Study design

This was a randomised, double-blind (both participants and investigators were blinded to the participants’ treatment sequences), placebo-controlled, two-period, crossover intervention study performed at the Radboudumc (Nijmegen, the Netherlands). The study was approved by the local ethics committee and national competent authority and conducted in accordance with the principles of the Declaration of Helsinki, the Medical Research Involving Human Subjects Act, and applicable ICH Good Clinical Practice guidelines. All participants provided written informed consent.

Study population

Participants were recruited from the outpatient clinic of the Radboudumc and through websites of patient associations. Adults (aged ≥18 years) with type 2 diabetes were eligible for participation when they were treated with insulin for at least 1 year and had a serum magnesium concentration ≤0.79 mmol/l. Key exclusion criteria were HbA1c >100 mmol/mol (>11.3%), a total daily insulin dose of >2 U/kg body weight, MDRD-GFR <45 ml/min per 1.73 m−2, BMI <18 or >40 kg/m2, pregnancy, chronic diarrhoea and self-reported alcohol consumption >14 units weekly. People using magnesium supplements were not excluded from participation provided that they agreed to stop these at least 1 week before screening. Participant enrolment ran from March to August 2022 and we believe that our study sample is representative of the larger population of interest. Sex/gender and race/ethnicity were self-reported.

Study procedure

Participants first came for a screening visit, which included medical history, standard physical examination and BP measurement. Blood was sampled for determination of HbA1c, potassium, calcium, albumin, lipid profile and creatinine. Urine was sampled for determination of magnesium and creatinine.

After inclusion, participants were randomly assigned to treatment with magnesium gluconate or matching placebo for 6 weeks in a crossover design (electronic supplementary material [ESM] Fig. 1). Randomisation was done using a randomisation list (sequentially numbered) and performed by the pharmacy department, to ensure the double-blind nature of the study. The two treatment sequences were evenly distributed among the trial population using block randomisation. Magnesium gluconate and placebo were administered orally as a liquid solution of 50 ml three times a day, resulting in a daily dose of 150 ml (15 mmol magnesium, equivalent to 360 mg). Possible dose-related side-effects were monitored weekly by telephone consultation and the dose could be reduced step by step if necessary. If participants did not tolerate a minimum dose of 50 ml daily, they were withdrawn from study participation. During the final week of each treatment period, participants wore a blinded continuous glucose monitoring (CGM) device (Dexcom G6; Dexcom, San Diego, CA, USA) and activity tracker (activPAL micro; PAL technologies, Glasgow, UK) for 7 days [19]. On the day CGM and the activity tracker started, each participant’s BP was measured for 30 min using automated BP monitoring.

At the end of each treatment period of 6 weeks, participants underwent a hyperinsulinaemic–euglycaemic clamp (target glucose 5.5 mmol/l) [20]. Briefly, participants attended the research facility at 08:00 hours after an overnight fast, having abstained from alcohol, caffeine and smoking for 24 h and from strenuous exercise for 48 h. They were asked to interrupt or reduce their long-acting insulin dose 24 h before the clamp and omit their morning insulin dose and other glucose-lowering medication if applicable. After arrival, body weight was measured and the participants filled out a questionnaire that assessed potential symptoms of hypomagnesaemia [1, 21, 22] (ESM Methods). These potential symptoms included: muscle cramps, myalgia, muscle weakness, stiffness, tingling, restless legs, palpitations, fatigue, sleeping problems and difficulty concentrating. A urine sample was collected for measuring magnesium and creatinine. One catheter was inserted in retrograde fashion into a forearm vein for frequent blood sampling and this forearm was placed in a heated box (~55°C) to arterialise venous blood. Another catheter was placed into the antecubital vein of the contralateral arm for infusion of glucose 20% (wt/vol.; Baxter, Deerfield, IL, USA) and insulin (insulin aspart; Novo Nordisk, Bagsvaerd, Denmark). Baseline hyperglycaemia was corrected as needed with a mean i.v. bolus of 4 U of insulin (range 0–16 U). Subsequently, insulin was administered at a stable dose of 120 mU m−2 min−1, which is likely to fully suppress hepatic glucose production, and glucose 20% was administered at a variable dose to obtain stable euglycaemia (target glucose concentration of 5.5 mmol/l for 120 min), based on plasma glucose measured every 5 min using Biosen C-Line (EKF Diagnostics, Cardiff, UK). At baseline, blood was sampled for determination of HbA1c, magnesium, potassium, calcium, albumin, lipid profile, creatinine and insulin. After 90 and 120 min, blood was sampled to measure insulin.

Study outcomes

The primary outcome was the mean glucose infusion rate (GIR) during the final 30 min of the clamp (i.e. M value) [23]. Secondary outcomes included the insulin sensitivity index, calculated as the M value divided by the insulin concentration (M/I), mean daily insulin dose administered on the 3 days before the clamp, HbA1c, hypomagnesaemia-related symptoms determined by the questionnaire, mean BP measured for 30 min using automated BP monitoring and lipid profile, as well as CGM and activity tracker outcomes. CGM outcomes included mean glucose concentration, CV (i.e. SD divided by mean glucose concentration), occurrence of hypoglycaemic events, and time spent above range (i.e. glucose >10.0 mmol/l) and in range (i.e. glucose ≥3.9 and ≤10.0 mmol/l), as derived from CGM downloads. A hypoglycaemic event was defined as a glucose concentration <3.9 mmol/l for at least 15 consecutive minutes and a new event was calculated if the glucose concentration had been above this level for at least 15 min [24]. All CGM outcomes were calculated using R version 4.1.2 (R Foundation for Statistical Computing, Vienna, Austria). All raw activity tracker data were analysed by a modified version of the previously developed script of Winkler et al [25], using SAS version 9.4 (SAS Institute, NC, USA), and included sitting time, active time (i.e. standing and stepping) and number of steps. Physical activity intensity was classified as light (i.e. metabolic equivalent of task [MET] score <3) or moderate-to-vigorous (i.e. MET-score ≥3) [26].

Measurements

HbA1c was measured in EDTA whole blood on a Tosoh G11 in variant HbA1c mode with ion-exchange chromatography and absorbance detection (Sysmex, the Netherlands). Magnesium, lipid profile, calcium, albumin (bromocresol purple), creatinine (enzymatic) and potassium (ion-selective electrode) in plasma (and urine) were measured on a random access analyser Cobas 8000 (Roche Diagnostics, the Netherlands). Plasma insulin was measured using an in-house radioimmunoassay, using guinea pig anti-human insulin antibody and 125I-labelled human insulin tracer. In this assay, calibrated on WHO international standard 83/500, bound–free separation was performed by second antibody/polyethylene glycol precipitation of antibody-bound insulin [27].

Statistical analysis

To identify an increase of at least 20% in GIR after magnesium supplementation with a statistical power of 80% [15], we calculated that 14 participants would be required to detect a difference at a significance level of 0.05. We used random effects models to account for the two measurements for each participant with period and treatment as independent variables. Continuous variables were analysed using a multilevel mixed-effects linear regression model performing restricted maximum-likelihood estimation, and binary outcomes were analysed using a logistic random effects model. Data that were not normally distributed were log transformed. We performed a sensitivity analysis for data that were near normally distributed using the Wilcoxon signed rank test, which showed similar results to the mixed model. Correlations were analysed using Spearman’s rank-order test. No adjustments were made to account for multiple testing. Data were analysed using SPSS version 27 (IBM, NY, USA) and Stata version 17 (StataCorp, TX, USA). All data are expressed as mean ± SEM, unless otherwise specified. A p value <0.05 was considered statistically significant.