In this report, we describe the conduct and findings of the feasibility phase of GLINT, a large-scale clinical outcome trial to quantify the effects of metformin on the incidence of CVD and cancer among people with NDH and high CVD risk. We have demonstrated that most aspects of the trial design and conduct appear to be feasible, but some aspects will require adaptation for the main trial, drawing on the lessons learned and to ensure that there is sufficient power to reliably address the primary end point.
Acceptability of the study to general practitioners and patients
It is clear that primary care practitioners were sufficiently interested in the research question addressed in this study to become involved. Feedback at practice recruitment meetings confirmed that the topic was relevant and that there was uncertainty about management of the growing number of patients who would be eligible for the study, further justifying the study aims. In addition, the limited direct workload for the practices associated with participation was deemed appropriate. However, the delays introduced by sponsorship changes meant that initial enthusiasm had waned somewhat by the time that the initiation of the trial took place. The sponsor requirement that each general practice that recruited ≥ 20 participants would need to undergo a formal monitoring visit is unlikely to be feasible in the full trial.
The trial design, procedures, treatment and duration (individuals were recruited to a trial with an anticipated duration of > 5 years) appeared to be sufficiently acceptable to individuals such that 9.7% of those receiving a letter from their GP about the study attended the baseline assessment, signed a consent form and agreed to take part. This compares favourably with recent similar trials, such as A Study of Cardiovascular Events iN Diabetes (ASCEND),35 involving the distribution of the study drug by post. Once randomised, only one participant withdrew consent for follow-up for end points via their GP and routine data sources.
Efficiency of recruitment methods
Between March and November 2015, we recruited and obtained consent from ≥ 500 people to take part in the trial, 249 of whom were eligible and were randomised. As such, we demonstrated the feasibility of participant recruitment. However, simply scaling up the same recruitment procedures is unlikely to be feasible, as discussed in Recruitment from general practices. There were inefficiencies, in particular duplicate data entry, in using local databases to manage the information and study co-ordination prior to randomisation and using the InferMed MACRO system to host data entry and data storage and co-ordinate all participant contacts post randomisation. For the full trial, all study procedures should be co-ordinated by one central trial management system.
Recruitment from general practices
The majority of participants (61.4%) were recruited from general practices. Although the GLINT feasibility study was dependent on the co-operation of 10 sites and 21 PICs, the level of support required to recruit and manage those practices is not sustainable on a larger scale. Personal visits from the principal investigators to recruit practices, and technical support from the study team to assist with searching of electronic medical records and sending out invitations to potential participants, is not feasible for a UK-wide endeavour. Alternative approaches, such as mass mailshots within federations of practices and research-accredited pharmacies and liaisons with existing screening and prevention programmes, will need to be used.
The availability in electronic medical records of recent information concerning inclusion/exclusion criteria varied but for most variables this was < 50%. Qualifying laboratory test results were available for fewer than half of the participants but, when available, closely corresponded with the information obtained at the baseline visit. Consequently, in the early phase of recruitment, the proportion of individuals agreeing to take part who were subsequently randomised was significantly less than 50%. Following the application of a diabetes risk score search in the practices, and stratification of practice populations using the risk score, this proportion increased to 49% for the whole study (i.e. one out of two individuals who attended the baseline assessment at a study centre was randomised).
Recruitment via the NHS Health Check programme
The NHS Health Check programme did not contribute as many participants as had been expected. First, NHS Health Checks are centrally mandated but implemented by local authorities. Staff undertaking NHS Health Checks in Cambridgeshire and Leicestershire were focused on programme delivery and meeting targets. In 2015, they were not in a position to facilitate opportunistic recruitment to GLINT by referring individuals to the study centre for eligibility assessments or by providing eligible individuals with our brief information sheet and contact details. Fewer NHS Health Checks were conducted in recruited practices than expected. There have been differences in uptake between regions and it was noted that the uptake in Cambridge was lower than in Leicester. Between 2013 and 2015, in Leicestershire, the proportion of eligible people invited for a Health Check was 70%,36 whereas, in Cambridgeshire, only 62% of eligible people were invited during this period. The proportion of people who responded to the invitation for a Health Check was 38% in Cambridgeshire and 40% in Leicestershire. In Cambridgeshire, the NHS Health Check programme has been targeted at more deprived communities whose general practices are less likely to become involved in research. Without these NHS Health Checks, the information required to identify potentially eligible individuals was rarely available in the medical records.
Measurement of glycaemia, a key eligibility criterion in the feasibility phase of GLINT, has also varied because of local decision-making about how NHS Health Checks are delivered. A greater proportion of patients in the Leicestershire practices had a recent value for HbA1c level in their medical records than in the Cambridgeshire practices. The Health Checks diabetes filter is currently under consultation. The new recommendations include the use of existing diabetes risk scores, such as the Cambridge22 and Leicester37 diabetes risk scores, which, as we have shown previously38,39 and in this study, perform reasonably well in identifying individuals at high risk of diabetes and CVD.
Recruitment from research databases
We demonstrated the relative efficiency of recruiting from existing research databases by targeting invitations to those with known values for eligibility criteria; 8.4% of individuals identified by this route were eventually randomised. However, results from tests relating to eligibility criteria that were undertaken several years ago are unlikely to remain valid. Hence, the efficiency of this approach is, to some degree, dependent on how recently the individuals in the research database were assessed.
We demonstrated that information for key eligibility criteria, for example concerning renal function, was available in the general practice electronic medical records for nearly half of the randomised participants and that the values for variables extracted from the medical records were very similar to those obtained at the baseline visit at the study centres.
In conclusion, none of the recruitment strategies assessed would allow sufficient numbers to be recruited rapidly for a large outcome study. Although we will continue to explore links with NHS programmes to facilitate recruitment to GLINT (e.g. the NHS Health Check programme and the NHS Diabetes Prevention programme), we should not depend on NHS referrals to facilitate recruitment. Oxford investigators have experience of other recruitment methods using secondary care electronic health records, which have enabled thousands of patients to be randomised into cardiovascular outcome trials, and these methods will be adopted for recruitment into the main study. In addition, mass participant recruitment to trials from primary care is increasingly being enabled by organisations such as NorthWest Ehealth [http://nweh.co.uk/products/farsite (accessed 4 January 2018)], a not-for-profit partnership between the NHS and the University of Manchester. NorthWest EHealth undertakes searches of electronic medical records across hundreds of practices and then arranges mailshots to potentially eligible individuals, at a fraction of the cost per recruited participant of the feasibility study and requiring limited input from research and practice staff. Furthermore, much of this recruitment activity is eligible for NHS support cost funds. Given the low frequency of recent HbA1c values in the medical records, we would also propose dispensing with this as an inclusion criterion and, instead, identify people mainly on the basis of cardiovascular risk.
The original GLINT proposal was based upon modelled CVD risk using data from the EPIC-Norfolk study.32 The feasibility study demonstrated that the CVD risk of recruited participants (and, hence, the number of events) was likely to be lower than this. Furthermore, the adherence to metformin was lower than expected and observed in previous studies. Finally, the original sample size calculation was based on an effect size estimate of a 17% reduction in risk of events, which we updated to a 15% risk reduction following the feasibility study. For these reasons, we needed a larger sample size for the full trial than originally estimated. As part of our application for funds to support the full trial, we updated the sample size calculations as below.
Randomisation of at least 20,000 participants with at least 5 years’ follow-up and until at least 2700 unrefuted primary outcomes occur. This calculation is based on (1) an estimated hazard ratio with full compliance of 0.85, (2) approximately 78% compliance overall and (3) 90% power at 2p < 0.05.
One of the key logistical challenges in the feasibility study concerned the co-ordination of baseline assessments, which included obtaining consent. To scale up from 249 participants to around 20,000 participants, we propose to remove the need for face-to-face consultations to obtain consent for the majority of participants, instead utilising the internet, correspondence and telephone calls, as per A Study of Cardiovascular Events iN Diabetes (ASCEND).35
Randomisation
We have demonstrated the feasibility of the randomisation procedure, which efficiently generated groups that were well matched for the majority of baseline characteristics. We had originally planned to undertake focus groups at the end of the feasibility study in which we would have addressed the acceptability of the study procedures, including randomisation. However, the significant delays that we experienced in relation to changes in sponsorship meant that this was not possible. Among recruiting practices and participants, no concerns were raised relating to the issue and process of randomisation. Furthermore, randomisation was not referred to by any participants in the free-text response section of the questionnaire nor during the regular contacts with study co-ordinators. We have therefore inferred that the randomisation process was acceptable.
Characteristics of recruited participants
Participants were generally elderly (mean age 70 years) and predominantly male and overweight or obese. The majority were ex-smokers or current smokers and were prescribed antihypertensive medication. To some extent this reflects the variables included in the risk score used to identify potentially eligible individuals. Participants were predominantly white. We expected that recruited participants would be representative of the overall population in the areas we recruited from and that a larger proportion would have come from an Asian background. However, the participant demographics reflect those of the populations of recruited practices. In the case of the Leicester site, the surgeries that took part were located around the county of Leicestershire, rather than in the city centre. Recruiting general practices from more-deprived inner-city areas with greater ethnic diversity remains a challenge. The mean modelled 10-year CVD risk based on practice records was 30%, confirming that it is feasible to recruit individuals at high risk of CVD events. However, although CVD risk scores exhibit reasonable discrimination, their calibration is less good and modelled risk tends to overestimate observed risk. Furthermore, just over half of the participants were prescribed statins. Consequently, the expected number of actual CVD events is likely to be lower than estimates based on modelled risk. This has implications for the size of the main study and the characteristics of those who might be invited in order to maximise event rates.
Adherence to study medication
Delivering the study drug by post was feasible and efficient. A very small proportion of medication packs had to be re-sent. Concerns that the medication packs might not fit through some letter boxes did not materialise. However, there was a high level of discontinuation of the study drug within 6 months, although the frequency did not differ between study groups. In the DPP, a similar proportion of participants had discontinued metformin (Glucophage 850 mg, twice daily) at 6 months and this proportion remained stable for a further 2.5 years, at which point it declined to between 65% and 70%.12 In the DPP, adherence to placebo was consistently higher than adherence to metformin, which contrasts with our study, in which we used a prolonged-release preparation. Reports of AEs associated with the study medication were similar for metformin and placebo, although the side effects experienced by the placebo group were regarded as more ‘bothersome’. There were some anecdotal reports about the large size and unpleasant texture of the study medication. Furthermore, among those participants who discontinued the study medication, a greater proportion in the metformin group than in the control group attributed side effects as the reason for discontinuation. Among participants taking the study drug, most were able to tolerate the full 1500-mg dose per day and had taken their medication on 85% of the preceding 14 days. Nevertheless, the level of discontinuation of the study drug in both groups suggests the need for a pre-randomisation run-in period in the full trial.
Pre-randomisation run-in to improve treatment adherence, participant retention and questionnaire response rates
For the feasibility study, we did not include a pre-randomisation run-in period but will do so for the main study. The main advantages of this are twofold. For interventions that are not universally well tolerated, such as metformin, a run-in period allows all potential participants to take the study drug before being randomised so that only those people who tolerate metformin at the end of the run-in period are randomised. This increases the number of people who need to be screened but has been shown to improve post-randomisation adherence. Second, the run-in period allows the exclusion of those people who are initially enthusiastic about trial participation but who then lose interest and stop complying with the study procedures (e.g. questionnaire completion) and are more likely to drop out post randomisation, resulting in a serious impact on study power.
Follow-up by questionnaire and electronic records/registers
We obtained appropriate consent for the use of routine data and registers for the tracking of outcomes and, although we did not pursue the formal application for HSCIC (now NHS Digital) approvals, we have demonstrated that such an approach is feasible in parallel studies.31 We also demonstrated the feasibility of collecting sufficient data for independent adjudication concerning potential end points. We demonstrated the feasibility of collecting data on functional status, health utility and health service use. Questionnaire response rates were reasonable at 4 months but fell to 75% by the end of the study, although, for a significant proportion of participants, the end-of-study questionnaire coincided fairly closely with a previous questionnaire. It is clear that using internet-based data collection instruments is not feasible for all participants and hence postal paper questionnaires would still be required. The declining response rate underlines the value of collecting outcome data using routine sources, of including study procedures in a run-in phase in a main trial and of maintaining contact with study participants through newsletters.
Safety monitoring
In the feasibility study, we included study centre visits at 3 and 6 months, mainly for the purpose of collecting blood samples to monitor safety. Attendance levels were high (86% at 6 months). Given the small sample size, the results of analyses of differences between study groups should be interpreted with caution.
Renal function and risk of lactic acidosis
There were small declines in renal function over 6 months, as one might expect in this study population with a mean age of 70 years. However, there were no differences between groups. The concern in relation to metformin has not been that it is associated with renal AEs, but rather that, in the presence of impaired renal function, the risk of lactic acidosis, a rare but serious condition, is increased. This remains unproven but, nevertheless, significantly impaired renal function is a contraindication to the use of metformin. During the course of the feasibility study, the SmPC was changed and stipulated that metformin should not be prescribed if the eGFR is < 45 ml/minute/1.73 m2 and that eGFR should be monitored. Also during the course of the feasibility study, a systematic review40 was published demonstrating the safety of metformin among people with impaired renal function. The review showed that drug levels generally remain within the therapeutic range and lactate concentrations are not substantially increased when used in patients with mild-to-moderate CKD (eGFRs of 30–60 ml/minute/1.73 m2). Furthermore, the overall incidence of lactic acidosis in metformin users (3–10 per 100,000 person-years) was similar to the background rate in the population with diabetes. A recently conducted placebo-controlled trial41 of 173 patients, in which an eGFR of < 45 ml/minute/1.73 m2 was an exclusion criterion, showed no difference in lactate levels over 18 months. A Cochrane Database Systematic Review by Salpeter et al.42 and a meta-analysis of English and non-English literature by the same authors43 reported that the risk of lactic acidosis is essentially nil in the context of clinical trials of metformin, including those that did not specify kidney disease as an exclusion criterion. It is apparent that patients who develop lactic acidosis while taking metformin typically have an acute supervening illness, for example sepsis, acute kidney or liver failure or cardiovascular collapse, which precipitates the metabolic decompensation causing lactic acidosis. Therefore, even though metformin is renally excreted, and its clearance is impaired in mild-to-moderate CKD, drug levels are still largely maintained within a therapeutic range when the eGFR is > 30 ml/minute/1.73 m2 and do not seem to significantly affect circulating lactate levels. Observational studies40 suggest a potential benefit of metformin on macrovascular outcomes, even in patients with renal contraindications for its use. The recognition of metformin’s safety has led to its common use in patients with CKD.
There has been increasing pressure for regulatory agencies to relax their guidelines regarding metformin prescribing in renal impairment, with an updated joint position statement44 issued by the American Diabetes Association and the European Association for the Study of Diabetes in 2015 suggesting that the renal safety guidelines in the USA may be ‘overly restrictive’. Following a comprehensive review of the medical literature, the US Food and Drug Administration issued updated guidance45 in April 2016 that relaxed the renal threshold for metformin to allow its use in patients with an eGFR of > 30 ml/minute/1.73 m2, with consideration of dose reduction and monitoring of renal function in those patients at highest risk. The European Medicines Agency issued similarly updated guidance46 in October 2016 to allow metformin use in patients with an eGFR of > 30 ml/minute/1.73 m2.
Having considered the data on changes in renal function from the feasibility study and the totality of evidence for the safety of metformin in patients with impaired renal function, we intend to revise the entry criteria for the full trial to exclude only those with severe renal dysfunction, defined as requiring ongoing follow-up in a specialist nephrology clinic, or with an eGFR of < 30 ml/minute/1.73 m2 and we will limit the maximum dose of the study drug to two tablets (equivalent to 1000 mg of metformin) for those participants aged > 80 years or with an eGFR of 30–45 ml/minute/1.73 m2. Interim monitoring in a subset of patients at high risk of progression of renal dysfunction is being considered in the revised trial design.
Vitamin B12
Metformin was not associated with reductions in plasma vitamin B12 levels over 6 months. At baseline, 6.83% of participants had a vitamin B12 level below the laboratory reference range and this had fallen to 0.94% at 6 months. Changes in vitamin B12 levels have been reported among people with insulin-treated T2D receiving a mean dose of 2050 mg per day of metformin.47 However, it is uncertain whether or not these findings are generalisable to people with NDH who are treated with a lower dose. Previous studies47,48 have shown that metformin is associated with reductions in serum vitamin B12 levels, but these have remained within the normal range and have not been associated with adverse health outcomes such as anaemia, neuropathy or reduced quality of life. Based on these data, it would be premature to introduce vitamin B12 monitoring for all T2D patients receiving metformin48 or for participants in our trial. We have no robust data showing that vitamin B12 monitoring in individuals receiving metformin would yield clinical benefits.
Cardiovascular risk factors, health utility and functional status
Compared with placebo, treatment with metformin was associated with small improvements in the following cardiovascular and metabolic risk factors over 6 months: HbA1c, ALT, total cholesterol, LDL cholesterol and triglycerides. Changes in HbA1c level were consistent with those observed in the DPP12 and the more recent Carotid Atherosclerosis: MEtformin for Insulin ResistAnce Study (CAMERA).41 Any observed reduction in the risk of CVD and cancer attributable to metformin is unlikely to be solely as a result of its effect on glycaemia. We showed that metformin was not associated with any detrimental effects on health utility or functional status within 4 months. One limitation of the feasibility study is that we did not include follow-up measures of weight and blood pressure, although the small sample size would have constrained interpretation of between-group differences.
Serious adverse events
Systems for collecting information about SAEs and reporting them to the sponsor and IDMC functioned well, although the number of events was small. A greater number of SAEs were reported by participants in the metformin group than by participants in the placebo group; however, none was considered to be related to the study medication.
Ensuring an adequate event rate for study power
The feasibility study was not expected to provide information on cardiovascular event rates but the types of patient recruited suggest that, if limited to a primary prevention population, the cardiovascular event rate is likely to be much lower than the 2% that was originally modelled.
In our original proposal we suggested recruiting some individuals with pre-existing CVD and some without. However, some reviewers and funders strongly recommended that GLINT should be a primary prevention study. We believe that the main study question, quantifying the effects of metformin on the risk of CVD events, can be feasibly answered only by recruiting mainly secondary prevention patients with a prior history of CVD. We intend to revise the GLINT inclusion criteria to recruit a cohort of around 80% secondary CVD prevention participants and 20% primary CVD prevention participants.
People with pre-existing CVD are at higher CVD risk, and are more easily identified in both primary and secondary care and recruited to studies, than individuals with a high modelled risk but no prior history of CVD. The Oxford Clinical Trial Service Unit, with whom we plan to work on the main study, has extensive experience of identifying high-risk patients with CVD from hospital-based electronic records. In the 1990s, it randomised 20,000 patients using these methods in the Heart Protection Study49 and, in subsequent years, 12,000 UK post-myocardial infarction patients were randomised in SEARCH (Study of the Effectiveness of Additional Reductions in Cholesterol and Homocysteine),50 8000 UK patients (out of 25,000 worldwide) were randomised in the HPS2-THRIVE (Treatment of HDL to Reduce the Incidence of Vascular Events) study51 and 8000 (out of 30,000 worldwide) were randomised in the HPS3-REVEAL (HPS3/TIMI55-REVEAL Randomized EValuation of the Effects of Anacetrapib through Lipid-modification) study.52 We will use similar methods, alongside primary care registers, to recruit approximately 20,000 patients to be randomised cost-effectively into GLINT.
