GLP-1RAs) and sodium-glucose cotransporter-2 inhibitors (SGLT2is) are central therapies for type 2 diabetes, and their ocular safety remains unresolved. The cardiovascular outcome trial SUSTAIN-6 reported more adjudicated diabetic-retinopathy complications with semaglutide than placebo, concentrated among patients with pre-existing retinopathy and concurrent insulin use, whereas large real-world datasets have not reproduced a consistent signal. Mechanistic and meta-analytic work increasingly attributes the trial signal not to the molecule itself but to the magnitude and rapidity of the accompanying fall in glycated hemoglobin (HbA1c), the long-recognized phenomenon of transient "early worsening" of retinopathy after rapid glycemic correction, first characterized in the Oslo study and subsequently in the DCCT.

Central hypothesis and definition of the exposure:

The study formalizes the driver of early worsening as a continuous, patient-level exposure termed glycemic velocity: the rate at which HbA1c falls after treatment initiation, computed as (HbA1c at baseline - HbA1c at Month 3) divided by elapsed time and expressed in percentage points per month. It is analyzed per +1 SD, with a clinically anchored secondary scale of +0.5 percentage points per month and a pre-specified test for non-linearity or a threshold effect. Baseline HbA1c and the absolute magnitude of HbA1c reduction are retained as separate covariates so that the effect of rate is distinguished from that of magnitude and of starting level. The complementary mechanistic hypothesis is that a transient disturbance of retinal and choroidal perfusion (relative hypoxia) is the shared intermediate underlying GLP-1RA-associated ocular signals, and that its magnitude tracks with glycemic velocity.

Rationale for the design:

Drug choice is made by the treating physician on clinical grounds; the study assigns, provides, and alters no therapy, adding only observation, imaging, and data collection. A new-user, active-comparator structure is used because enrolling only treatment-naive initiators removes prevalent-user and immortal-time biases, while an active comparator rather than non-users minimizes confounding by indication and the healthy-user effect. SGLT2i was selected as the comparator because it shares the glucose-lowering indication and a broadly comparable position in contemporary treatment algorithms, yet is not classically implicated in early retinopathy worsening, providing the contrast needed to separate a drug-specific retinal effect from the drug-independent consequence of rapid glycemic correction. Because glycemic velocity is a continuous measured variable present in both groups, the primary hypothesis remains testable irrespective of how patients were allocated to drug class. Enrollment into each group is capped to preserve balance for the between-group comparison. Both eyes are imaged, with the eye-visit as the unit of observation and statistical accounting for inter-eye correlation.

Visit schedule and data acquisition:

Patients are identified at the point of treatment initiation and referred the same day, or within the permitted baseline window, for screening and baseline imaging. Assessments occur at Baseline (V0), Month 1 (V1, plus or minus 2 weeks), and Month 3 (V2, plus or minus 2 weeks), with a pre-planned companion analysis of longer-term trajectory over an extended follow-up described in the protocol. All imaging is performed on a single OCT-angiography/OCT platform by trained operators using a fixed acquisition protocol, and only scans meeting a pre-specified signal-strength and artifact threshold are analyzed. Quantitative metrics are produced by validated device software, with choroidal indices derived by standardized image binarization. Two independent graders, masked to drug class, glycemic data, and visit sequence, perform ETDRS grading and quality control, with senior adjudication of discrepancies; inter- and intra-grader reliability is quantified on a randomly selected double-graded subset using intraclass correlation coefficients and weighted kappa, against a pre-specified target. Data are captured in REDCap with role-based access rights and a full audit trail, and the metabolic and imaging datasets are reconciled only at the analysis stage.

Sample-size justification:

Enrolment is governed by the between-group contrast in the primary outcome: detecting a difference of 2.0 percentage points (SD 3.5; Cohen's d approximately 0.57) at a two-sided alpha of 0.05 with 80% power requires approximately 49 participants per group. This exceeds the requirement for the primary association (approximately 62 participants for a partial correlation of r = 0.35, by Fisher's z-transformation) and therefore governs, and is inflated by approximately 20% for anticipated attrition and ungradable imaging. Repeated measures and the use of both eyes further increase effective precision. A non-inferential feasibility review after 40 participants complete Month 3 assesses imaging gradability, recruitment rate, and the plausibility of the assumed effect sizes; no efficacy stopping rule applies.

Statistical analysis:

The primary analysis fits a linear mixed-effects model of the primary outcome over time, with the glycemic velocity by time interaction as the principal test, adjusted for drug class and for pre-specified confounders (age, sex, diabetes duration, baseline retinopathy severity, baseline HbA1c, magnitude of HbA1c reduction, blood pressure, eGFR, axial length, and OCT-angiography signal strength), and with random intercepts for participant and for eye nested within participant. The model uses all available eye-visit observations under a missing-at-random assumption. A causal mediation analysis in the counterfactual framework decomposes the total effect of drug class on the primary outcome into a natural indirect effect transmitted through glycemic velocity and a natural direct, unmediated effect; exposure-mediator interaction is permitted, and inference uses bootstrap confidence intervals with E-value sensitivity analysis for unmeasured mediator-outcome confounding. Baseline balance between groups is summarized using standardized mean differences rather than significance tests, and propensity-score methods (overlap weighting or matching) serve as a confounding-control sensitivity analysis. Pre-specified effect modification by baseline retinopathy severity and by insulin co-therapy is tested through interaction terms and interpreted as hypothesis-generating. Analyses follow a statistical analysis plan finalized and signed before database lock, are conducted in R or Stata, and are reported in accordance with STROBE and, for the mediation component, AGReMA.