Canagliflozin, dapagliflozin and empagliflozin monotherapy for treating type 2 diabetes: systematic review and economic evaluation

Sulfonylureas

The SUs are insulin secretagogues, which means that they work largely by stimulating insulin release by the beta cells in the pancreas. Once the beta cell capacity falls, the SUs become less effective. There is some evidence that the duration of effectiveness is longer with gliclazide than glibenclamide. 17

The main AEs of the SUs are weight gain and hypoglycaemia. A population-based study from Tayside found an incidence of severe hypoglycaemia amongst people on SUs of 0.9 per 100 patient-years. 18 This rate is similar to the 0.8% seen in the meta-analysis by Schopman et al. 19 Monami et al. ,20 in a good-quality meta-analysis of 69 trials involving SUs, reported a cumulative incidence of at least one episode of severe hypoglycaemia of 1.2%, but this was based on 24 trials because the others did not report any severe hypoglycaemia. There was some evidence that hypoglycaemia was less common with gliclazide than with other SUs. 20 Schopman et al. 19 reported that 0.1% of patients on gliclazide had severe hypoglycaemia and that 1.4% had plasma glucose (PG) under 3.1 mmol/l at some point in trials that ranged in duration from 24 to 104 weeks. Schernthaner et al. 21 from the 27-week GUIDE (GlUcose control in type 2 diabetes: Diamicron MR vs. glimEpiride trial), using modified-release (MR) gliclazide, reported that 3.7% of patients had at least one PG of < 3 mmol/l, but that none needed assistance. Compared with the glimepiride arm, there were about 50% fewer hypoglycaemic episodes, despite a reduction in HbA_1c of 1.2% on gliclazide and 1.0% on glimepiride.

The Schopman meta-analysis19 reported that, overall, 0.8% of patients on SUs had a severe hypoglycaemic episode, but the proportions ranged from 0.1% for gliclazide to 2.1% for glipizide. In the ORIGIN trial, 75% of patients on standard treatment (25% of whom were on SUs) never had any hypoglycaemia. 22

In the very large (11,140 patients) ADVANCE (Action in Diabetes and Vascular Disease: Preterax and Diamicron Modified Release Controlled Evaluation) trial, gliclazide MR was used in two arms, intensive and standard. In the intensive arm, the aim was to achieve HbA_1c of 6.5% or less. 23 This was achieved in 65% in the intensive arm and 29% in the standard arm. Severe hypoglycaemia event rates were 0.07 per 1000 patient-years in the intensive arm and 0.04 per 1000 patient-years in the standard arm. Minor hypoglycaemic events occurred at rates of 12 and 9 per 1000 patient-years in intensive and standard arms, respectively.

These rates of hypoglycaemia on SUs are much lower than the 7% reported for severe hypoglycaemia by the UK Hypoglycaemia Study Group,24 but the patients in that study were recruited only from secondary care clinics.

In the Netherlands, the guideline for the management of type 2 diabetes advises that gliclazide is the SU of choice, partly because of its safety in renal failure. 25,26 A meta-analysis of SU trials concluded that severe hypoglycaemia was rare with gliclazide, especially if the dose does not exceed 240 mg daily. Non-severe hypoglycaemia was seen mainly in those on 320 mg daily. 25

Simpson et al. 27 argued that as different SUs had different tissue selectivity and risk of hypoglycaemia, the cardiovascular risk might also vary. They carried out a systematic review and NMA, and used glibenclamide as the reference risk. Compared with people taking glibenclamide, those on gliclazide had a relative risk for total mortality of 0.65 [95% credible interval (CrI) 0.53 to 0.79]. For cardiovascular mortality, the risk ratio (RR) for gliclazide was 0.60 (95% CrI 0.45 to 0.84), whereas other SUs showed no significant difference from glibenclamide.

Schramm et al. 28 used Danish record linkage data to compare the mortality and cardiovascular risks amongst patients on monotherapy with SUs and repaglinide, with those on metformin. The risks were higher on most SUs, but not for gliclazide or repaglinide.

The risk of severe hypoglycaemia with SUs may have been overestimated, but it remains a problem that can lead to hospital admission, as well causing anxiety and interrupting usual activities.

The Scottish Intercollegiate Guidelines Network (SIGN) recommends that SUs should be considered as first line in patients who cannot take metformin. 29 The 2015 American Diabetes Association (ADA) position statement expresses no preference amongst SUs, pioglitazone, flozins and gliptins in people who cannot take metformin. 30

If SUs were the same price as the newer drugs such as the gliptins or the flozins, they would probably be superceded. But they are very cheap, and have been used for so long that all their AEs are known.

In this report, based on the evidence reported above, we use gliclazide as the SU of choice. There are two forms of gliclazide, standard and MR. The Diamicron Study Group reported these to be clinically equivalent in a 10-month study in 800 patients. 31 The MR form was given once a day, and 30–120 mg was equivalent to 80–320 mg of the standard form taken twice daily. No severe hypoglycaemia occurred. Mild or moderate hypoglycaemia was seen in 5% of those on the MR form. Once-daily administration may help adherence, but the MR form costs more – £62 a year at 60 mg a day, £89 at 90 mg. The standard form costs about £28 a year.

Pioglitazone

Pioglitazone, the only glitazone used in the UK, can cause oedema, which can precipitate congestive heart failure (CHF) and fractures. CHF is a common cause of admission to hospital, and the second commonest first presentation of cardiovascular disease (CVD) (after peripheral arterial disease). 32 A fivefold risk of macular oedema (MO) has also been reported. 33

There is an increased risk of fractures amongst people taking pioglitazone. The fractures were originally reported as being atypical fractures of long bones,34 but Scottish data also show an increase in hip fractures. 35

More recently there has been concern over bladder cancer. Pioglitazone use has now been discontinued in France.

However, the evidence is inconsistent. A Canadian study using UK data36 reported an increased risk of 1.83 (95% confidence interval (CI) 1.10 to 3.05). A French study37 reported a doubling of a very small risk of bladder cancer. The large Kaiser Permanente study from the USA reported an increase in risk with pioglitazone with RR of 1.18 but this was not statistically significant. 38 The PrOactive (PROspective pioglitAzone Clinical Trial In macroVascular Events) trial reported a RR of 2.83 (p = 0.04) but once cases of bladder cancer diagnosed in the first year were excluded there was no difference. 39 It was argued that cancers diagnosed within a year of starting the drug must have been there before. However, Gale has argued that pioglitazone could be acting as a growth promoter in latent tumours. 40

A very large study by Levin et al. ,41 mainly in the UK, Finland and British Columbia (one million people with type 2 diabetes, almost 6 million person-years of observation), found no increased risk of bladder cancer, providing further reassurance.

It should be noted that diabetes itself has been reported in a very large meta-analysis to increase the risk of bladder cancer with RR 1.35 (95% CI 1.17 to 1.56), though this applied only to those within 5 years of diagnosis. 42 Amongst those with duration of over 5 years, the RR was 1.08.

The European Medicines Agency (EMA) issued a statement in 2011 saying that there was a small increased risk of bladder cancer, but that, on balance, pioglitazone could still be used as a second- and third-line treatment. 43 The Medicines & Healthcare products Regulatory Agency (MHRA) concurred. 44

Patients should be screened for haematuria before starting pioglitazone and then at least annually afterwards.

There are some cardiovascular benefits from pioglitazone (the reverse of what was seen with rosiglitazone) with a reported reduced risk of myocardial infarction (MI), but there is clearly an increased risk of heart failure,34,39 and regular monitoring with B-type natriuretic peptide (BNP) seems advisable for the safest use of this drug. 45 Patients are advised of possible side effects and advised to stop if oedema or shortness of breath develops. If there are concerns regarding heart failure, echocardiography is often carried out, to check that left ventricular function is satisfactory, before starting pioglitazone.

Despite its side effects, including progressive weight gain by as much as 5 kg, pioglitazone can be a valuable diabetes therapy, as it is an insulin sensitiser and allows reduction in insulin resistance, still known to be a major factor in the pathogenesis of type 2 diabetes and glucose intolerance. Early studies using genetic profiling showed that the Pro12Ala of the peroxisome proliferator-activated receptor gamma (PPARG) gene showed a population attributable risk of approximately 50% and, taken together with clinical risk factors, might define those most at risk of renal sodium retention and oedema. Unfortunately, probably because of the fact that the PPARG agonists also show greater metabolic efficacy in those with the Pro12Ala variant, this approach has not been developed in clinical practice, as those who would benefit most would have to be excluded. 46

Many people with type 2 diabetes are considerably overweight and may develop non-alcoholic fatty liver disease (NAFLD). Pioglitazone has been reported to improve NAFLD,47 so if attempts at weight loss are unsuccessful and the NAFLD is progressing, pioglitazone may need to be considered for this group of patients. NAFLD is a spectrum of disease ranging from an increased fat content in the liver (steatosis) to inflammation (non-alcoholic steatohepatitis) and possibly on to cirrhosis. NAFLD is strongly associated with insulin resistance.

Despite its AEs, pioglitazone is still widely used, though its use may be declining, with new initiations falling in recent years. The Health and Social Care Information Centre Report gives figures for items prescribed in 2013–143 (Table 1).

The strongest argument for using pioglitazone is the very low cost, but the costs of AEs need to be considered.

Dipeptidyl peptidase-4 inhibitors

The first two of these to reach the market, sitagliptin and vildagliptin, were appraised for NICE CG87, and recommended for use in combination therapy. 8 There are now five DPP-4 inhibitors with slightly different licensed indications. Others are coming, including two that are taken only once a week, trelagliptin and omarigliptin, both now licensed in Japan.

The CG87 guidance8 is reproduced in Box 2.

Repaglinide

Repaglinide acts on the same receptor in the pancreas as the SUs (and another receptor) but is shorter-acting and was therefore thought to be particularly useful in controlling hyperglycaemia after meals. Like the SUs, its AEs include significant weight gain and hypoglycaemia.

The relevant recommendation in CG878 was ‘to consider offering a rapid-acting insulin secretagogue to a person with an erratic lifestyle’. This presumably related to unpredictability of mealtimes, when there would be a case for using a shorter-acting meglitinide analogue instead of a SU.

The cost of repaglinide treatment will depend on dosages used. It was designed to be taken to reduce postprandial hyperglycaemia, which means it should be taken at mealtimes. The NICE guideline costing16 assumes a total daily dose of 4 mg. If that comprised 2 × 2-mg tablets twice a day, the annual cost would be about £48. However, that assumes that people take it at only two meals. If a third 2-mg dose was added, the annual cost would be £72. But if the third dose was only 1 mg (say to cover a small breakfast or lunch), the annual cost would be £92, because the 1-mg tablets are almost double the price of the 2-mg ones. The variability in doses used in the repaglinide studies makes comparison with the SUs difficult.

Sodium–glucose co-transporter 2 inhibitors

The SGLT2 inhibitors, hereafter referred to as the flozins, have a unique mechanism of action. In the non-diabetic state glucose is allowed through the filter in the renal glomeruli but is fully reabsorbed in the renal tubules through sodium–glucose co-transporter mechanisms. Glycosuria (glucose in the urine) occurs when the renal threshold for glucose (blood glucose of approximately 10 mmol/l) is exceeded. The main transport mechanism responsible for glucose reabsorption, SGLT2, is found in the proximal kidney tubule. This is encoded by the gene for the solute carrier family 5 (sodium–glucose co-transporter), member 2 (SLC5A2). Some people have a mutation in the SLC5A2 gene that causes a defective SGLT2 protein, resulting in glycosuria. Individuals who have this mutation do not have significant problems related to the glycosuria, such as urinary tract infections (UTIs), and they have a normal life expectancy with no increase in cardiovascular mortality or urogenital cancers. 48 This implies that blocking the transport mechanism should not cause problems.

The flozins block the SGLT2 system and so mimic the effect of the SLC5A2 mutation and reduce the reabsorption of renal filtered glucose back into the bloodstream, thereby lowering blood glucose levels. Owing to their insulin-independent mode of action, they do this without weight gain or hypoglycaemia. 49

For uncertain reasons, the SGLT2 inhibitors do not block all glucose reabsorption. Around 160–180 mg of glucose is filtered into the urine each day, and the SGLT2 system reabsorbs 80–90% of that. The amount blocked appears to vary amongst the different drugs, with dapagliflozin 10 mg blocking only about a third of reabsorption. 50,51 Even very large doses of dapagliflozin (such as 100 mg) do not block all glucose reabsorption in people with type 2 diabetes. 52 But none of the SGLT2 inhibitors block reabsorption of over half of the filtered glucose load.

There is also a sodium–glucose co-transporter 1 (SGLT1) transport mechanism, which is present both in the kidney and the gut. In the kidney, it is much less important than SGLT2. Inhibition of gut SGLT1 reduces absorption of glucose there, and it has been suggested that canagliflozin may have a dual action. This was reported first in healthy volunteers53 but has since been reported in a study of people with type 2 diabetes. 54

Because these drugs act through an insulin-independent mechanism, they can be effective when other drugs that depend entirely (sulfonylureas and meglitinides) or in part (gliptins and GLP-1 analogues) on stimulating insulin release have lost effectiveness. In type 2 diabetes, the capacity of the pancreatic beta cells to produce insulin often falls over time.

In addition to improving glycaemic control, the SGLT2 inhibitors also reduce blood pressure (BP). In a meta-analysis of 27 randomised controlled trials (RCTs) with 12,960 patients, Baker et al. 55 reported a mean reduction in systolic blood pressure (SBP) of 4 mmHg.

Inclusion criteria

Search strategy

Searches were run in Ovid MEDLINE, EMBASE and Web of Science from the inception of the databases until February 2015. Thereafter, weekly auto-alerts were run in PubMed in process and EMBASE until September 2015 to check for newly emerging studies. The searches were not restricted by language or publication type. The full search strategy is shown in Appendix 1.

Selection of studies

Two reviewers independently checked titles and abstracts of the search results against the inclusion criteria. Studies were retrieved in full if they appeared to fulfil the inclusion criteria or when eligibility could not be determined from the search results alone.

Assessment of study quality

The quality of the RCTs was assessed using the Cochrane risk of bias tool, which included the following items (rated as adequate, unclear, not reported or inadequate):

• method of randomisation

• allocation concealment

• blinding of participants and personnel

• blinding of outcome assessment

• incomplete outcome data (> 20% dropout regarded as inadequate)

• intention-to-treat (ITT) analysis

• selective reporting

• similarity at baseline

• other (e.g. power analysis).

Overall quality was expressed in terms of proportion of items rated as ‘adequate’.

Quality was assessed by one reviewer and checked by a second reviewer.

Data extraction

Data were extracted using a predesigned data extraction table, with one reviewer extracting and another reviewer checking the data.

Results were expressed as means and standard deviations (SDs). Standard errors (SEs) and CIs were converted to SDs using the equations provided in the Cochrane handbook. 75 Results for lipids were expressed as millimoles per litre (mmol/l). Cholesterol values expressed in milligrams per decilitre (mg/dl) were converted to mmol/l by dividing by 38.67 and lipid values expressed in milligrams per decilitre were converted to millimoles per litre by dividing by 88.57.

Data summary

Data were summarised using text and tables.

The following subgroup analyses were considered:

• BMI < 25, 25–29, 30 kg/m² and over

• baseline HbA_1c.

Search results

Seven studies76–84 were included in the final analysis. We will usually refer to them by first author and year. They were:

Canagliflozin:

• CANTATA-M (CANagliflozin Treatment and Trial Analysis – Monotherapy) 201384

• Inagaki 2014. 76

Dapagliflozin:

• Ferrannini 2010 (with Bailey et al. 2015)77,78

• Ji 201479

• Kaku 2014. 80

Empagliflozin:

• Lewin 201581

• Roden 2013–14. 82,83

A list of excluded studies, and reasons for exclusion, is in Appendix 2.

Characteristics of included studies

A summary of study characteristics is shown in Table 2.

Details can be found in Appendix 3.

Glycated haemoglobin

Weight

Lipids

Systolic blood pressure

Urogenital tract infections

Although most UTIs are mild and easily resolved with appropriate antibiotic treatment, more severe infections can be devastating, resulting in bacteraemia, sepsis and death. Because of the frequency with which they occur, UTIs also impose a substantial economic burden on health-care systems. 88

Symptoms of UTI include dysuria (a burning feeling when urinating); frequency of urination; urgency (a feeling of an intense urge to urinate); pain or pressure in the back or lower abdomen; nausea and/or vomiting; cloudy, dark, bloody or strange-smelling urine; feeling tired or shaky; and fever or chills.

The presence of glucose in the urine (glycosuria) creates a suitable environment for the growth and proliferation of bacteria. Glycosuria also promotes increased adherence of bacteria to uroepithelial cells, in particular Escherichia coli. 89 By blocking renal glucose reabsorption, SGLT2 inhibitors cause glycosuria, and increase the risk of UTI in patients. 89

Glycosuria in patients with type 2 diabetes predisposes these patients to develop GTIs, in particular genital mycotic infections, that is vulvovaginal candidiasis in women and candida balanitis in men, as it provides a favourable growth environment for otherwise commensal genital microorganisms. Candida albicans is the most common cause, but Candida glabrata is also an important cause in women with type 2 diabetes. 90

Symptoms of genital candidiasis can include itching; burning; genital discharge; pain during sexual intercourse; soreness; redness in the genital area; and rash.

Both UTIs and GTIs are more common in females. 91

Frequencies of urinary tract infections

The trials of different drugs reported different rates of UTIs, but a recent meta-analysis of 19 trials found no significant differences in risk amongst the three drugs. 120

When do urinary tract infections occur?

Several trials report cumulative incidence of UTIs. Kaku80 reports 2.3% at 24 weeks and 3.6% by 52 weeks. So 1.3% of UTIs occur in months 7–12. Ferrannini77 reports 5.7% at 24 weeks and 8.6% at 102 weeks. So 2.9% occur from week 24 to week 102. Roden82,83 reports 6.7% at week 24 and 9.4% by week 76. So 2.7% occurred from week 24 to week 76.

Patients on SGLT2 inhibitors who have more than one UTI will be switched to another drug. For modelling purposes, we will assume that:

• Sixty per cent of flozin-induced UTIs will occur in the first 6 months.

• All flozin-induced UTIs will occur in the first 2 years.

• Two UTIs will trigger a change of therapy.

How were these cardiovascular benefits achieved?

Glycated haemoglobin was 0.57% lower in the empagliflozin group than the placebo arm at week 12, but steadily narrowed thereafter to 0.35% at week 206. Given the weak relationship between glycaemic control and CVD,68 this difference seems unlikely to have caused the difference.

Systolic blood pressure fell by about 5.5 mmHg in the empagliflozin group and by about 2 mmHg in the placebo group by week 16, but the difference between the empagliflozin and placebo groups narrowed thereafter to about 2 mmHg. There was no difference in BP lowering between doses. Diastolic blood pressure (DBP) fell by about 2.5 mmHg in all three arms. By 206 weeks, there was no difference in DBP between empagliflozin and placebo groups.

For some complications of diabetes, BP control is as important as glycaemic control, as was shown by the UKPDS study, where ‘tight’ BP control (with a mean BP of 144/82 mmHg, it was not really tight) reduced overall mortality by 32% (RR 0.68, 95% CI 0.49 to 0.94). 138

Weight was reduced by 2 kg in the empagliflozin group and by 1.2 kg in the placebo group.

Changes in lipids were small. On empagliflozin 25 mg, LDL-C rose (placebo adjusted) from baseline 2.2 mmol/l to about 2.3 by 12 months, stayed there until about 136 weeks then fell to about 2.21 mmol/l, just below the placebo level. On empagliflozin 25 mg, HDL-C rose by about 0.05 mmol/l then fell slightly. The placebo level rose by about 0.01 mmol/l. (Figures derived from graph – data not provided in text.) The baseline TC/high-density lipoprotein (HDL) ratio was 3.5, perhaps because so many were on statins and other lipid-lowering drugs. These lipid changes seem insufficient to explain the mortality results.

However, the combination of factors might have more effect than the individual ones, and the reduction in BP, though small, is similar to that seen in the Heart Outcomes Prevention Evaluation (HOPE) trial,139 in which a reduction of 2–4 mmHg in BP from ramipril (an ACEI) was thought by the HOPE authors to be sufficient to explain about one-quarter of the observed 25% reduction in cardiovascular events, in another high-risk group with vascular disease and/or diabetes.

Discontinuation rates from study drugs due to adverse events are reported as 19.4% for placebo and 17.3% for empagliflozin in the paper but as 13.0% and 11.5% in appendix H.

Subgroup analysis for the primary composite outcome shows:

• Statistically significant benefit in Asians (a mixed group, with about 44% from North East Asia) but not in white people, though death from cardiovascular causes is significantly reduced in white people. This implies that white people gained less for non-fatal MI and stroke. The Asian group is rather heterogeneous and no details are given of risks in East Asians versus South Asians. There are differences in the balance of insulin deficiency and insulin resistance.

• Statistically significant benefit in those with baseline HbA_1c under 8.5% but not in those above that.

• Statistically significant benefit in those with BMI of < 30 kg/m² but not in those above that level.

• Statistically significant benefit in those on insulin but not in those on non-insulin regimens; 51% in both groups were on insulin.

• Greater benefit in those aged over 65 years.

In a number of subgroups, reductions in cardiovascular death were statistically significant when reductions in the primary outcome were not. To recall, the primary outcome was cardiovascular death, non-fatal MI and non-fatal stroke. Of 282 primary outcome events in the placebo group, 49% were cardiovascular deaths. Of 490 primary outcome events in the empagliflozin groups, 35% were cardiovascular deaths. So a greater proportion of events in the empagliflozin group was of non-fatal events. The fact that in some subgroups, cardiovascular death rates are significantly different when the composite primary outcome is not, is explained by a lack of any statistically significant differences in non-fatal MI and non-fatal stroke.

Non-fatal MI was diagnosed on the basis of symptoms plus one or more of:

• troponin or creatine kinase-MB

• ECG changes

• imaging of new non-viable or non-motile myocardium.

This study has attracted worldwide interest. It contrasts with the equivalent studies with the DPP-4 inhibitors, which did not show any reduction in cardiovascular outcomes. They were:

• SAVOUR for saxagliptin (The Saxagliptin assessment Of Vascular Outcomes Recorded in patients with diabetes mellitus). 140

• EXAMINE for alogliptin (Examination of cardiovascular outcomes with alogliptin versus standard of care in patients with type 2 diabetes mellitus and acute coronary syndrome). 141

• TECOS for sitagliptin (Trial Evaluating Cardiovascular Outcomes with Sitagliptin). 142

There was no difference in cardiovascular outcomes in SAVOUR except that more patients on saxagliptin than on placebo (3.5% vs. 2.8%: HR 1.27, 95% CI 1.07 to 1.51) were admitted to hospital with heart failure.

The findings in EXAMINE were similar – no difference in end points – except an increase in heart failure in a subgroup analysis of patients with no heart failure at baseline (2.2% on alogliptin, 1.3% on placebo; HR 1.76, 95% CI 1.07 to 2.90).

The TECOS results were again similar – no differences in a composite primary end point of cardiovascular death, non-fatal MI and non-fatal stroke nor in other end points including hospital admission for heart failure and death from any cause.

These results were seen as providing reassurance, in the wake of the rosiglitazone story with an increase in cardiovascular events. 143 They could also be seen as disappointing in that they did not reduce the most important complication of diabetes, the excess of CVD. However, as has been pointed out by Hirshberg and Katz,144 these trials ran for only a few years and showed only small changes in HbA_1c (reductions of 0.27–0.36% compared with placebo), so reducing the chance of showing reductions in cardiovascular events.

The subgroup analyses in EMPA-REG OUTCOME (Empagliflozin Cardiovascular Outcome Event Trial in Type 2 diabetes Mellitus Patients) trial are interesting. 136 Younger, lighter, better-controlled patients did better, as did the Asian group. There could be overlapping features here in that the East Asians tend to be lighter. Further details will no doubt be released but with such a very large study, further analysis is bound to take time.

The differences observed do not seem sufficient to justify the very optimistic media coverage, such as reports that ‘Lilly’s Jardiance diabetes pill could be a $6 billion-a-year blockbuster’. 145

It is worth noting that the EMPA-REG OUTCOME trial136 involved patients at high cardiovascular risk who had had diabetes for many years and who were on complex regimens for their diabetes. The results are not applicable to people starting monotherapy with empagliflozin.

Selection of trials

We applied the following selection criteria:

• Trials of 24–26 weeks, starting with placebo as the common comparator.

• Trials only of selected drugs. For example, we did not include all sulfonylureas but focused on gliclazide, which should be the sulfonylurea for comparison in trials of newer agents. 26 We did not include all DPP-4 inhibitors, originally intending to focus on sitagliptin.

• Baseline HbA_1c of 7.5% or more, based on the NICE guideline for treatment intensification. 8 There are some RCTs in patients with lower baseline HbA_1c values, but they have less scope for lowering HbA_1c.

• Dropout rates of no more than 20%.

However, the first of these criteria had to be relaxed in order to include gliclazide, as we found no trials of gliclazide against placebo. We had to indirectly link gliclazide with placebo via linagliptin and pioglitazone. All of the other drugs included had trials against placebo. Unfortunately, we found no satisfactory trials of repaglinide for inclusion.

We searched the lists of trials used for the NICE guideline group on type 2 diabetes, but carried out additional searches specifically for gliclazide trials, as the guideline group pooled trials of sulfonylureas.

Evidence on repaglinide

The annex to the NICE guideline CG87 lists seven studies on repaglinide but only three gave 24-week data. 16

Abbatecola et al. 146 report a randomised trial comparing repaglinide and glibenclamide. The main outcome measure was cognitive function, with the hypothesis being the tighter control of postprandial PG would reduce cognitive decline, in patients aged 60–78 years, mean age 74 years. [Note that according to the British National Formulary (BNF), repaglinide is ‘not recommended’ in people over 75 years. 147]

The baseline HbA_1c in patients in this trial was quite low – 7.25%. So it is an exclusion for our purposes. The final HbA_1c is not given in the text, but from the graph is about 6.6% with no difference between the drugs.

In the Jovanovic 2000 trial,148 patients were randomised to placebo, and repaglinide 1 mg or 4 mg daily. Under 30% of patients were drug naive, and 10% had been on two glucose-lowering drugs. Baseline HbA_1c was 8.6% in the placebo group and rose to 10% at 24 weeks. In those who had been on previous combination treatment, HbA_1c rose by 1.8% on placebo, and fell by inconsequential 0.07% and 0.05% on repaglinide 1 mg and 4 mg. Dropout rates were very high – 60% in the placebo group, of which half had to start rescue treatment, and 23% and 31% in the repaglinide groups. Given that 70% had been on prior drug therapy, the high proportion in the placebo group requiring rescue treatment is not surprising, but it devalues any conclusions drawn from this study.

The third 24-week study used by NICE was by Saleem et al. 149 from Lahore. This compared the effects on HbA_1c of repaglinide and glibenclamide. It says that 50 patients were ‘randomly selected’ for each group but gives no details of how this was done, or on allocation concealment. Blinding was not feasible because of different dosing frequencies – once or twice daily for glibenclamide, preprandially up to three times a day for repaglinide. No patients are reported to have dropped out. The recruitment period in this study (March 2006 to March 2007) overlaps with that for another paper by the same group (Shah et al. 2011150), which reports only PG, in 200 patients. The changes in fasting plasma glucose (FPG) and 2-hour PG are almost identical in the two studies. 149,150 The Shah paper150 has no HbA_1c data. Saleem et al. 2011149 report a reduction by 24 weeks in HbA_1c of 0.6% on repaglinide and 0.4% on glibenclamide. The final repaglinide dose was 4.27 mg daily, and the final glibenclamide dose was 8.8 mg (identical to the Shah et al. article150). The Shah article150 states that dosages were reported as being adjusted based on glucose levels, so it is not clear why the final glucose levels are so different, with a reduction in FPG in the repaglinide group which is almost double that in the glibenclamide group. No details of source of funding are given. We think that the patients in the Saleem study149 may be a subset of those in the Shah study. 150

We also note the trial by Jibran et al. 151 This paper is very similar to the Saleem et al. 2011149 paper, but has no authors in common. The numbers of patients are the same, and values for baseline age, weight and BMI have identical means and SDs. The result tables are identical in means and SDs. Much of the text is the same. The patients are said to have been recruited in different time periods.

Glibenclamide is a first-generation SU and was not included in our NMA, so the Saleem149/Jibran151 and Abbatecola146 trials are not included.

The NICE guideline group also considered evidence on repaglinide at 12 months, from four trials. These included the Abbatecola and Saleem trials146,149 mentioned above, and two better-quality ones by Derosa et al. 152 and Marbury et al. 153 The Derosa trial152 compared repaglinide with glimepiride, in patients with mean HbA_1c of 8.0%, and showed a reduction of 1.2% at 12 months. We use the effect sizes from this study in our modelling, for changes in HbA_1c, SBP and weight. However, we prefer gliclazide to glimepiride, so the Derosa trial152 is not included in our NMA. The Marbury trial153 recruited both patients who had never had any glucose-lowering drugs (13%) and those who had previously been treated with sulfonylureas and other drugs (87%). The reduction in HbA_1c was much greater in the pharmacotherapy-naive group – 1.3% at 12 months, similar to the Derosa152 results. In previously treated patients, the HbA_1c actually rose by 0.3%. Mean baseline HbA_1c was quite high at 8.7% so the results may be less applicable to patients treated according to the NICE guidelines with close monitoring and prompt intensification when HbA_1c exceeded 7.5%. The sulfonylurea comparator was glibenclamide, so the trial is not used in our NMA.

The network diagram is shown in Figure 2. The included trials are listed in Table 8.

FIGURE 2.

Network meta-analysis diagram. b.i.d., twice daily.

Summary measures

The primary measures of treatment effect were the mean differences (MDs) in change from baseline for HbA_1c, weight gain and SBP. A negative value indicates improvement in the outcome. In the case of missing values for SD of change from baseline values, the SD was imputed as described in detail in the Cochrane Handbook. 75 In brief, we assumed a correlation of r = 0.5 between baseline and follow-up to estimate SD for change from baseline.

Data synthesis and model implementation

We used a Bayesian NMA method to analyse all the data, preserving randomised treatment effects within trials and accounting for correlation between comparisons with three arms or four arms using the freely available software, WinBUGS 1.4.3 (MRC Biostatistics Unit, Cambridge, UK). The statistical heterogeneity in treatment effect estimates was estimated using between-study variance (i.e. square root of the SD of underlying effects across trials) with 95% CrI. 164 To estimate inconsistency in the networks of evidence, we calculated the difference between indirect and direct estimates whenever indirect estimates could be constructed with a single common comparator. 164 Inconsistency was defined as disagreement between direct and indirect evidence with a 95% CrI excluding 0 for MD. 165 The model convergence was assessed using trace plots and the Brooks–Gelman–Rubin statistic. 166 The analysis was undertaken using two Markov chains, which were run simultaneously. The model was found to be converging adequately after 20,000 samples for both chains. We ran the model further using 70,000 samples and the results presented in the paper are based on these samples as we discarded the first 20,000 samples.

We used both the fixed- and random-effects models. The Bayesian deviance information criterion (DIC) was used to compare the two models to see which was appropriate to compare treatment effects. The DIC measures the fit of the model while penalising it for the number of effective parameters. The model with the lowest DIC value was considered as the most appropriate NMA model. Based on DIC values obtained from the two models and also because of small number of studies available for the NMA, a fixed-effects model was chosen. Owing to small number of studies, it would have been difficult to estimate between studies variance if a random-effects model was implemented.

All results are reported as posterior medians of MDs with corresponding 95% CrIs. CrIs are the Bayesian equivalent of classic CIs. A 95% CrI can be interpreted as there being a 95% probability that the parameter takes a value in the specified range. Drugs were not ranked, but were considered in terms of effect sizes and uncertainties.

Glycated haemoglobin (haemoglobin A_1c)

Networks of eligible comparisons for HbA_1c are provided in Figure 3, showing predominantly pairwise comparisons of drugs with placebo. There were eleven comparisons (ten drugs plus placebo).

FIGURE 3.

Network plot: HbA_1c.

Figure 4 and Table 9 display a caterpillar plot of the MD and 95% CrIs for all comparisons for mean change in HbA_1c (at 24 weeks) from baseline. All SGLT2 inhibitors were significantly more effective than placebo in reducing mean change in HbA_1c from baseline, with the reduction ranging from –1.19% to –0.59%. Canagliflozin 100 mg and 300 mg and pioglitazone were significantly more effective in reducing mean change in HbA_1c from baseline than linagliptin 5 mg, dapagliflozin 10 mg and vildagliptin 50 mg. The reductions in HbA_1c from baseline were similar for linagliptin 5 mg and dapagliflozin 10 mg. The between-study variance was small, suggesting no heterogeneity, but the CrIs were wide, which reflects the small number of studies available for pairwise comparisons. Analyses based on direct versus indirect comparisons showed no evidence of inconsistency between direct and indirect evidence in the network for HbA_1c.

FIGURE 4.

Pairwise comparisons of all drugs for HbA_1c.

Weight gain

Networks of eligible comparisons for the weight gain are provided in Figure 5, showing predominantly pairwise comparisons of drugs with placebo. There were 11 comparisons (10 active drugs plus placebo).

FIGURE 5.

Network plot: weight gain.

Figure 6 and Table 10 display a caterpillar plot of the MD and 95% CrIs for all comparisons for mean change in weight gain from baseline. Sitagliptin 100 mg, vildagliptin 50 mg, gliclazide and pioglitazone were associated with significant weight gain compared with placebo, with the weight gain ranging from 0.74 kg to as much as 3.79 kg. Compared with placebo, canagliflozin 300 mg, canagliflozin 100 mg, empagliflozin 25 mg, empagliflozin 10 mg and dapagliflozin 10 mg were associated with significant weight loss, ranging from –2.91 kg to –1.58 kg. Compared with all other drugs in the network, canagliflozin 300 mg was associated with significant weight reduction. The between-study variance was small, suggesting no heterogeneity, but the CrIs were wide, which reflects the small number of studies available for pairwise comparisons. Analyses based on direct versus indirect comparisons showed no evidence of inconsistency between direct and indirect evidence in the network for weight gain.

FIGURE 6.

Pairwise comparisons for weight gain.

Systolic blood pressure

Networks of eligible comparisons for the SBP are shown in Figure 7, showing predominantly pairwise comparisons of drugs with placebo. There were seven comparisons.

FIGURE 7.

Network plot: SBP.

Figure 8 and Table 11 display a caterpillar plot of the MD and 95% CrIs for all comparisons for mean change in SBP from baseline. Canagliflozin 100 mg and 300 mg, empagliflozin 25 mg, dapagliflozin 10 mg and empagliflozin 10 mg were significantly more effective in reducing mean change in SBP from baseline compared with placebo and sitagliptin 100 mg. Canagliflozin 300 mg gave the largest reduction in mean change in SBP from baseline compared with placebo (–5.65 mmHg). The between-study variance was small, suggesting no heterogeneity, but the CrIs were wide, which reflects the small number of studies available for pairwise comparisons. Analyses based on direct versus indirect comparisons showed no evidence of inconsistency between direct and indirect evidence in the network for SBP.

FIGURE 8.

Pairwise comparisons for SBP.

One question was whether or not canagliflozin is more potent than other SGLT2 inhibitors, due to its dual effect on SGLT2 and SGLT1 receptors. In monotherapy, both doses of canagliflozin lowered HbA_1c slightly more than both doses of empagliflozin, which does not have a significant effect on SGLT1 receptors. Nor does canagliflozin 100 mg. This suggests that the SGLT1 effect does not explain all the differences in HbA_1c results. It may explain some of the difference between the two doses of canagliflozin, or it may not be clinically significant.

However, irrespective of the mechanism, one finding is that canagliflozin 300 mg does have a greater effect on HbA_1c than dapagliflozin and empagliflozin. Indeed, the 100-mg dose also has a clinically significantly greater reduction in HbA_1c than dapagliflozin 10 mg (but see caveats to follow).

Table 12 compares the effects and AEs of the two doses of canagliflozin, with effects taken from our NMA and AEs from the published studies. Effects are compared with placebo. However, we do not know whether the reduction seen with canagliflozin 300 mg in the trials would be as great in patients who responded insufficiently to the 100-mg dose.

If in patients in whom a SGLT2 inhibitor is considered the appropriate choice, it is considered worth trying canagliflozin 300 mg if the 100-mg dose does not have enough effect; it would be logical to also try canagliflozin 300 mg if dapagliflozin or empagliflozin are insufficiently effective. The licence implies that they would have to switch to canagliflozin 100 mg first, if only briefly. However, the same caveat would apply – the HbA_1c reduction seen with canagliflozin 300 mg might be less amongst patients who have not responded sufficiently to starting doses.

Table 13 shows that the differences in effects of the two empagliflozin doses are slight, using figures from the Roden and Lewin trials. 81–83

Again, a caveat is required. Those who do not respond to empagliflozin 10 mg may not achieve as great a reduction in HbA_1c after increasing to 25 mg daily, as in Table 13. The differences are, in any case, mostly not clinically meaningful.

In the NMA reported here, dapagliflozin reduced HbA_1c significantly less than canagliflozin 100 mg, but it should be noted that the Kaku 2014 trial80 of dapagliflozin recruited patients with mean baseline HbA_1c of 7.5%, whereas most trials had baseline HbA_1c of around 8%. To summarise, the placebo-adjusted HbA_1c reductions in the trials at 24–26 weeks were, for the starting dosages:

• canagliflozin 100 mg:

  • CANTATA84 0.91%

  • Inagaki76 1.03%

• dapagliflozin 10 mg:

  • Ferrannini77,78 0.66%

  • Ji79 0.82%

  • Kaku80 0.39%

• empagliflozin 10 mg:

  • Roden82,83 0.74%.

Hence the Kaku trial,80 while qualifying for our NMA based on the baseline HbA_1c of 7.5%, will be reducing the mean effect of dapagliflozin.

When interpreting weight changes, the baseline BMIs need to be considered. The trials in China and Japan recruited people with BMIs in the 25- to 26-kg/m² range, whereas the European trials had mean BMIs ranging from 28 to almost 34 kg/m². The pattern of type 2 diabetes differs in East Asians, with lower BMI and a more insulin-secretory defect. 167 This does not apply to South Asians (Indian subcontinent) in whom insulin resistance is more important.

Another factor to be considered in interpretation is that in the dapagliflozin trials, HbA_1c fell in the placebo groups, by 0.29%, and by 0.23% in the Ji79 and Ferrannini77 trials. In the Ferrannini trial,77 weight fell significantly, by 2.2 kg. In the placebo groups in the canagliflozin trials, HbA_1c rose by 0.29% (Inagaki76) and 0.14% (Stenlöf, CANTATA-M84). Ferranini et al. 77 (and the AstraZeneca submission, which talks of a ‘motivated placebo group’ on p. 58) suggested that the reduction in HbA_1c in the placebo group might have been due to improved adherence to lifestyle advice in that group, but as the placebo tablets matched the dapagliflozin ones, this seems unlikely.

Dual therapy

The NICE guideline on type 2 diabetes envisages, for patients who cannot take metformin, dual therapy with one of the following combinations:

• pioglitazone and a sulfonylurea

• pioglitazone and a gliptin

• sulfonylurea and a gliptin.

In the interest of simplicity, we have chosen the sulfonylurea as the second drug, except after gliclazide monotherapy, when we use pioglitazone. The sulfonylurea was preferred to pioglitazone because of the latter’s safety record. Pioglitazone is preferred to a DPP-4 inhibitor only on cost grounds.

We have assumed that patients are at the maximum tolerated dose of each monotherapy drug before moving to dual therapy.

Triple therapy

Moving to triple therapy is more complicated, as after some of the dual regimens, pioglitazone and a gliptin are still available, and the GLP-1 analogues and insulin enter the frame. It is not possible to review all options.

At this stage, the NICE guideline recommends that insulin-based treatment should be considered. 16

In the interests of simplicity, our base case is therefore to bring in NPH insulin for triple therapy. We therefore have sequences as follow:

• empagliflozin 25 mg → empagliflozin + gliclazide → empagliflozin + gliclazide + NPH insulin

• canagliflozin 300 mg → canagliflozin + gliclazide → canagliflozin + gliclazide + NPH

• dapagliflozin 10 mg → dapagliflozin + gliclazide → dapagliflozin + gliclazide + NPH

• sitagliptin 100 mg → sitagliptin + gliclazide → sitagliptin + gliclazide + NPH

• pioglitazone 45 mg → pioglitazone + gliclazide → pioglitazone + gliclazide + NPH

• gliclazide → gliclazide + pioglitazone → gliclazide + pioglitazone + NPH.

Some patients will progress to needing short-acting insulin to control blood glucose after meals. We assume that once patients move to a basal-bolus insulin regimen, the sulfonylurea will be stopped.

Note that we have not introduced any of the flozins beyond monotherapy, as those situations were dealt with in the three single technology appraisals (STAs).

An alternative to bringing in insulin as third drug, is to consider the GLP-1 analogues. These are simpler for patients to manage, involving a once-a-week injection, and a low risk of hypoglycaemia.

The National Institute for Health and Care Excellence has adopted a very restrictive position on the GLP-1 analogues, based on a minimum BMI, and stopping rules requiring both a 1% reduction in HbA_1c and 3% weight loss, but that was not based on any cost-effectiveness analysis and is due for review.

Only one long-acting GLP-1 analogue has been appraised by NICE – long-acting exenatide. If we bring that in as a third drug, basal insulin would be the fourth drug, with sequences as follow:

• empagliflozin 25 mg → empagliflozin + gliclazide → empagliflozin + gliclazide + exenatide LA → empagliflozin + gliclazide + exenatide + NPH insulin

• canagliflozin 300 mg → canagliflozin + gliclazide → canagliflozin + gliclazide + exenatide LA → canagliflozin + gliclazide + exenatide + NPH

• dapagliflozin → dapagliflozin + gliclazide → dapagliflozin + gliclazide + exenatide LA → dapagliflozin + gliclazide + exenatide + NPH

• sitagliptin 100 mg → sitagliptin + gliclazide → sitagliptin + gliclazide + exenatide LA → sitagliptin + gliclazide + exenatide + NPH

• pioglitazone 45 mg → pioglitazone + gliclazide → pioglitazone + gliclazide + exenatide LA → pioglitazone + gliclazide + exenatide + NPH.

One problem with deriving effect sizes for modelling is that most trials recruit patients with rather poorer control than would be expected amongst patients who are being followed up according to NICE type 2 guidelines that recommend measuring HbA_1c at 3–6 monthly intervals until HbA_1c is stable (and presumably satisfactory).

So if the above guideline is being followed, patients whose HbA_1c rises above the 7.5% intensification threshold, should have that detected within a few months, before it has gone much higher. Their HbA_1c levels might be in the 7.5–8.0% range. Whereas most trials of intensification to dual or triple therapy recruit patients with much higher HbA_1c, often in the 8.7–9.05% range but sometimes well over 9%.

The importance of this is that reductions in HbA_1c tend to be larger when baseline HbA_1c is higher. So the effect sizes in HbA_1c seen in most trials will be larger than expected in management of type 2 diabetes according to the NICE guideline, with close monitoring and prompt intensification.

So we need to be selective in the trials from which we extract data, rather than using effect sizes from broad-spectrum meta-analysis.

The generalisability of trials to routine care has been examined by Thomsen et al. 7 in Denmark. They looked at the effects of adding a second drug to metformin in a large population-based cohort, and concluded that the results were similar to those seen in the trials. The mean HbA_1c at intensification was 8.0%. They observed reductions in median HbA_1c of 1.2% with sulfonylureas, 0.8% with DPP-4 inhibitors, 1.3% with GLP-1 receptor agonists and 2.4% with insulin. However, these differences reflect different baseline HbA_1c levels, notably 9.5–10% amongst those who started insulin. Despite intensification, 41% had not achieved HbA_1c < 7.5% 6 months later.

Thomsen et al. 7 also noted that the threshold for intensification had fallen over the years, from about 8.8% in 2000–3 to about 8.1% in 2010–12 (estimated from graph in supplementary material; figure 3). If this has also occurred in the UK, it reinforces the need to be selective in extracting effect sizes for modelling. In past studies, patients with type 2 diabetes were often left poorly controlled for several years before intensification,189,190 but this may be happening less nowadays, with improved control promoted by the Quality Outcomes Framework of payments to general practices for demonstrating performance against HbA_1c control indicators,191 including DM007 for the HbA_1c indicator. The three bands are now 59, 64 and 75 mmol/mol. All of them (not just the tightest) probably encourage initiation of insulin in practice.

Company submissions

There are three company submissions:

• Boehringer Ingelheim for empagliflozin

• AstraZeneca for dapagliflozin

• Janssen for canagliflozin.

An overarching summary of the companies’ and the AG’s modelling assumptions, inputs and results is presented at the end of the economics section, permitting an easy read across. Readers may wish to work through this overarching summary first, before turning to the more detailed summaries presented below for more clarity around specific points of the individual modelling exercises.

All of the submissions contain modelling exercises with long-term time horizons of around 40 years, which for the majority of patients will be a lifetime horizon. They all undertake a cost–utility analysis using the appropriate perspectives of the NHS and Personal Social Services (PSS) for costs and the patient for benefits, and discount costs and benefits at 3.5%.

Boehringer Ingelheim designed a front end to the UKPDS OM1 model. The Boehringer Ingelheim submission has a great deal in common with the modelling of recent NICE CGs for type 2 diabetes, and the AG modelling for the current assessment, both of which design a front end to the UKPDS OM1.

AstraZeneca uses the Cardiff Diabetes Model (CDM), which uses many of the UKPDS68193 equations and so has much in common with the UKPDS OM1 model, but updates the calculation of the probabilities of having an event to use the UKPDS82,194 which is the basis of the OM2.

Janssen differs from AstraZeneca and Boehringer Ingelheim in using the ECHO-T2DM model. Its base case has assumptions that differ quite noticeably from those of the other two submissions. There is also relatively little detail in the Janssen submission, with most of the detail being contained in the appendices to the submission and the submitted electronic copy of the model.

In light of the above, the review of the company submissions below provides a reasonably in depth review of the Janssen modelling. This is followed by shorter reviews of the Boehringer Ingelheim and the AstraZeneca modellings, which are more in line with the AG modelling.

Patient characteristics

Patient characteristics at baseline were drawn from a pooled analysis of the CANTATA-M84,92 and Japanese canagliflozin76 studies, resulting in baseline estimates of 56 years of age, 52% male, 8.016% HbA_1c, 128 mmHg SBP, 29.7 kg/m² BMI, 200 mg/dl TC, 118 mg/dl LDL, 48 mg/dl HDL, 175 mg/dl triglycerides and a mean eGFR of 86 ml/minute/1.73 m². Note that the inclusion of the Japanese study76 pulls down the mean BMI because the mean BMI was 25.6 kg/m². Those in CANTATA-M84,92 had a higher BMI, mean 31.6 kg/m².

Based upon the submitted electronic model input sheets, the mean disease duration was 4.6 years, and was assumed to range uniformly between 0 years and 9.2 years. The proportion of patients with background diabetic retinopathy (BDR) was 0.7%; microalbuminuria, 0.1%; and symptomatic neuropathy, 1.5%. The proportion of patients with IHD was 1.2%, MI 0.8%, stroke 0.1% and amputation 0.1%. The other complication rates were zero.

Sequences modelled and treatment effectiveness

The Janssen submission modelled the following treatment sequences (Table 16).

Repaglinide was included only as a scenario analysis. For canagliflozin 100 mg it appears that two arms were modelled: one that intensified by adding gliclazide. The other was classed as confidential. But the Janssen submission is slightly ambiguous about this.

Clinical effectiveness estimates for the monotherapies were mainly drawn from the 26-week NMA, though infection rates were drawn from the canagliflozin trials, with the flozins being assumed to have the same rate as canagliflozin 100 mg and the other comparators the same rate as the placebo arm. Based upon the electronic input sheets submitted by Janssen and the model having an annual cycle, the 26-week estimates were assumed to apply at the end of the first cycle. Note also that the estimates of the electronic input sheets are stated as being relative to placebo, and that there does not appear to be a placebo effect within the electronic input sheets. As a consequence, it appears that the treatment effects relative to placebo rather than the absolute treatment effects have been applied within the Janssen modelling.

The appendices to the submission present two NMAs: with and without repaglinide. The central estimates of these appear to the AG to be virtually identical, with the exception of the rates of severe and non-severe hypoglycaemic events for pioglitazone and sitagliptin. Why these should differ between the two analyses is not clear (Table 17).

As already noted, the definition of hypoglycaemia used a cut-off of below 4.0 mmol/l, which is above the foot of the normal range of 3.5 mmol/l.

The submission appears to state that when a patient intensified by adding another therapy that the clinical effectiveness estimates of that therapy were applied. When a patient intensified therapy by switching to another therapy, rebound was assumed with the clinical effectiveness estimates of the initial therapy being removed prior to applying the clinical effectiveness estimates of the therapy to which the patient was switching. Rebound consequently appears to be to take the patients back to their baseline values for the risk factors. And the clinical effectiveness of a treatment was assumed to be the same whether it was being used as a monotherapy or was being added to other therapies.

At the progression to insulin the clinical effectiveness estimates in Table 18 were applied, the values being taken from the copy of the electronic model input sheet that was submitted. The source of these estimates was not clear to the AG.

Treatment intensification and discontinuation

Treatment intensification occurred if the patient breached the 7.5% HbA_1c treatment intensification threshold. Some details in the Janssen submission were classed as confidential and cannot be replicated here.

A key issue is whether or not the subset of patients who do poorly when treated with canagliflozin 100 mg will, when they switch to canagliflozin 300 mg, experience the same peak HbA_1c effect as estimated from the entire patient group in the trial, which was randomised to canagliflozin 300 mg from the start.

Patients could also discontinue their current treatment because of adverse events and contraindications. It is unclear to the AG whether those discontinuing one treatment were assumed to intensify to monotherapy with another agent or to dual therapy with another agent.

The submission states that the treatment effects associated with discontinuations were immediately reversed at discontinuation. If treatment effects were limited to, for example, one-off reductions in HbA_1c it is easy to see how this treatment effect could be reversed and rebound could occur. But treatment effects are not limited to one-off effects. The evolution of HbA_1c after the initial one-off effect is also treatment specific so has to be counted as a treatment effect. The AG assumption is that rebound was to baseline values, but this is not unambiguous and may have been treated differently for different risk factors, for example HbA_1c and weight.

The Janssen submission is ambiguous about whether or not the model handles treatment intensifications and treatment discontinuations in the same manner. It is important to know whether or not rebound also occurs when treatments are withdrawn at treatment intensification, particularly for the intensification from repaglinide to pioglitazone, and for the intensification to insulin when patients are assumed to discontinue their oral therapies. Appendix 4 to the main submission states that:

Treatment intensification algorithms triggered when biomarker threshold levels are exceeded determine AHA and concomitant medication use (i.e. anti-hypertensive, dyslipidaemia) over time. Each drug (or drug combination) is described by a profile that includes price, initial treatment effects, bio-marker evolution, ‘rebound’ effects applied upon discontinuation, AE rates, non-compliance rates, contraindications, and disutility (if any). A treatment sequence of rescue medications can be specified by the user, including at least one line of rescue insulin therapy. Agents can be continued or discontinued at bio-marker failure (HbA_1c for AHA, SBP for anti-hypertensive agents, and any of the cholesterol components for anti-dyslipidaemia agents).

To the AG, this suggests the intensifications, or discontinuation at biomarker failure of HbA_1c in the above, may be treated in the same manner as discontinuations and treatment switches, and have rebound applied if specified by the user.

This raises the possibility of repaglinide rebounding at treatment failure, so adding 1.28% to the then current 7.5% patient HbA_1c. Intensifications to pioglitazone with its –0.78% effect and then the subsequent intensification to add gliclazide and its effect of –0.59% may not reverse this rebound, given the incorporation of annual drift. If this applied, patients on repaglinide could spend little to no time on subsequent oral intensifications before intensifying to insulin. But the AG assumption is that rebound is to the baseline value rather than to the baseline value plus annual drift, and that this rebound applies to all of the risk factors and not just HbA_1c.

It is not obviously reasonable to assume that there will be rebound when patients start insulin. The clinical effectiveness estimates for insulin may reflect the overall effect of a switch to insulin including any discontinuations of existing therapy.

Glycated haemoglobin evolution

A major difference in assumptions in the Janssen submission compared with the other submissions and the AG modelling is that rather than apply the UKPDS68193 equation to evolve HbA_1c, treatment-specific linear evolutions were assumed. The argument for this is that though most NICE assessments in diabetes have used the UKPDS68193 equations to evolve HbA_1c, these evolutions encompass treatment intensifications. As a consequence, if treatment effects are being associated with treatment intensifications in the modelling, using the UKPDS68193 evolution will tend to double count these treatment effects.

The values of these for the monotherapies were based upon values taken from the ADOPT trial as reported in Kahn et al. 174 The majority of the monotherapies under consideration were not used in the ADOPT trial, which used rosiglitazone, metformin and glyburide (the North American name for glibenclamide). Janssen assumes that values from treatments within the ADOPT trial apply to the monotherapies under consideration in Table 19.

Applying the glibenclamide progression rate to gliclazide may be pessimistic given the 6-year difference in start of insulin on gliclazide and glibenclamide, in favour of gliclazide. 17

Given the annual rates of drift and the initial HbA_1c treatment effects estimated in the NMA, it is apparent that the annual rates of drift are likely to be as, if not more, important than the HbA_1c treatment effects estimated in the NMA. Owing to the NMA estimating HbA_1c from 24-week data, half the annual drift is added to the estimated treatment effect to provide the 52-week estimate.

At intensification, if another treatment is added the annual rate of HbA_1c drift is assumed to be the average of the HbA_1c annual drifts of the two treatments being used as dual therapy (Table 20).

For those intensifying to insulin, Janssen derives an annual rate of drift of 0.15% from the UKPDS82. 194

The impact of these different annual drifts in HbA_1c is applied in tandem with the initial treatment effects. For instance, at the central Janssen treatment effect estimates and a baseline value for HbA_1c of 8.0% it appears that for those who do not discontinue for other reasons the following evolutions are implied up to the point at which HbA_1c rises above 7.5% and treatment is intensified (Figure 9).

FIGURE 9.

Janssen modelled HbA_1c drift by treatment at central values.

Immediately apparent is that for gliclazide (the topmost line) the treatment effect of –0.59% when coupled with half of the annual drift, that is an increase of 0.12%, means that the patient is above 7.5% at the end of the first cycle.

Turning to canagliflozin 100 mg (the solid line), despite its initial treatment effect of –0.97% being somewhat less than the –1.28% of repaglinide, they both breach the 7.5% HbA_1c intensification threshold at the fourth year due to the differences in annual drift. The slower drift for canagliflozin 300 mg also means that despite an initial treatment effect of –1.20% it breaches the 7.5% HbA_1c intensification threshold by 6 years.

Thereafter, it should be borne in mind that repaglinide is assumed to be replaced by pioglitazone. Owing to the withdrawal of treatment the AG reading of the Janssen submission is that in the model this causes the HbA_1c to rebound to the original baseline value of 8.016%, and not to have the full rebound of the repaglinide 1.28% treatment effect applied. The –0.59% pioglitazone effect is then applied with the pioglitazone-specific HbA_1c rate of drift applied, with a further intensification to pioglitazone plus SU after this. In other words, in the repaglinide arm there is the initial repaglinide evolution of HbA_1c in the above example for 4 years, which is then followed by exactly the same HbA_1c evolution as in the pioglitazone arm, only with this being lagged by 4 years.

The same annual drifts apply for empagliflozin, dapagliflozin and sitagliptin as for canagliflozin. They also intensify by adding gliclazide, and then on to insulin. Again, ignoring discontinuations for other reasons, the linear evolution of HbA_1c means that any difference in the timings of the first intensification is also reflected in the timings of the intensification to insulin. It appears that the assumption of a linear evolution of HbA_1c will maintain absolute differences in HbA_1c between treatments, at least until the patient intensifies to insulin. Upon intensification to insulin it appears that the ECHO-T2DM model applies the linear evolution for HbA_1c, but then permits the patient to increase his/her insulin dose in order to stabilise his/her HbA_1c. The Janssen submission is not particularly clear on this point, but it appears that this means that HbA_1c may eventually converge between treatments once the patient has started insulin.

Pioglitazone benefits from a slower annual rate of drift, and continues to derive some benefit from this source even after the intensification of adding gliclazide. At central parameter values and again ignoring discontinuations it will have a permanent HbA_1c benefit over all the other comparators with the possible exception of canagliflozin 300 mg.

Reconciling the above with the evolutions of HbA_1c reported in the Janssen submission is difficult, given the annual cycle of the model. But it should be borne in mind that the Janssen curves are averaged over a large number of patients and PSA iterations, and include the effects of discontinuations for other reasons (Figure 10).

FIGURE 10.

Janssen figure 13: submission reported evolutions of HbA_1c.

For instance, the mean first year effect for the gliclazide arm as shown by the topmost curve is around a 0.95% reduction in HbA_1c, which is somewhat greater than the –0.59% mean estimate for gliclazide. Those discontinuing for reasons other than HbA_1c apparently in effect switch to pioglitazone, the estimate for which is –0.78%. Similarly, the reduction of around 1.4% in HbA_1c for canagliflozin 300 mg, as shown by the bottom curve, is also difficult to reconcile with its central estimate of –1.20%. Perhaps both discontinuations and their treatment effects and intensifications and their treatment effects are included in the year 1 estimates, though this timing could be questionable given the annual cycle of the model. It remains difficult to reconcile the above with the central estimates for treatment effectiveness.

At central estimates around one-third of the canagliflozin 300 mg patients would be required to not receive a boost of the –0.59% estimate for the intensification to gliclazide to achieve the reduction of 1.4% shown in the above figure. This also requires that none discontinue, rebound and receive a lesser treatment effect from whatever alternative to which they switch. Any discontinuations in the canagliflozin 300-mg arm would seem to require an even larger boost to the 1.2% canagliflozin 300-mg treatment effect among those not discontinuing if the reduction of around 1.4% is to be arrived at.

Even with the effects of second-order sampling, it is not clear to the AG how the above central estimates for the evolutions of HbA_1c have been arrived at. Perhaps there are additional placebo treatment effects in addition to the treatment effects relative to placebo and the AG has not managed to identify these. In light of the above, while having read the Janssen submission and its appendices, the AG does not really understand how the model is implementing the changes in HbA_1c and how the central estimate for canagliflozin 300 mg is to reduce the patient HBA_1c, on average, from 8.016% at baseline to around 6.63% at 1 year: a reduction of around 1.4%.

The above figure also sees the HbA_1c values converge between the arms. Janssen states that this is ‘Because of differences in the timing of requirements for rescue medication, HbA_1c, SBP, and lipid curves tend to converge as patients with higher values benefit from treatment-related improvements earlier’. To the AG this does not obviously explain the convergence, or why the curve for gliclazide converges by essentially having zero increase from the tenth to the thirtieth year. A possible explanation might be the intensification to insulin with patients then being permitted to increase their insulin dose in order to stabilise their HbA_1c.

The argument that the UKPDS68 equation 11193 includes the effects of treatment intensifications does have some force. But it should be borne in mind that the UKPDS68 equation 11193 explicitly includes a parameter for whether or not the patient is in his/her second year of diagnosis. This could be viewed as a proxy for the clinical effectiveness of the first treatment for diabetes being introduced, though the second year of diagnosis might be a little early for some patients. As a consequence, modelling could as an alternative apply treatment-specific effects and still apply the UKPDS68 equation 11193 thereafter, only ignoring the parameter related to whether or not the patient is in his/her second year. For the patient baseline characteristics outlined in the Janssen submission, the annual increases implied by the UKPDS68 equation 11193 in the years shortly after the second year are around 0.18%, which is broadly central to the rates Janssen takes from the ADOPT study as reported in Kahn et al. 174 This may be preferable to extrapolating linear rates from the ADOPT trial, particularly as these rates are being applied to treatments which were not used in the ADOPT trial, and to periods beyond the 5-year follow-up of the ADOPT trial, and with some distinctly ad hoc averaging for the rates for dual therapy.

The AG has some sympathy with the argument that HbA_1c drift may initially be treatment specific. A linear evolution might even be the most reasonable functional form, particularly given the coefficients reported by Kahn et al. 174 But the AG does not view it as reasonable to assume a linear evolution of HbA_1c throughout and that there will be no convergence between treatments, or at least none until a patient starts insulin therapy. This may artificially preserve differences between treatments, when the UKPDS68193 evolution clearly implies a convergence. The AG is also uncomfortable with the assumed linear rates of drift that have been imputed from Kahn et al. ,174 given that none of the monotherapies under consideration was studied by Kahn et al.

A scenario analysis where the model applies the UKPDS68193 evolution of HbA_1c was presented.

Evolution of other risk factors

Linear drifts were also assumed for the other biomarkers but these were not differentiated by treatment: 0.30 mmHg for SBP, 0.03 mg/dl for lipids resulting in a flat TC/HDL evolution. These annual rates of drift were apparently derived from the UKPDS. The UKPDS was conducted largely before the use of statins, and it can be argued that alternative evolutions to those of the UKPDS are now appropriate. But given that the annual rates of drift were apparently derived from the UKPDS, it is unclear to the AG why the equations of the UKPDS68193 were not applied.

For a patient’s SBP, the common annual rate of drift will tend to maintain the absolute differences between the arms. As far as the AG can ascertain, it also appears that unlike HbA_1c this absolute difference for SBP will be maintained even after insulin therapy has been started.

Weight was associated with an annual gain of 0.1 kg. Table 12 of the submission states that weight drifts upwards but there is no suggestion of a base-case assumption of convergence; table 13 goes on to suggest that for the base case the patient’s weight was assumed to converge at treatment discontinuation, with this being in line with figure 26 of the appendix. If this convergence applied, it is unclear whether it was at the first intensification or was when the oral therapies were being discontinued and the patient switched to insulin. A scenario analysis of a slower convergence over 2 years after treatment discontinuation was also presented.

Treatment discontinuation: renal impairment

In accordance with the canagliflozin Summary of Product Characteristics (SmPC), canagliflozin 100 mg was modelled as being discontinued if the eGFR fell below 45 ml/minute/1.73 m²; canagliflozin 300 mg was modelled as being discontinued if the eGFR fell below 60 ml/minute/1.73 m². The SmPC states that if the eGFR falls below 60 ml/minute/1.73 m² the patient should have his/her dose adjusted to 100 mg. Those intensifying from canagliflozin 100 mg to canagliflozin 300 mg would have already failed on canagliflozin 100 mg. The licence states that canagliflozin 100 mg should always be used before canagliflozin 300 mg. So dose reduction may be of limited relevance and discontinuation of canagliflozin 300 mg may be the appropriate assumption.

Empagliflozin has similar restrictions in its SmPC, with it being possible to increase the dose from 10 mg to 25 mg if the eGFR is more than 60 ml/minute/1.73 m². If the eGFR falls below 60 ml/minute/1.73 m² the patient should have his/her dose adjusted to 10 mg, and if the eGFR falls below 45 ml/minute/1.73 m² the patient should discontinue empagliflozin.

The dapagliflozin SmPC has slightly different restrictions and is not recommended for patients with an eGFR of less than 60 ml/minute/1.73 m².

The sitagliptin SmPC does make the administered dose depend on renal impairment, requiring it to be reduced to 50 mg in those with moderate renal impairment, and 25 mg in those with severe renal impairment. But discontinuation does not appear to be required.

As far as the AG can see the submission does not state what assumptions, if any, have been made about discontinuing empagliflozin and dapagliflozin based upon the patient’s eGFR. Table 12 of the submission is explicit in its consideration of the canagliflozin SmPC for discontinuations related to renal impairment and to the pioglitazone SmPC for discontinuations related to CHF, but makes no reference to the empagliflozin SmPC or the dapagliflozin SmPC. Appendix 4 of the submission mentions that empagliflozin and dapagliflozin also have treatment rules based upon eGFR. The input sheets to the electronic model suggest that empagliflozin 10 mg, empagliflozin 25 mg and dapagliflozin are assumed to be discontinued as per canagliflozin: if the eGFR drops below 60 ml/minute/1.73 m². This appears to be incorrect for empagliflozin 10 mg, and illogical for empagliflozin 25 mg.

The use of the ECHO-T2DM model was, in part, justified by Janssen on grounds of the need to properly account for discontinuations due to renal impairment. In light of this, a scenario analysis that does not apply these discontinuations would have been useful in order to assess the importance of attempting to model this.

Treatment discontinuations: adverse events

Patients could discontinue treatments as a result of adverse events. It is not clear to the AG what was assumed to happen to these patients. They may have discontinued their current treatment and switched to whatever is next in the sequence, for example from flozins to SU and then on to NPH insulin, or they may switch to an alternative monotherapy, or they may have switched to an alternative dual therapy.

Hypoglycaemic events

Hypoglycaemia event rates were derived from the pooled 26-week data of the two canagliflozin trials. The NMA 26-week data provided the estimates for the other comparators with the exception of gliclazide. No hypoglycaemia rates were available for gliclazide, and, as a consequence, rates for glimepiride were adjusted by relative risks of 0.43 for symptomatic hypoglycaemia and 0.45 for severe hypoglycaemia. The AG is unclear whether or not these rates were adjusted to be annual rates and so to be in line with the annual model cycle.

Hypoglycaemia event rates were further modified by increasing the risk of hypoglycaemia for low HbA_1c values. The relationship underlying this was based upon a large data set of the Diabetes Control and Complications Trial (DCCT) study among patients with type 1 diabetes. A 1% drop in HbA_1c below the mean value of the clinical study was associated with an increased hazard of hypoglycaemia of 1.43. However, patients in the DCCT intensive arm were on either MDI or continuous subcutaneous insulin infusion (CSII) via insulin pumps.

Adverse events

The UTI and GTI event rates were derived from the pooled 26-week data of the two canagliflozin trials. The AG is unclear whether or not these rates were adjusted to be annual rates and so to be in line with the annual model cycle. The other flozins were assumed to have the same rates as canagliflozin 100 mg, with the other comparators being assumed to have the same rate as the pooled placebo arms of the canagliflozin trials.

Quality of life

A systematic literature review was conducted, with Janssen preferring the CODE-2 (Cost of Diabetes in Europe–Type 2) data of Bagust and Beale199 over the UKPDS62198 as a result of it providing greater richness for the microvascular complications. Bagust and Beale199 is also the source for the QoL coefficient for BMI.

Janssen reports that no appropriate studies were identified for adverse events, and as a consequence a time trade-off (TTO) study among 100 members of the UK general public was conducted to determine the QoL impacts from UTIs and GTIs. This TTO study also explored hypoglycaemia, GI symptoms and hypovolaemic events, but the estimates for these were disregarded (Table 21).

Based upon the references cited, these values apparently contributed to a regression analysis, which arrived at the final QALY decrements. Unfortunately, the AG has not been able to source this regression analysis.

The health state descriptors suggest that the estimates relate to ongoing infection and are not time limited, as is appropriate in a TTO study, but this implies that event health states need to be adjusted by the average duration of UTIs and GTIs, as occurs in the Janssen submission. The method of this adjustment does not appear to have been presented. At mean values, 1 week of a moderate UTI would roughly correspond with a –0.0012 QALY decrement; 2 weeks of a severe UTI would roughly correspond with a –0.0073 QALY decrement; and 1 week of a mycotic infection would roughly correspond with a –0.0046 QALY decrement.

The QoL values applied for the baseline characteristics and microvascular complications are presented in Table 22. Those for macrovascular complications, obesity, hypoglycaemic events and adverse events are presented in the overarching summary comparison of the companies’ and AG’s inputs (see Table 22).

Costs

Direct drug costs were sourced from BNF 69147 and are not presented here for reasons of space. Note that the Janssen analysis, in common with the other company analyses, applies the £608 cost for canagliflozin 300 mg as a result of the submission predating the equalisation of the canagliflozin list prices at the canagliflozin 100 mg list price of £477.

The costs of blindness, IHD, MI, CHF and stroke were derived from the UKPDS84195 but are not presented here again for reasons of space. Similar costs are presented within the section on the AG modelling. A variety of other costs are sourced from a variety of NICE guidelines and other sources. Again, for reasons of space, and because they have very little impact upon the modelling, these are not presented here. A full table of event costs is presented in the appendices of the Janssen submission in table 33, starting on p. 69. Costs are also summarised in the overarching comparison of the companies and AG modelling exercises.

Adverse event costs are given in Table 23.

Some additional costs for GTIs might be anticipated if a patient’s partner is also treated.

Results

The QALY losses in the model are presented in tables 42 and 43 of the appendices, which are summarised in Table 24. The AG interpretation of this is that canagliflozin 100 mg has been taken as the reference for survival with the absolute QALY losses associated with the various events also being presented. The net absolute QALY difference between canagliflozin 100 mg and each of the comparators is then presented, with the percentage contribution of the various events to this net QALY effect being presented alongside. As the complications of diabetes other than neuropathy make little contribution to this, they have been grouped together for reasons of space.

The above shows that within the modelling survival differences account for a reasonable proportion of the estimated differences in mean QALYs between the comparators: about one-quarter for the non-canagliflozin comparators. But the direct QoL impacts of weight, hypoglycaemia and, to a lesser extent, neuropathy account for the majority of the differences. The QALY losses from the other complications of diabetes are relatively insignificant. A similar analysis can be presented for the cost differences (Table 25).

As can be seen from Table 25, the main differences in cost arise from different treatment costs for both the oral drugs and insulin, with these being in part driven by the survival differences alluded to above. The only real exceptions to this are for the comparison with pioglitazone, for which an additional £156 treatment cost for CHF is anticipated compared with canagliflozin 100 mg.

Ranking treatments in order of increasing total costs, the Janssen base-case cost-effectiveness results are shown in Table 26. Note that the following table presents the ICERs relative to the cheapest comparator among those being considered. Not all of these ICERs are presented by Janssen, with some having been derived by the AG, so are subject to rounding errors (see Table 26).

Pioglitazone is the cheapest owing to its acquisition cost. As a consequence, although other treatments are estimated to be more effective compared with pioglitazone their cost-effectiveness is poor. Both gliclazide and sitagliptin are estimated to be dominated by it. Sitagliptin being dominated by pioglitazone may be due to its assumed faster rate of HbA_1c drift. Canagliflozin 100 mg is estimated to have an ICER of £79,537 per QALY compared with pioglitazone. Canagliflozin 100 mg also provides more QALYs at a cheaper cost than empagliflozin and dapagliflozin, so dominating them. Canagliflozin 100 mg followed by canagliflozin 300 mg has a better cost-effectiveness against pioglitazone than canagliflozin 100 mg, so extendedly dominates canagliflozin 100 mg. But it is, in turn, extendedly dominated by canagliflozin 300 mg, which is estimated to have a cost-effectiveness compared with pioglitazone of £47,456 per QALY.

The cost-effectiveness of sitagliptin and the flozins is estimated to be more reasonable when compared with gliclazide. But canagliflozin 100 mg extendedly dominates sitagliptin and dominates empagliflozin and dapagliflozin.

Across the nine comparators, the probabilistic modelling suggested that if the willingness to pay is zero, pioglitazone has a 100% probability of being the most cost-effective treatment. This probability declined as the willingness to pay increased, until at around a willingness to pay of £55,000 it ceased to have the highest probability of being the most cost-effective treatment. At this point canagliflozin 300 mg overtook it, with a probability of being the most cost-effective of around 25%.

At high willingness-to-pay values of £200,000 per QALY it appears that the probabilities of being the most cost-effective have stabilised. Canagliflozin 300 mg remains the highest with a probability of around 33%.

The others increase their probabilities as the willingness to pay rises, but still converge only to values under 20%, with the values for empagliflozin 10 mg, dapagliflozin 10 mg, sitagliptin 100 mg and gliclazide never rising above 10%.

Sensitivity analyses

The univariate sensitivity analyses presented by Janssen vary parameters by an arbitrary ±20% and are consequently of limited interest. Full results of these are presented in tables 46–48 of the appendices to the submission. The main result of interest is that the modelling is sensitive to the annual rate of HbA_1c drift that is assumed for canagliflozin: deterministic sensitivity analyses (DSAs) number 6 lower value and upper value (6L and 6U), which, respectively, decrease and increase the base case 0.14% annual rate of drift by 20%. The cost-effectiveness estimates under DSA 6L and DSA 6U for canagliflozin 100 mg compared with:

• pioglitazone are £45,862 per QALY and £211,000 per QALY, respectively

• gliclazide are £593 per QALY and £8751 per QALY, respectively

• sitagliptin are dominance and £8528 per QALY, respectively.

In the opinion of the AG these changes are likely to be due more to the time spent on therapy and its immediate effects upon treatment cost, weight, adverse events and hypoglycaemia than to any changes in the modelled complications of diabetes. An exception to this might be the modelled rates of neuropathy.

Scenario analyses

A range of scenario analyses as summarised in table 13 on p. 52 of the Janssen submission is presented, with summary results for all of these in tables 19 and 20 of the Janssen submission. The main points of interest identified by the AG are summarised below.

For the comparison with empagliflozin and dapagliflozin the main scenario analyses of interest are those that:

• revise the patient characteristics at baseline from those that Janssen pools from its trials to those of The Health Improvement Network (THIN) database, which is the database that underlies the patient characteristics of the modelling for the draft NICE CG: Sc5. The THIN database has anonymised patient data from 562 general practices covering 6.2% of the population. 200

• apply the UKPDS68193 HbA_1c evolution equation and UKPDS62198 QoL values, while also assuming that patients can intensify to NPH insulin but not to basal-bolus insulin: Sc6; and,

• apply the UKPDS68193 HbA_1c evolution equation: Sc14.

These scenario analyses remove the dominance of canagliflozin 100 mg over empagliflozin and dapagliflozin with the cost-effectiveness estimates typically changing to lie between £5000 and £10,000 per QALY.

Scenario analysis 14 is rather more dramatic in terms of the cost-effectiveness estimates. But it would probably be more accurate to describe it as showing broad clinical equivalence, but additional costs from canagliflozin compared with dapagliflozin, empagliflozin 10 mg and empagliflozin 25 mg of £198, £150 and £65, respectively.

Patient characteristics

Patient characteristics at baseline were mainly drawn from the NMA, resulting in baseline estimates of 55 years of age, 54.6% male, 7.5% HbA_1c due to NICE CGs, though the NMA mean of 8.2% was used as a scenario analysis, 128.3 mmHg SBP, 195 mg/dl TC, 46 mg/dl HDL and a weight of 80 kg.

The submission does not appear to state what the baseline prevalence of the complications of diabetes was. The submitted electronic model sets these to zero.

Sequences modelled and treatment effectiveness

The comparators were grouped into their class as per the AstraZeneca NMA, for example the cost-effectiveness of flozins as a group was estimated compared with the gliptins as a group. Note that of the glitazones only pioglitazone was considered, rather than a class effect being applied. Repaglinide was not considered as a comparator owing to a lack of evidence.

AstraZeneca argues that allowing the intensifications to differ between the arms would not permit a fair assessment of the cost-effectiveness of alternative monotherapies, but would rather be a comparison of the cost-effectiveness of alternative treatment sequences. Consequently, the AstraZeneca submission modelled the following treatment sequences despite this not reflecting UK clinical practice (Table 28).

Intensified insulin was assumed to involve a 50% dose escalation.

Clinical effectiveness estimates for the monotherapies were drawn mainly from the 24-week NMA. Infection rates were not meta-analysed, but were drawn from a weighted pooled mean of incidence data at 24 weeks from the papers included in the NMA. Clinical effectiveness estimates for insulin were drawn from Monami et al. 201 for NPH and from Waugh et al. 202 for intensified NPH (Table 29). (The AG does not know how these figures for ‘intensified NPH’ were obtained. Usually if NPH was insufficient, short-acting insulin would be added at mealtimes.)

Treatment intensifications and discontinuations

A patient is modelled as intensifying treatment, first to NPH and then to intensified NPH, when their HbA_1c breaches the 7.5% intensification threshold. The AG assumption is that the monotherapies are withdrawn at treatment intensification, but this is not explicit within the AstraZeneca submission.

Patients may also discontinue due to adverse events. The AG was unable to identify what was assumed for these patients: whether they switched to an alternative monotherapy and, if so, which, or they intensified to NPH insulin.

Glycated haemoglobin evolution

In common with the AG modelling, the evolution of HbA_1c is based upon equation 11 of the UKPDS68. 193 Treatment intensification occurs if a patient’s HbA_1c breaches the 7.5% intensification threshold. This leads to a sawtooth evolution of HbA_1c, as described in more detail in the section on the AG modelling.

Evolution of other risk factors

The evolution of SBP and the TC/HDL ratio was also based upon equations 12 and 13 of the UKPDS68. 193 The section on the AG modelling describes this in some detail so it is not further described here.

Weight loss with the flozins is assumed to be maintained for 2 years, after which it is assumed that patients rebound to their starting weight. A similar assumption appears to have been made for the gliptins, though the weight loss is maintained for only 1 year. The weight increases associated with the other treatments are assumed to be retained. Weight is also assumed to increase by 0.1 kg annually.

Quality of life

The QoL for a patient without any complications is a function of patient age, as drawn from analysis of EuroQol-5 Dimensions (EQ-5D) data from the Health Survey for England 2003:203

This results in a baseline QoL of 0.882, with this slowly declining over time.

Quality-of-life decrements associated with the complications of diabetes were drawn from the UKPDS62198 with the exception of that for ESRD, which was drawn from the standard UKPDS OM1 source. QoL decrements for hypoglycaemic events were drawn from Currie et al. 204 but note that it appears that the coefficient for symptomatic event was applied to the number of symptomatic events rather than to their logarithm. The QoL impacts of increasing BMI was drawn from Bagust and Beale. 199 These sources and values are all as per the AG modelling, so are not tabulated here for reasons of space.

Note that the submission suggests that the BMI disutility is applied for all BMI changes and is not limited in its effects to changes in BMI when the patient BMI is greater than 25 kg/m². If this applies it may have biased the analysis in favour of the flozins by valuing reductions on patients’ BMI among those with a BMI of less than 25 kg/m². But given the mean BMI at baseline of 29.2 kg/m² this may not be a particular concern.

A systematic literature review was conducted for UTI and GTI QoL decrements. For UTIs the average of the values of Barry et al. 205 of –0.3732 for pyelonephritis and of –0.2894 for dysuria appears to have been coupled with an assumed duration of around 3 days to yield a QALY decrement of –0.00283 per UTI. Apparently no values were found for GTIs and, as a consequence, these had the same disutility applied.

Costs

The direct drug costs were sourced from the BNF 69. 147 For both the flozins and the gliptins, weighted average costs based upon their UK market share were used. This resulted in a mean annual flozin cost of £482 and a mean annual gliptin cost of £429. The annual cost of pioglitazone was £19 and the annual sulfonylurea cost was based upon gliclazide at an annual cost of £66. The cost of gliclazide suggests to the AG that MR gliclazide has been assumed, as the standard version would be around half of the cost that was applied.

The costs of the complications of diabetes in the first year and for subsequent years for blindness and amputation were based upon the UKPDS84. 195 This is the same source as the AG, though the AG arrives at somewhat lower values. There may be a suggestion that indexation by AstraZeneca was based on 2007 prices, when the UKPDS84195 is in 2012 prices. But the source of the discrepancies is unclear.

The AG calculations suggest that the UKPDS84195 average inpatient costs and outpatient costs for those without any of the modelled complications have not been included within the AstraZeneca modelling. If this is the case it would be a quite serious omission, and would tend to bias the analysis in favour of the more effective treatment.

AstraZeneca may have used the UKPDS84195 bespoke costing template to derive costs for a representative baseline patient, but this seems unlikely to be the source of the discrepancies between AstraZeneca and the AG. The AstraZeneca mean age at baseline is 55 years, whereas the AG, in order to be able to implement the costs probabilistically, has taken the costs example of the UKPDS82194 for a 60-year-old man. Costs are typically increasing with age in the UKPDS84. 195

Table 5.10 of the AstraZeneca submission also does not include a cost for fatal IHD events despite these being within the UKPDS84195 and seeming to be associated with deaths in the UKPDS82194 and the UKPDS OM2. The UKPDS65,196 which goes along with the UKPDS68193 and the UKPDS OM1, does not itemise a cost for fatal IHD events. But the AG understanding is that the AstraZeneca CDM modelling is based upon the UKPDS82194 and, as a consequence, does not understand why fatal IHD events have had a zero cost assigned.

For reasons that are unclear, AstraZeneca chose to revert to the costs of the UKPDS65196 for the ongoing costs among those with a history of IHD, CHF and stroke, and probably MI as well. This seems peculiar to the AG, given that the UKPDS82194 and the UKPDS65196 are very similar in their format, with the UKPDS82194 also presenting cost estimates for those with a history of IHD, CHF, stroke and MI.

End-stage renal disease was costed using the estimate of Baboolal et al. 206 for continuous ambulatory peritoneal dialysis. Previous NICE assessments have also used this reference, though have also tended to use the higher cost estimates within Baboolal et al. 206 for hospital haemodialysis. AstraZeneca argued that the use of the peritoneal dialysis cost was conservative.

Severe hypoglycaemia was costed using the Hammer et al. (2009) reference,207 which is the reference used for the AG modelling. UTIs and GTIs were assumed to involve one GP appointment, costed at £46 using the Personal Social Services Research Unit (PSSRU) Unit Costs of Health and Social Care 2014208 (Table 30).

Results

The CDM modelling for the base-case results is shown in Table 31. The AG assumption is that this is based upon deterministic modelling, that is with no second order sampling.

Within the pairwise comparisons, compared with the sulfonylureas the flozins offer some additional benefit of 0.027 QALYs, but there are reasonable additional costs of £1397 associated with this resulting in a cost-effectiveness estimate of £52,047 per QALY.

The gains from the flozins compared with pioglitazone are larger at 0.095 QALYs, which may be sufficient to justify the additional cost of £1912, which results in a cost-effectiveness estimate of £20,089 per QALY.

When compared with the gliptins, the flozins provide only a small additional gain of 0.018 QALYs, but this is also at a relatively modest additional £106 cost, which results in a cost-effectiveness estimate of £5904 per QALY.

The probabilistic pairwise modelling suggests that the flozins have a probability of being cost-effective (using the NICE threshold of £30,000 per QALY) compared with the gliptin of 66%, compared with pioglitazone of 51% and compared with the sulfonylureas of 13%.

Ranking results in order of increasing cost are provided in Table 32.

As would be anticipated from the pairwise comparisons, it appears that the sulfonylureas have an acceptable cost-effectiveness compared with pioglitazone of £7574 per QALY. The gliptins offer minimal patient benefit compared with the sulfonylureas, 0.009 QALYs or the equivalent of around an additional 4 days’ survival, but with a reasonable increase in costs of £1291 and a cost-effectiveness estimate of £143,000 per QALY. As already noted, the flozins cost-effectiveness estimate compared with the gliptins is good at £5904 per QALY, but is poor against the sulfonylureas at £52,047 per QALY.

A variety of univariate sensitivity analyses were presented, which varied the clinical effectiveness estimates to their upper and lower CI limits, the disutilities for BMI changes to their upper and lower CI limits; the disutilities for complications by ±10%; and the total non-drug costs by ±25%.

Only the changes to the BMI disutility had any marked impact, with these impacts being mainly for the comparisons with pioglitazone and sulfonylurea. The lower confidence limit for the disutility of weight gains improved the cost-effectiveness estimate compared with pioglitazone from £20,089 per QALY to £14,626 per QALY, whereas the upper confidence limit worsened it to £32,065 per QALY. The lower confidence limit for the utility of weight losses worsened the cost-effectiveness compared with the sulfonylureas from £52,047 per QALY to £62,810 per QALY, whereas the upper confidence limit improved it to £4434 per QALY.

Owing to the CDM of AstraZeneca modelling only pairwise comparisons, the probabilistic modelling is presented only for the pairwise comparisons.

Compared with the gliptins, at willingness-to-pay values of £0, £20,000 and £30,000 per QALY the flozins are estimated to have a probability of being the most cost-effective of around 42%, 65% and 68%, with the cost-effectiveness acceptability curve (CEAC) converging to a little over 70% at high willingness-to-pay values.

Compared with pioglitazone, at willingness-to-pay values of £0, £20,000 and £30,000 per QALY the flozins are estimated to have a probability of being the most cost-effective of around 0%, 50% and 80%, with the CEAC converging to a little over 95% at high willingness-to-pay values.

Compared with sulfonylurea, at willingness-to-pay values of £0, £20,000 and £30,000 per QALY the flozins are estimated to have a probability of being the most cost-effective of around 0%, 12% and 26%, with the CEAC still slowly increasing to a little over 60% at a willingness to pay of £100,000 per QALY.

Scenario analyses

A range of scenario analyses were undertaken, mainly varying the HbA_1c values at baseline and HbA_1c thresholds for intensifying treatment, altering the assumptions around maintenance of weight effects and the drug costs that were applied (Table 33).

The scenario analyses around adverse events and discontinuations for the comparison with pioglitazone were reported as having the same values as the corresponding analyses for sitagliptin, so appear to be typographical errors.

Less stringent thresholds for intensification of therapy tended to worsen the cost-effectiveness estimates. This seems likely to be mainly due to patients remaining on their monotherapy for longer and the associated increase in the direct drug costs, and not due to differences in the modelled complications of diabetes.

Weight convergence between the flozins and pioglitazone somewhat worsens the cost-effectiveness estimate, because of this removing the weight gains associated with pioglitazone. Note that this convergence is imposed only at around the seventh or eighth year of the modelling.

Results compared with the gliptins are not particularly sensitive to whether the sitagliptin price is applied rather than the gliptins’ prices weighted by their market shares, due to AstraZeneca estimating sitagliptin to have a market share of the gliptin monotherapy of 56% with the similarly prices saxagliptin and linagliptin having market shares of 16% and 25%, respectively. Alogliptin is notably cheaper but is estimated to have less than a 1% market share.

Patient characteristics

Patient characteristics were drawn from patients within the Clinical Practice Research Datalink (CPRD). This identified 9211 UK patients who started their first oral antidiabetic treatment in 2014. Although not all of the codes used for the search appear to have been specific to type 2 diabetes, for example diabetes mellitus, the minimum age at diagnosis was 23 years, with a mean of 60 years and a SD of 12 years, and the requirement to be starting an oral therapy should have restricted the sample to patients with type 2 diabetes.

The average duration of diabetes was 2.9 years, 57% being male. The mean HbA_1c was 8.49%, SBP 134 mmHg, HDL 1.2 mmol/l and LDL 4.0 mmol/l. The mean BMI was 31 kg/m².

The presence of existing complications was included: 6.63% for atrial fibrillation, 3.18% for PVD, 2.21% for MI, 1.92% for CHF, 1.62% for stroke, 6.13% for IHD, 0.29% for amputation, 0.23% for blindness and 0.05% for renal failure.

Sequences modelled and treatment effectiveness

Model A considers only the first year of treatment with monotherapy and then adds the UKPDS costs and complications to this.

Model B considers the treatment sequences in Tables 34 and 35 with patients intensifying to the next line of therapy when their HbA_1c breaches the 7.5% intensification threshold.

Pioglitazone 45 mg was chosen because of it being the most commonly prescribed dose. The differences in cost between pioglitazone 30 mg and pioglitazone 45 mg are minimal.

Note that if repaglinide was insufficient, it would be replaced, as for dual use it is licensed only with metformin. So it would not be logical to add gliclazide to repaglinide since they act largely on the same receptors. Note also that 1 mg is a small dose of repaglinide.

Clinical effectiveness data were based upon 52-week data where available, though this was apparently not available for canagliflozin or dapagliflozin. Sitagliptin clinical effectiveness estimates were apparently based upon 24-week data, though the submission does not state that 52-week data were not available. The effect upon SBP and rates of UTIs were also based upon 24-week data because of 52-week data not being available.

Hypoglycaemia rates were based upon the sulfonylurea 16.4% annual rate from the NMA (for pooled sulfonylureas) coupled with odds ratios for each of the comparators against sulfonylurea. A ratio of non-severe to severe hypoglycaemia event rate of 17.2 was based upon the 0.009 annual rate of severe hypoglycaemia event rates of Leese et al.,18 coupled with the overall rates of hypoglycaemia in the NMA.

Urinary tract infection event rates were based upon the annual placebo rates of 3.5% from the NMA coupled with odds ratios for each of the comparators.

Treatment intensifications and discontinuations

Within model A there are no treatment intensifications. Treatment with the monotherapies is for 1 year only, after which it appears that the UKPDS OM1 model is appended to this.

Within model B, treatment is intensified when a patient is modelled as breaching the 7.5% HbA_1c threshold.

It appears that discontinuations of therapy for reasons other than breaching the HbA_1c threshold of 7.5% have not been modelled.

Glycated haemoglobin evolution

The evolution of HbA_1c is based upon equation 11 of the UKPDS68. 193

Within model A it appears that the patient’s baseline HbA_1c has the treatment effect of the initial monotherapy applied, with the UKPDS OM1 and hence relevant equation of the UKPDS68193 being used to model its evolution thereafter. But, as stated in the introduction, the AG has not been able to identify how model A interacts with the UKPDS OM1.

Within model B it appears that the patient’s baseline HbA_1c has the treatment effect of the initial monotherapy applied. Equation 11 of the UKPDS68193 is then used to model the evolution of the patient’s HbA_1c until it breaches the 7.5% intensification threshold, at which point the treatment effect of the first intensification is applied. Equation 11 of the UKPDS68193 is then applied until the second intensification occurs with the associated treatment effect. HbA_1c is then modelled as progressing as per equation 11 of the UKPDS68. 193

Evolution of the other risk factors

The evolution of SBP and the TC/HDL is based on the UKPDS OM1 and hence on the relevant equation of the UKPDS68. 193

Within model A, it appears that the patient’s baseline SBP had the treatment effect of the initial therapy applied, with equation 12 of the UKPDS68193 being used to model its evolution thereafter. For the TC/HDL ratio, as a result of there being no treatment effects estimated, it appears that equation 13 of the UKPDS68193 was used to model the evolution throughout. But, again, as stated in the introduction, the AG has not been able to identify how model A interacts with the UKPDS OM1.

Within model A the direct impacts of weight changes upon QoL were evaluated only during the first year. For the UKPDS modelling it appears that weight losses were assumed to rebound to baseline after 1 year, whereas weight gains were assumed to remain indefinitely.

Within model B it appears that the patient’s baseline SBP had the treatment effect of the initial therapy applied, with equation 12 of the UKPDS68193 being used to model its evolution except for when a treatment intensification took place at which point the treatment effect of the intensification was applied. For the TC/HDL ratio, as a result of there being no treatment effects estimated, it appears that equation 13 of the UKPDS68193 was used to model the evolution throughout.

Within model B, weight losses from treatment were assumed to apply at 52 weeks and then to rebound to baseline at 104 weeks. Weight gains from treatment were assumed to be maintained indefinitely. A 0.1-kg annual weight gain from natural history was also applied.

Quality of life

The QoL at baseline and the QoL decrements associated with the complications of diabetes were drawn from the recent paper by Alva et al. ,197 which reanalysed the updated UKPDS data set and in some sense updated the values of the UKPDS62,198 which is the paper that the AG modelling relies upon. The values and a commentary upon this are presented later in the comparison of modelling inputs used by the companies and the AG.

In common with the other companies, the QoL impact of hypoglycaemic events was drawn from Currie et al. 204 Similarly, the QoL decrement of –0.0061 per BMI point above 25 kg/m² was drawn from Bagust and Beale. 199

The QoL decrement of –0.00283 per UTI was based upon the estimates of Barry et al. 205

Costs

Treatment costs were based upon the March 2015 Monthly Index of Medical Specialities (MIMS).

The costs of diabetes without complications and the costs of the complications of diabetes were taken from the UKPDS84. 195 Boehringer Ingelheim appears to have applied only the inpatient costs of the UKPDS84195 and ignored the outpatient costs.

A cost of £380 per severe hypoglycaemic event was drawn from the draft NICE CG for type 2 diabetes, which is similar to that of the other company submissions and AG value.

Urinary tract infections were associated with a £36 cost, based upon the ERG report52 for the previous STA of dapagliflozin for type 2 diabetes combination therapy.

Note that the UKPDS costs of model B that are reported in Tables 36 and 37 are around half of those of model A. The reason for these discrepancies is far from obvious.

Model

The protocol specified that either the UKPDS OM1 or the UKPDS OM2 would be used by the AG. For some of its outputs the OM1 is quite different from the OM2 in its predictions. But the OM2 was not made available to the AG in time for the assessment and so, as specified in the protocol, the OM1 has been used.

As already noted, the OM1 predicts roughly double the number of MIs over 10 years, and the rates of IHD are also noticeably higher than those of the OM2. The 10-year mortality is also higher with the OM1. Compared with the OM2, the OM1 will tend to overpredict event rates and so overstate the benefits and cost savings arising from any avoidance of the complications of diabetes that are associated with the more effective treatment. Being more recent and more reflective of current practice, the OM2 would consequently have been much preferable had it been available to the AG.

The OM1 was used for the modelling that underlies the NICE CG for diabetes. During its development the Guideline Development Group (GDG) reviewed in detail 10 type 2 diabetes cost-effectiveness models. These included the JADE209 and CORE210 models, but not the ECHO-T2DM model. Based upon validation and consistency with the NICE reference case, the GDG very much preferred the OM1, in no small part because of it being based upon a single RCT rather than drawing a range of modelling inputs from disparate sources.

The AG has developed a front and back end to the OM1. Briefly, for each patient and treatment strategy that is simulated the AG front end models the patient’s progression from monotherapy through the various treatment intensifications over a 40-year time horizon in annual cycles. This, in turn, introduces the patient’s evolutions of HbA_1c, SBP, TC/HDL, BMI, hypoglycaemia event rates, adverse events and treatment costs. The evolutions of the patient’s HbA_1c, SBP and TC/HDL are then fed into the OM1, which models the complications of diabetes and patient lifespan, and outputs the costs and QoL impacts of living with diabetes and the patient’s survival curve. The AG back end takes the OM1 survival curve and uses this to condition the evolutions of the patient’s BMI, hypoglycaemia event rates, adverse events and treatment costs. The cost and QoL impacts of these are then summed with the cost and QoL impacts outputted by the OM1.

In slightly more detail, patients start on monotherapy but intensify their treatment if their HbA_1c is modelled as breaching the 7.5% threshold. Intensifications typically add another treatment to a patient’s existing treatment(s). This permits treatment sequences to be modelled, starting with monotherapy but with subsequent treatment intensifications, these intensifications eventually leading to first basal insulin and then basal-bolus insulin. Each treatment within a sequence is associated with treatment costs, weight changes, hypoglycaemic events and adverse events. The AG modelling also permits treatments to be associated with a discontinuation rate in their first year, with patients who discontinue being assumed to switch to another treatment at the same line of therapy.

For each patient who is modelled, the modelled treatment sequences lead to a modelled evolution of HbA_1c, SBP and TC/HDL ratio. These, together with the patient’s baseline characteristics, are fed into the OM1. The OM1 then models the rates of the complications of diabetes, such as CHF, and the patient survival, which results in estimates for the discounted costs and QALY impacts of the complications of diabetes over the modelled lifetime of the patient. A survival curve is also drawn from the OM1 model. Owing to the model being an individual patient simulation, any given patient is run through the model many times, say 1000 inner loops, in order to reduce Monte Carlo sampling error. In effect this is the same as running a cohort of 1000 identical patients through the model. The OM1 survival curve is the proportion of this cohort, or 1000 inner loops, that is modelled as surviving. This survival curve conditions the AG front end evolutions of treatment sequences and the cost and QALY impacts of their treatment costs, weight changes, hypoglycaemic events and adverse events.

For the deterministic model run, the OM1 correctly outputs the relevant survival curve. Unfortunately, for the PSA iterations it appears that the OM1 does not output the relevant survival curve. As a consequence, the relevant survival curve has had to be imputed from the OM1 annual discounted QALY estimates by an initial run of the model with the baseline QoL set equal to unity and the QoL decrements of the complications all set to zero. The resulting annual discounted QALYs were then undiscounted to arrive at the patient-specific survival curve. (For the PSA each patient was run with 100 inner loops, and, as a consequence, the imputed survival curve had a granularity of 1%.) But this also meant that the PSA had to run the OM1 model twice for each strategy for a given patient, for a given PSA iteration. This also required that the same random number seed be used for each of these model runs in order for the imputed survival curve to be consistent with the second run of the model that estimated the strategy’s costs and benefits. The OM1 permits the random number seed to be only 1 of 100 values. The AG model randomly assigned this value during the PSA, keeping this value constant between the two model runs for a given patient, for a given PSA iteration. Having to run the model twice for a given patient for a given PSA iteration also significantly increased the time it took to run the PSA.

An element that the OM1 cannot address is the requirement for patients receiving a flozin to have their dose of it reduced or discontinued based upon renal function and eGFR rates. Although the AG has a number of issues with the Janssen modelling, the use of the ECHO-T2DM model did permit this to be explored, though the AG has not reviewed the implementation of this in any detail. It would be interesting to know the impact that turning off these discontinuations would have upon the cost-effectiveness estimates of the Janssen modelling. If this is significant enough to affect the conclusions that would be drawn from the Janssen modelling it could suggest additional modelling uncertainty from the AG use of the OM1.

The AG visual basic modelling has the advantage of permitting up to 12 treatment strategies to be simultaneously compared with one another, with the correlation between treatments’ effects being properly taken into account. Each PSA iteration also uses the same set of parameter values and random number seed across all of the treatment strategies being modelled. This, in turn, permits the correct characterisation of uncertainty within the probabilistic modelling.

Model runs

The NICE CG group for type 2 diabetes concluded after a number of model runs that with a patient cohort of 35,000 or more there was little to be gained from running more than 100 inner loops to reduce Monte Carlo sampling error. As a consequence, probabilistic results were based upon 1000 PSA iterations, each with a patient cohort of 50,000 with 100 inner loops. For deterministic model runs, that is those without any second order sampling, the modelling for the CG increased the number of inner loops to 1000, as recommended within the OM1 manual.

The AG has adopted the same approach.

Probabilistic sampling

The risk factor evolution parameters of equations 11–13 of the UKPDS68193 were received as 1000 bootstrap samples from the UKPDS group (Professor Alastair Gray, University of Oxford, June 2015, personal communication). The UKPDS OM1 also permits only 1000 bootstraps.

The other parameters within the modelling were sampled by the AG. Clinical effectiveness was sampled within the NMA, with this outputting 10,000 look-up values for the various clinical effectiveness parameters. But as a result of the OM1 permitting only 1000 bootstraps and the time taken to run the PSA, a subset of 1000 was sampled from the 10,000 look-up values of the NMA. It was checked that these subsets had means that were similar to the central estimates of the NMA.

Other parameters were sampled by the AG using the distributions outlined below.

Patient characteristics at baseline

The patient characteristics at baseline are taken from the draft NICE CG for diabetes. This undertook extensive analysis of the THIN database, supported by some additional data from the Health Survey for England (Table 38).

These were sampled once for the modelling using the full variance–covariance between the characteristics, as per the modelling for the current draft NICE CG for diabetes.

The AG has adopted these values with the exception of the baseline the baseline TC/HDL ratio. TC/HDL has been assumed to be 3.0 owing to NICE guidelines on atorvastatin use in people with diabetes. This change in therapy may be partly the cause of the differences between the OM1 and the OM2. A scenario analysis applies the values of the NICE CG and evolves these according to the UKPDS68193 equation 13.

AstraZeneca argued that the baseline HbA_1c should be 7.5% in order to be in line with NICE guidelines. But the patients modelled are starting their first drug treatment after, on average, having been diagnosed with diabetes for 2.0 years. The mean HbA_1c at diagnosis was estimated to be 8.2%. It seems unlikely that most patients will have successfully controlled their diabetes through diet and exercise and got below 7.5% if they were above it at diagnosis, only to subsequently lose this control. As a consequence, the base case will apply the baseline HbA_1c values as estimated within the draft NICE CG. A scenario analysis applies a common 7.5% HbA_1c at baseline across the 50,000 patients simulated.

Sequences modelled

As outlined in the assessment protocol, in line with NICE guidelines patients will intensify their treatment if their HbA_1c breaches the 7.5% intensification threshold. As a consequence, the modelling needs to take into account the clinical effects and costs of these intensifications. Based upon expert opinion the AG has modelled the treatment sequences in Table 39.

Clinical effectiveness

The clinical effectiveness estimates are drawn from the Warwick AG NMA as presented in the clinical effectiveness section and from a review of the literature. Events rates are annual unless otherwise stated (Tables 40 and 41).

For the flozins the UTI rates and GTI rates are half-yearly.

For the intensifications, because of a lack of data the addition of a treatment is assumed to have the same clinical effectiveness regardless of what it is being added to (Tables 42 and 43).

Adjusting the glycated haemoglobin effect for a patient’s baseline glycated haemoglobin

The NICE CG modelling estimated two alternative models for the change at 1 year in HbA_1c. The first corresponded to the base-case approach of the AG, though metformin was the reference treatment in the NICE CG NMA rather than placebo, as in the AG NMA.

• Estimate a reference treatment’s absolute change in HbA_1c from baseline, t₀, to the end of the first cycle, t₁: Δ_abs.

• Estimate the difference between the reference treatment and the other treatments at the end of the first cycle: ΔTx_rel.

• H₁ = H₀ + Δ_abs + ΔTx_rel.

For instance, suppose that the change between t₀ and t₁ for metformin Δ_abs = –1.49 and that the difference between metformin and canagliflozin at t₁ was ΔTx_rel = –0.51. A patient with a baseline H₀ = 9.00 would be estimated to have H₁ = 9.00 – 1.49 – 0.51 = 7.00, whereas a patient with a baseline H₀ = 7.00 would be estimated to have H₁ = 7.00 – 1.49 – 0.51 = 5.00.

But a strong correlation was observed between the trials’ metformin absolute effect between t₀ and t₁ and their mean baseline HbA_1c. As a consequence the NICE CG explored adding an additional term to Δ_abs to make the change also a function of the baseline HbA_1c, H₀. This led to the following adjusted model for the HbA_1c at the end of the first cycle:

with ΔTx_rel being taken from the NICE CG NMA.

The Δ_abs and β were not estimated during the NMA, but were separately estimated using the same data for metformin as was used in the NMA. The adjusted model simplifies to the unadjusted model by setting β = 0. This resulted in the following coefficients (Table 44), and the following unadjusted and adjusted treatment effects for metformin (Figure 11).

FIGURE 11.

National Institute for Health and Care Excellence CG adjustment to reference treatment HbA_1c effect by baseline HbA_1c.

In essence, the adjusted function adds the difference between the two intercepts –1.49 and –0.78, or a constant 0.71, and the β (H₀ –7.5) to the unadjusted H₁. In other words for a given baseline H₀, the H₁ of the adjusted function is a constant difference from the H₁ of the unadjusted function, regardless of the treatment effect relative to metformin ΔTx_rel. As β = –0.50 is negative, the reduction in HbA_1c between t₀ and t₁ is larger for those with a high baseline H₀. The application of the adjusted function means that more patients will see a treatment reduce their HbA_1c to below the NICE treatment intensification threshold of 7.5%. It also prevents patients with a low baseline H₀ being modelled as falling to perhaps unrealistically low values of HbA_1c.

The adjusted function was preferred for the NICE CG because of a superior information criterion and because the influence of β was judged to be significant with its 95% CrI all lying below zero.

For the patient with a baseline H₀ = 9.00, the adjusted model estimates that under canagliflozin his/her H₁ = 9.00 – 0.78 – 0.50 × (9.00 – 7.50) – 0.51 = 6.96 in contrast to the estimate of H₁ = 7.00 of the unadjusted model. Similarly, for the patient with a baseline H₀ = 7. 00, the adjusted model estimates that under canagliflozin their H₁ = 7.00 – 0.78 – 0.50 × (7.00 – 7.50) – 0.51 = 5.96 in contrast with the estimate of H₁ = 5.00 of the unadjusted model.

The NICE CG function is for metformin monotherapy. It was also estimated using a very different data set than the current AG NMA. Any read across from it to the current assessment is consequently almost submerged in caveats. But if a hypothetical placebo in the monotherapy metformin trials would have a reasonably constant relative effect, Tx_rel, at t₁ compared with metformin, and this placebo effect could reasonably be read across to the current patient group, it would be reasonable to explore the impact of the above relationship in the current assessment. For a deterministic analysis this simply requires 0.71 – 0.50 (H₀ – 7.5) to be added to the overall unadjusted H₁ treatment effect estimated for each of the active treatments within the AG NMA. This will be explored as a scenario analysis.

Treatment discontinuations

Those discontinuing in the first year for reasons other than their HbA_1c not falling below the 7.5% threshold are assumed to switch to another monotherapy:

• From:

  • flozins to gliclazide

  • sitagliptin to gliclazide

  • pioglitazone to gliclazide

  • gliclazide to pioglitazone

  • repaglinide to pioglitazone.

Note that those discontinuing are in effect assumed to switch to the alternative monotherapy, and its associated subsequent sequence of treatments. These sequences were retained in part due to data availability and in part due to a desire not to introduce new sequences with a different number of possible intensification steps.

But these subsequent sequences may also contain the treatment that the patient was intolerant of as a monotherapy. This affects only those discontinuing from pioglitazone and those discontinuing from gliclazide. In light of this, a scenario analysis will be undertaken, whereby, among those discontinuing and switching treatment, the intensification step to a treatment, to which the patient was intolerant as a monotherapy, is omitted.

The modelling of the evolution of the risk factors

For HbA_1c the base case applies the treatment effect in the first year of therapy. HbA_1c is then evolved according to the UKPDS68193 equation 11. But this is with the proviso of the UKPDS68193 equation 11 parameter for a patient being in their second year at diagnosis not being applied. Given the average patient duration of 2 years since diagnosis, the AG is of the opinion that including the UKPDS68193 equation 11 parameter for a patient being in their second year since diagnosis would tend to double count the treatment effect of starting a monotherapy. HbA_1c is evolved according to the UKPDS68193 equation 11 until the treatment intensification threshold of 7.5% is breached.

At this point, the patient intensifies treatment and receives the associated treatment effect. HbA_1c is then once more evolved according to the UKPDS68193 equation 11 until the treatment intensification threshold of 7.5% is breached, at which point another treatment intensification occurs. When the patient is on the last line of treatment, HbA_1c evolves according to the UKPDS68193 equation 11 with no further treatment intensifications.

Should a patient discontinue and move on to an alternative treatment at the same line of therapy, the treatment effect of the first line of therapy of the first year is removed, 1 year’s evolution according to the UKPDS68193 equation 11 added, and the treatment effect of the alternative treatment applied.

The paragraphs that follow are purely for illustration. The data are hypothetical and bear no relation to the actual inputs used in the AG modelling.

Figure 12a shows how this results in a sawtooth evolution of HbA_1c. It applies to a patient aged 40 years with a current baseline HbA_1c of 7.6%, who at diagnosis was aged 30 years and had a HbA_1c of 7.0%. It assumes four strategies, with initial reductions in HbA_1c of 1.8%, 1.6%, 1.4% and 1.2% for strategies 1–4, respectively. It also assumes that two further treatment intensifications are possible, these having the reductions in HbA_1c of 2.0% and 1.5% across the four strategies.

FIGURE 12.

Example of the modelled evolution of HbA_1c: UKPDS68. 193

The modelled evolution of strategies 2–4 are very similar, with the first treatment intensification at year 6. The slightly greater initial reduction in HbA_1c of strategy 1 is sufficient for HbA_1c not to breach the treatment intensification threshold of 7.5% until 1 year later, and, as a consequence, the first treatment intensification does not occur until year 7.

Figure 12b illustrates how a discontinuation could affect the modelled evolution of HbA_1c for strategy 1. This still assumes a reduction in HbA_1c from the original treatment of 1.8%, but from the alternative treatment that the patient discontinues, a reduction of only 0.5%. As before, this also assumes that two further treatment intensifications are possible, with reductions in HbA_1c of –2.0% and –1.5% in the sequence to which the patient has discontinued.

In light of the Janssen submission, the model has been constructed to permit a scenario analysis of HbA_1c having a linear increase and for the annual rate of increase to be treatment specific. The following illustrates the same initial treatment effects for strategies 1–4, but for the annual linear increase while on first-line treatment to be 0.1%, 0.2%, 0.3% and 0.4%, respectively. For those discontinuing from strategy 1, the annual linear increase while on the first-line treatment is assumed to be 0.05%. Subsequent to treatment intensification, the evolution of HbA_1c is assumed to revert to the UKPDS68193 equation 11, but the model has the facility to impose treatment-specific annual linear increases in HbA_1c at any or all treatment lines (Figure 13).

FIGURE 13.

Example of the modelled evolution of HbA_1c: linear evolution.

A similar approach is taken for the modelling of the evolution of SBP and can be for the TC/HDL ratio, though the base case holds the TC/HDL ratio constant at 3.0. The UKPDS68193 specifies equations 12 and 13, respectively. Treatment intensifications are still determined by the modelled HbA_1c.

Figure 14 illustrates the modelled evolution of SBP for the same patient as before, with HbA_1c being modelled to evolve according to the UKPDS68193 equation 11, with a SBP at baseline of 130 mmHg and at diagnosis of 120 mmHg. It assumes initial treatment effects of –20 mmHg, –15 mmHg, –10 mmHg and –5 mmHg for strategies 1–4, respectively. The intensifications result in treatment effects of –15 mmHg and –10 mmHg. For the patient modelled as discontinuing during strategy 1, the alternative treatment effect is –5 mmHg.

FIGURE 14.

Example of the modelled evolution of SBP.

Figure 15 illustrates the UKPDS modelled evolution of the TC/HDL ratio for the same patient as before, with HbA_1c being modelled to evolve according to the UKPDS68193 equation 11, with a TC/HDL ratio at baseline of 4.0 and a ratio at diagnosis of 3.5. It assumes initial treatment effects of –1.0, –0.8, –0.6 and –0.4 for strategies 1–4 respectively. The intensifications result in treatment effects of –0.8 and –0.6. For the patient modelled as discontinuing during strategy 1, the alternative treatment effect is –0.4.

FIGURE 15.

Example of the modelled evolution of the TC/HDL ratio.

The modelled evolution of HBA_1c, SBP and the TC/HDL ratio for a range of inputted patient characteristics was cross-checked with that modelled by the UKPDS OM1 at central UKPDS68193 parameter values. The modelled evolutions values were typically around 99.99% of the values simulated by the UKPDS OM1 model.

The evolution of the patient BMI is based upon a mean annual increase in weight of 0.1 kg, as has typically been used in previous NICE assessments of treatments for diabetes, and is apparently originally sourced from the 2006 NICE CG on obesity. 211 Over the course of NICE assessments and guidelines development for treatments for type 2 diabetes, quite a lot of discussion has focused upon what is reasonable to assume about the duration of weight effects. There has been some argument that initial weight losses associated with treatment may tend to be transient, whereas initial weight gains associated with treatment may tend to be more permanent. In light of this, five scenarios are modelled:

• treatment weight changes maintained, with no rebound to natural history

• treatment weight gains maintained, weight losses rebound to natural history after 1 year

• treatment weight gains maintained, weight losses rebound to natural history at intensification

• treatment weight changes rebound to natural history after 1 year

• treatment weight changes rebound to natural history at intensification.

Figures 16 and 17 illustrate weight changes being maintained for a patient of 85 kg at baseline. Initial hypothetical treatment effects are 5 kg, –5 kg, 10 kg and –4 kg for strategies 1–4, respectively. The intensifications result in treatment effects of 3 kg and 7 kg. For the patient modelled as discontinuing during strategy 1, the alternative treatment effect is 4 kg, with the additional weight gains in the alternative sequence thereafter being assumed to be the same as in the original sequence.

FIGURE 16.

Example of the modelled evolution of patient weight: no rebound.

The above is largely self-explanatory. Within the evolution of weight for the patient under strategy 1 who discontinues, there is a drop in weight between year 1 and year 2. This arises as a result of the patient being assumed to come off his/her original treatment, which would have increased his/her weight by 5 kg and to move on to the alternative treatment which increases his/her weight by only 4 kg. Quite when the patient would discontinue, and, as a consequence, quite what the balance would be, in practice, between the 5-kg increase and the 4-kg increase during the first line of treatment, is a moot point.

FIGURE 17.

Example of the modelled evolution of patient weight: rebound scenarios. Tx, treatment.

Body mass index scenarios 2 and 3, which have weight losses rebounding to natural history, do not affect strategies 1 and 2, as they incur only weight gains.

Body mass index scenario 2 sees weight losses rebound to natural history after 1 year. So at year 2 both strategies 2 and 4 have rebounded to natural history and 85.2-kg weight. Thereafter, they follow the same weight profile, as their treatment intensifications also occur at the same time: years 6 and 11.

Body mass index scenario 3 sees weight losses rebound to natural history at treatment intensification. To model this, the rebound to natural history that would have applied in the previous year is first calculated – 85.5 kg – to which the intensification treatment effect of 3.0 kg is added to give a weight of 88.5 kg at year 6. An alternative way of looking at this is to view the intensification treatment effect of 3.0 kg being composed of two parts: 0.1 kg for natural history and 2.9 kg for additional treatment effect.

Body mass index scenario 4 sees all weight changes rebound to natural history after 1 year. The figure is largely self-explanatory, with all initial treatment effects being removed at year 2. Thereafter, strategies 2–4 intensify treatment at years 6 and 11, and see the weight gains associated with intensifications rebound to natural history at years 7 and 12. Strategy 1 intensifies later at years 7 and 12, hence the later rebounds to natural history.

Body mass index scenario 5 sees all weight changes rebound to natural history at treatment change. Strategies 2–4 intensify treatment at years 6 and 11, so rebound to natural history the following year. But the weight gains associated with the intensifications are then added to where the patients rebound. As a result, after the first intensification at year 6 the patient weight is 88.5 kg for strategies 2–4. For strategy 1 the intensification is 1 year later, with it joining the other strategies at a weight of 88.6 kg in year 7. Much the same happens at the subsequent intensification.

Note that the BMI scenarios 4 and 5 may be felt to be literally unrealistic, with weight gains rebounding to natural history after 1 year or at the next treatment change. But they may be better thought of as causing weight to converge between strategies either after 1 year or at treatment change. In terms of the modelling of the impact of BMI upon QoL this will not be exactly arithmetically correct because of the floor of 25 kg/m² on the BMI QoL coefficient of –0.0061. Given the baseline BMI of 31.6 kg/m² (SD 6.0 kg/m²), around 13% of patients are modelled as having a BMI of less than 25 kg/m². For some of these patients, the rebounds of scenarios 4 and 5 may not be exactly equivalent in their QoL impacts to weight converging between the strategies by some other means, for example at the value of strategy with the higher BMI. But the AG is of the opinion that any differences are likely to be minor.

Note that there is a minor error within the implementation of weight in the AG modelling. The UKPDS68193 requires that the BMI at diagnosis is used and this has been implemented correctly. But the AG modelling assumes that the BMI at diagnosis also applies at baseline. It can be argued that the BMI at baseline should be the BMI at diagnosis plus the natural history increase that would be implied by the duration at baseline. But as the mean duration of diabetes at baseline is estimated to be 2.0 years, and that this adjustment would affect results for only the subset of patients during the cycles they bordered, the 25 kg/m² threshold, this error will have negligible effects upon net results.

Diabetic ketoacidosis

There has been some suggestion that the flozins may increase the risk of DKA. But rates are low in absolute terms, with the EMA reporting 101 cases over about 500,000 patient-years of flozins use.

The Diabetics With Eating Disorders surveyed England’s primary care trusts (PCTs) in 2010. 212 Among the 45 PCTs that responded, the mean cost per DKA event was £1438, or £1552 in 2014 prices. But given an event rate of 1 per 5000 patient-years, the average increase in costs associated with this is minimal: around 30p per year of treatment. The typical duration of DKA events is also quite short, certainly less than 1 week, though there may be recurrence. But even with quite a large QoL decrement, given the absolute event rate and the short duration any overall average QALY impact will also be minimal.

There remains the possibility of an increased mortality with DKA among those with type 2 diabetes. This could have more of an impact upon the modelled average QALY in the flozin arms. But there is no simple means of incorporating this mortality into the OM1 modelling, this being a black box to the AG.

For the above reasons, DKA has not been incorporated into the economic modelling.

Quality of life: diabetes and the complications of diabetes

Given the use of the OM1 model, the AG draws the QoL value for those without any complications and QoL impacts for the complications of diabetes from the UKPDS62198 (Table 45). A value for renal failure is not given in the UKPDS62198 and, as a consequence, the AG has used the OM1 default value of –0.263 as drawn from Kiberd and Jindal. 213

Quality of life: weight

In common with most NICE assessments of treatment for type 2 diabetes and the draft NICE type 2 diabetes CG, the AG applies the utility decrement of –0.0061 (SE 0.001) of Bagust and Beale. 199 Within Bagust and Beale,199 this decrement applies if the patient BMI is above 25 kg/m².

The mean BMI within the UKPDS RCT was 27.7 kg/m², and it is from here that the mean baseline utility of 0.785 is drawn. As a consequence, as the modelling applies the –0.0061 QoL decrement when the patient BMI rises above 25 kg/m², it can be argued that the baseline utility of 0.785 should have 0.0061 × 2.7 = 0.0165 added to it. This modification is adopted for the AG base case.

Quality of life: treatment discontinuations

Treatment discontinuations were assigned a QALY decrement associated with nausea as drawn from Matza et al. 214 The with- and without-nausea QoL values of 0.85 and 0.89 were taken to apply yielding a mean decrement of 0.04, which the GDG thought a 6-week duration would be most reasonable estimate, this yielding a mean QALY decrement of –0.00462.

Quality of life: adverse events

The 2012 NICE CG on infection, CG139,215 undertook a systematic review of the literature for studies of the QoL impacts of symptomatic UTIs. This identified 11 studies, but a number of these are of limited relevance to the current assessment because, for example, being among patients with spinal cord injuries. Of the 11 studies, five appear most relevant to the current assessment.

Clinical guideline 139215 also undertook economic modelling of treatments for UTIs. But as a result of the available clinical effectiveness estimates being largely limited to those with spinal cord injury, the QoL values applied are of limited relevance to the current assessment. This modelling also considered progression from symptomatic UTIs through to first-line drug resistance and multidrug resistance, with these causing increased costs and mortality.

The AG modelling for the current assessment only considers the QoL and cost impacts of treating UTIs and GTIs, with the assumption that none will progress to a more serious condition. So there are caveats around these estimates, and for a given set of inputs they could be seen as being biased and on the low side.

Barry et al. 205 used the Index of Wellbeing (IWB) to estimate QoL among young women with UTIs. The IWB is a generic QoL instrument. The mapping function from the IWB to QoL values was apparently based upon 62 American nurses and non-medical graduate students ranking health states on a 16-point scale. Barry et al. 205 report that the IWB includes hospitalisation, self-care and ambulatory status, and permits the inclusion of the following symptoms: pain; bleeding; itching discharge from sexual organs; painful burning or frequent urination; burning or itching rash on large areas of the body; taking medication; fever or chills with aching all over; and pain in the chest, stomach, sides, back or hips. But Barry et al. 205 do not describe quite how the QoL values for the health states for their model have been derived. It appears to be based upon an index patient with a set of symptoms; that is expert opinion linked to the IWB. They estimated disutilities of 0.3732 for pyelonephritis, and 0.2894 for vaginitis and persistent dysuria. Their duration was estimated to be 10 days, 5 days and 5 days, respectively.

Gold et al. 216 catalogued 130 health states using the Health and Activity Limitation Index (HALex), with the score being based upon the answers to two questions of the US National Health Interview Survey. A multiattribute utility model resulted in QoL estimates of 1.00 for perfect health, and 0.73 for bladder infection and 0.66 for kidney infection. The derivation of these weights is not particularly clear within the paper.

Ackerman et al. 217 used the standard gamble to estimate QoL values among 13 men with moderate to severe benign prostatic hyperplasia. A variety of health states were described, with the QoL impacts of severe UTIs being estimated among these. The six risk-averse men reported an average value of 0.972 for a severe UTI, whereas the seven non-risk-averse men reported an average value of 0.893. Over the 13 respondents this suggests an average disutility per severe UTI of 0.071.

Ellis and Verma218 measured the impact of UTIs among 118 otherwise healthy Canadian women through a case-controlled analysis using the Short Form questionnaire-36 items (SF-36) with a recall period of 1 day. The mapping from the SF-36 to the EQ-5D QoL values based upon the algorithm of Ara and Brazier219 appears to have been undertaken by the NICE CG, as Ellis and Verma218 report only the mean values for the eight main elements of the SF-36. This resulted in those with no UTI having a mean QoL of 0.922 compared with 0.724 for those with a UTI.

Ernst et al. 220 used the Quality of Wellbeing (QWB) to estimate the QoL among 146 American women who were diagnosed with acute cystitis, and the effect of treatment upon QoL. Those with type 2 diabetes were excluded from the study. The QWB was administered 3, 7, 14 and 28 days after the initial visit. The QoL at baseline was 0.68 (SD 0.03) compared with 0.81 (SD 0.11) at the 28-day point. QoL among those cured compared with those not cured was statistically significantly different at the 5% level at days 3, 7 and 14, with respective QWB scores of 0.77 vs. 0.72, 0.82 vs. 0.71 and 0.83 vs. 0.76 (Table 46).

Of the above papers, Ackerman et al. 217 could be argued as coming closest to the NICE reference case. But the usefulness of these estimates is compromised by the small sample size. As a consequence, the AG will use the results of the Janssen TTO study for the base case of QoL impact of –0.19 for a UTI and –0.25 for a GTI. Nicolle et al. 98 estimated median durations of UTIs of between 11.0 days and 12.5 days, and, as a consequence, the base case will assume 2 weeks’ average duration.

Quality of life: hypoglycaemia

Following the lead of the draft NICE CG for type 2 diabetes, the source for the base case for the QoL decrements associated with hypoglycaemia events will be Currie et al. 204 This used two separate 3-month recall surveys among patients with diabetes (n = 408 and n = 897), undertaken at different time points, though 145 patients responded to both surveys.

The first survey was used to estimate a relationship between a patient’s score on the Hypoglycaemic Fear Survey (HFS) and the number of non-severe and severe hypoglycaemic episodes with coefficients of 1.773 (SE 0.230) and 5.881 (SE 1.553), respectively. The second survey was used to estimate the relationship between the HFS and the EQ-5D QoL with a coefficient of –0.008 (SE 0.001).

Given the 3-month recall period, the mapping between non-severe hypoglycaemia event rates and the patient’s score on the HFS requires that rates be converted to 3-monthly rates before the 1.773 HFS coefficient can be applied to arrive at the correct QALY decrement.

The authors of the draft NICE CG for type 2 diabetes also point out that the table 4 coefficient of Currie et al. 204 for non-severe hypoglycaemia events is based upon the natural logarithm of the event rate rather than the event rate. As such it is non-linear. To account for this the AG has followed the method of the authors of the draft NICE CG for type 2 diabetes and applied a Poisson distribution to give the spread of possible patient event rates prior to applying the coefficients of Currie et al. 204

The 5.881 HFS coefficient of table 4 of Currie et al. 204 for severe hypoglycaemic events was derived on a dichotomous basis, equal to 1 if there were any events reported during the previous 3 months and equal to 0 if there were none reported. The draft NICE CG gives a quite complicated formula for accounting for this using a binomial distribution, but this apparently simplifies to the quarterly probability times the utility decrement (Gabriel Rogers, NICE, 17 August 2015, personal communication).

Table 47 presents a range of estimates based upon this method.

But the values of Currie et al. 204 come with some major caveats. As Currie et al. 204 note regarding the two data sources, ‘These studies were commissioned by the pharmaceutical industry to inform drug developments around new treatments for diabetes that were found to reduce the frequency of hypoglycaemia’. The paper authorship also includes staff of Novo Nordisk and Sanofi-Aventis.

The values are based on results from two surveys, with a response rate of 31%. The hypoglycaemic episodes were recent events and perhaps therefore fresh in the memory. Forty-five per cent of respondents were on insulin. Respondents might have been more likely to have been concerned about hypoglycaemia than non-respondents.

Around one-third of respondents had type 1 diabetes mellitus (T1DM) with around two-thirds of respondents having type 2 diabetes. Quite what covariates were considered and quite how the paper arrived at the final regressions is not entirely explicit. Patient data from the first survey were removed if the patient also responded to the second survey, reducing the sample to 57% of the original, though the reasons for this and impacts of doing so are not clear. Similarly, the grouping of complications was also possibly subjective.

The 5.881 coefficient for severe hypoglycaemia episodes was also based on whether patients had had any severe hypoglycaemia events during the recall period. If, within this group, the mean number of severe hypoglycaemic episodes was more than one, it seems likely that the coefficient somewhat overestimates the impact of having one severe hypoglycaemia events within a quarter.

The patient number and demographics reported by Currie et al. 204 for the first survey are based upon the full 408 patients of this survey. But for the analysis, 175 of these patients were excluded because of also being in the second survey. As a consequence the demographics and events rates that were used when analysing the data subset of the first survey cannot be determined.

For the full 408 patients of the first survey, only 2.3% (n = 9) reported experiencing at least one severe hypoglycaemic event during the previous 3 months. This was somewhat less than the 8.6% (n = 77) proportion who reported experiencing at least one severe hypoglycaemic event during the previous 3 months in the second survey.

For severe hypoglycaemic event rates, Currie et al. 204 state that within the surveys ‘very few people > 1 event’ and they report a mean rate of ‘1.47 events per patient-year’. It seems likely that this mean rate was the average across the two surveys. It would have been useful to have known the mean rate for each survey, and for the small subset of the first survey that was actually analysed.

The relationship between having experienced at least one severe hypoglycaemic event in the last 3 months and the HFS index, that is the 5.881 coefficient, consequently appears to have been based upon, at most, nine patients reporting. The restriction of the subset analysed to 57% of the total sample of the first survey suggests that this number is likely to have been somewhat less than nine patients. This gives rise to the possibility of an outlier patient within this small subset having an unreasonable impact upon results. The construction of the subset was at investigator discretion.

The AG cannot further interrogate the data underlying the estimates of Currie et al.,204 and it is possible that they may be overestimates. Note that in common with previous analyses, the method of the table above in effect assumes that a patient experiences at most one severe hypoglycaemic event per quarter. There may be an argument for dividing the QALY decrement associated with severe hypoglycaemic events by 1.47, the mean event rate reported in Currie et al. 204

Costs: direct drug costs

Treatment costs are based upon the NHS drug tariff, and upon list prices for which there are no entries in the NHS drug tariff. Daily doses are assumed to be 60 mg for gliclazide MR, 45 mg for pioglitazone, 6.0 mg for repaglinide, 100 mg for sitagliptin, 10 mg for dapagliflozin, 25 mg for empagliflozin and 300 mg for canagliflozin. Insulins costs are based upon a requirement of 0.3 IU/kg when starting NPH, with this rising to 0.55 IU/kg when adding bolus which itself is required at 0.2 IU/kg.

The AG expert opinion also suggests that those receiving pioglitazone should have their BNP measured, perhaps initially 6-monthly but annually thereafter. A marginal cost of £21 has been taken from Craig et al. 45 and inflated to 2014 prices using a 1.25 multiplier from the PSSRU Unit Costs of Health and Social Care index. This has also been assumed to require a dedicated GP appointment, costed at £46 using the PSSRU Unit Costs of Health and Social Care 2014. 208

This results in the following treatment costs for the oral therapies (Table 48).

The AG modelled sequences differ from those of the company submissions in that patients add NPH insulin rather than switch to it and as a consequence the cost differences between the sequences are maintained over the horizon of the modelling. In light of this, a scenario analysis is undertaken, which withdraws the initial monotherapies when patients switch to NPH insulin. Note that this affects only the direct drug costs and not the clinical effectiveness estimates.

Costs: treatment intensifications and switches

Treatment intensifications due to breaching the 7.5% HbA_1c threshold and treatment switches due to intolerance are assumed to involve one 12-minute GP appointment. This is costed using the PSSRU Unit Costs of Health and Social Care 2014 at £46. 208

Costs: adverse events

The AG treatment assumptions are broadly in line with those of the Janssen submission (Table 49), with this resource use being confirmed by AG expert opinion. Medication for UTIs is assumed to be 7 days of trimethoprim 200 mg twice daily, for male GTIs fluconazole 200 mg and female GTIs three 200-mg clotrizamole pessaries (see Table 49).

These costs are largely based upon assumption and have consequently been treated deterministically within the probabilistic modelling.

Costs: hypoglycaemic events

The AG have followed the draft NICE CG when costing severe hypoglycaemic events.

Hammer et al. ,207 in an industry-sponsored study, surveyed 147 UK patients with type 2 diabetes using insulin, with 19 reporting at least one severe hypoglycaemic episode in the previous year, with 10 of these being treated by the NHS. Hammer et al. 207 acknowledge the non-random selection of their patient sample, but provide few details about it other than to note that it was predominantly through health-care professionals. Patients were surveyed using a structured questionnaire about the resource use associated with their events.

Patients were divided into three groups: those who had their severe hypoglycaemic event treated by family members; by medical practitioners in the community; and, in hospital. The mean direct costs by type, in 2007 prices, were £33 for those treated by family members, due to NHS follow-up costs; £231 for those treated by the NHS in the community; and £862 for those treated in hospital. Owing to the non-random sample selection, there is no definitive means to translate these into a weighted average cost. But the GDG of the draft NICE CG was reportedly happy to use the sample proportion treated by family members (9/19), coupled with an assumption that, of the remainder, 65% would be treated in hospital.

This results in a mean cost per severe hypoglycaemic event of £353 in 2007 prices, which, when uplifted by a 1.16 multiplier from the PSSRU Unit Costs of Health and Social Care 2014,208 results in an estimate of £411.

Assessment Group sensitivity analyses

All scenario analyses have been run deterministically with a cohort of 50,000 patients and 1000 inner loops to reduce Monte Carlo error. The sensitivity analyses around the –0.0061 QoL decrement per BMI point above 25 kg/m² and the rebound of treatments’ effect upon weight are presented for all analyses: 

• BMI 1 Natural history progression with no rebound

• BMI 2 Natural history progression with weight losses rebounding after 1 year

• BMI 3 Natural history progression with weight losses rebounding at treatment change

• BMI 4 Natural history progression with weight rebounding after 1 year

• BMI 5 Natural history progression with weight rebounding at treatment change.

The AG has also undertaken the following sensitivity analyses.

• SA01 At the third intensification patients switch to insulin plus gliclazide, and cease their other treatments.

• SA02 Applying the UTI and GTI rates to all cycles of the model.

• SA03 Assuming that all patients when starting monotherapy have a HbA_1c of 7.5%.

• SA04 Adjusting the HbA_1c treatment effect for patients’ baseline HbA_1c values as in the NICE CG.

• SA05 Not applying the discontinuation rates.

• SA06 Applying the NICE CG baseline TC/HDL values and the UKPDS68193 TC/HDL progression.

• SA07 Applying the UKPDS68193 year 2 parameter for the evolution of HbA_1c.

• SA08 Intensifying when adding gliclazide having a –0.47% HbA_1c effect.

• SA09 Applying the Janssen linear evolutions of HbA_1c for all treatments.

• SA10 Assuming that those discontinuing from a treatment omit any subsequent intensification step that reapplies this treatment.

• SA11 SA01 and SA08 combined.

Assessment Group base-case results

The disaggregate costs of the base case provided in Table 51.