In more mad science, we turn our attention to the sodium/glucose cotransporter 2 (SGLT2) inhibitors. Someone in a pharma lab clearly said “Hey, you know how diabetics have glucosuria when they’re hyperglycemic? Could we use that?” And so now, thanks to the miracles of modern science, you can pee like you’ve got a hemoglobin A1c (HbA1c) of 12% all the time.
Blocking a major receptor responsible for re-uptake of glucose from the proximal renal tubule, canagliflozin and its sister drugs were for a long time the red-headed stepchild of diabetes therapy. With a weird side effect profile and an ever-growing roster of competing medications, the argument for using SGLT2 inhibitors was tough to make. After all, a drug with a 10% rate of genital yeast infection is a hard sell next to, say sitagliptin, a drug you barely know you’re taking. But thanks to a series of trials looking into the cardiovascular effects of these medications, the class is getting a new look. The CANVAS study and it’s sister CANVAS-R trial, product of Bruce Neal and colleagues, make up the latest demonstration of the fact that glucosuria may be a good thing after all. I mean, under the right circumstances.
If You’re Only Going to Read One Paragraph
In this randomized double-blind placebo-controlled trial, an aggregate of patients on doses of 100mg and 300mg canagliflozin had lower rates of cardiovascular death, nonfatal MI, or nonfatal stroke than patients on placebo (HR 0.86; 95% CI 0.75 to 0.97; P=0.02 for superiority). There was also reduction in progression of albuminuria and reduced rates of GFR decline, though due to the sequence of hypothesis testing these results are not statistically significant (taken individually, they would meet this threshold). These findings are consistent with the EMPA-REG trial of empagliflozin versus placebo, where a hazard ratio of 0.86 was observed (driven mostly by a significant reduction in cardiovascular death). There is a troubling signal in CANVAS with regard to increased rates of amputation compared to placebo (6.3 vs. 3.4 participants per 1000 patient-years; HR 1.97), but adverse effects are otherwise consistent with previous reports, which is to say frequent but generally non-catastrophic. This is an indication that the cardiovascular benefit seen previously from empagliflozin is likely consistent across the SGLT2 class, and should make us more likely to consider these drugs in treating diabetic patients at elevated risk for adverse cardiovascular events.
A Bit of Context
So to talk about trials of cardiovascular outcomes in new antidiabetic medications, we have to talk about rosiglitazone. Rosiglitazone, a thiazolidinedione (no, I can’t pronounce that) similar to the still-in-use pioglitazone, was a promising anti-hyperglycemic drug. Approved in 1997, the drug’s sales spiked at just over 2.5 billion dollars in 2006. And then in 2007 a metanalysis found an association of rosiglitazone use with increased rates of MI (HR 1.43, P=0.03) and a nearly-significant trend towards increased death from cardiovascular causes (HR 1.64, 95% CI, 0.98 to 2.74; P=0.06). This led to about 13,000 lawsuits, which sounds like a number that I’m making up but is actually a real thing that happened.
Now, this is far from the whole story — unlike, say, Vioxx, the drug was not pulled from the market, and was in fact partially exonerated when a 2009 RCT failed to show any association with increased MI risk. Still, this was enough of a scare that the FDA in December 2008 released new guidelines for the development of antidiabetic therapies mandating that they receive meaningful evaluation of cardiovascular outcomes. Hence the new rash of cardiovascular outcomes trials for the SGLT2 inhibitors and GLP-1 agonists. But these trials, designed to prove noninferiority (i.e. that the drugs in question do not increase rates of MI over comparable therapy), took a bit of a turn starting with EMPA-REG OUTCOME.
It was just another day at the office for empagliflozin, an SGLT2 inhibitor like canagliflozin. And then, over a mean follow up of 3.1 years, a composite primary outcome of cardiovascular death, nonfatal MI, and nonfatal stroke happened at rates not just non-inferior to but indeed significantly superior to usual care plus placebo (HR 0.86; P=0.04 for superiority). There was no difference in all-cause mortality, and interestingly no difference in nonfatal MI or stroke — the observed difference was driven almost entirely by a reduction in cardiovascular death (ARR 2.2%, RRR 38%).
And in unison, pharmaceutical executives everywhere got more excited about their mandated cardiovascular safety trials. Rather than simply hoping for noninferiority, they could now hope for FDA approval for reduced rates of cardiovascular death. The GLP-1 agonist exenatide also hit the mark in the LEADER trial, though it’s cousin lixesenatide only made it to noninferior in the ELIXA trial.
This raised a (non-)dilemma for subsequent investigations. What’s the difference between a non-inferiority study and a study looking for superiority? Statistical design and the resulting power calculation. So while it may have been designed for non-inferiority, the fact that CANVAS combined patients from the dedicated cardiovascular outcomes trial with patients enrolled in a renal outcomes trial (CANVAS-R) into an amalgamated study in order “to maximize statistical power to detect plausible effects of canagliflozin on cardiovascular, kidney, and safety outcomes as suggested from evolving evidence about SGLT2 inhibitors,”? Yeah, if you’ve got the chance to prove that you generate better cardiovascular outcomes, you’ve kinda gotta take it.
Patients, Intervention, Comparator and Outcomes
CANVAS and CANVAS-R are two separately designed trials subsequently combined and reported as one — from here out I’m only going to report them as one trial, with the understanding that patients were initially recruited into two separate protocols (though the inclusion criteria were the same for both) and later underwent combined analysis. 4430 patients were enrolled through CANVAS and 5813 through CANVAS-R.
667 centers in 30 countries recruited patients at least 30 years of age with type 2 diabetes and A1c levels between 7.0% and 10.5%. Patients were required to have a history of symptomatic ASCVD or to be older than 50 with two of the following risk factors: duration of diabetes of at least 10 years, systolic blood pressure higher than 140 mm Hg on antihypertensives, current smoking, microalbuminuria or macroalbuminuria, or high-density lipoprotein (HDL) less than 38.7 mg/dL. Patients also had to take at least 80% of pills during a two week placebo run-in, performed to ensure reasonable medication compliance. Prospective participants were excluded for a history of DKA, an “unstable” medication regimen (i.e. changed in the last 8 weeks), severe hypoglycemic events in the past 6 months, or major adverse cardiac events (MACE) within the last 3 months. Patients with CKD were eligible as long as their enrollment GFR was > 30 mL/min. As always, dive into the supplementary appendix if you want to view the full list.
After the above run-in period, patients in CANVAS were randomized in a 1:1:1 ratio to canagliflozin 300mg daily, 100mg daily, or placebo. In CANVAS-R, patients were randomized 1:1 to 100mg vs placebo, with the option to increase to 300mg subsequently (71.4% of those patients did end up increasing the dose, so a little over 60% of intervention patients ended up on this higher dose).
The primary outcome is a pretty vanilla version of MACE; cardiovascular death, nonfatal MI, or nonfatal stroke. All cause mortality was a secondary outcome, as was progression of albuminuria (at least 30% increase). “Exploratory” outcomes included hospitalization for heart failure and a 40% decrement in estimated GFR.
OK, we’ve got “exploratory outcomes” and now we have to talk about stats for a second. The issue with multiple comparisons in a single trial is that the probability of alpha errors (observing a difference that isn’t actually present) isn’t additive, it’s multiplicative. The analogy that’s most accessible here is sequential coin flips. For the first flip you can say with confidence that the probability of it coming up heads is 50%. You can say that for the second flip, too — but if you’re looking at the probability that both flips are going to come up heads, you’re down to 25%. Similarly, when looking at multiple outcomes, the probability of a type 1 error accumulates. For example, if P = 0.05 on two separate observations, your probability of error in at least one of them is not 5% (1 – 0.95), but rather (1 – 0.95^2), or 9.75%. The odds that canagliflozin decreases rates of MACE is one thing, but the odds that it does that and decreases the rate of GFR decrement is analagous to two subsequent coin flips. So the authors prespecified the order in which they would do their analysis, and acknowledged that outcomes further down the chain could not reasonably be said to be free of type 1 error. If you’d like someone to explain this to you a little bit more eloquently, try Wikipedia. And if that makes you feel dirty, try the sources to which they refer. But this is important, because it’s the reason that a few observations that meet significance per se are not considered statistically significant on the basis of this trial.
10,142 patients were enrolled, with 9734 included in the final analysis. Median follow up was 126.1 weeks, with mean just over 3 years. Mean age was 63.3. Participants were 64.2% male and had diabetes for a mean of 13.5 years. 65.6% of patients had cardiovascular disease at baseline. 3.3% of patients were black, which doesn’t sound like much until you look at the ~1% we’ve been dealing with in other large cardiovascular outcome trials. Still a 78.3% white population. Otherwise, the patients were sick and multimorbid, fairly representative of a complex primary care population: 17.5% had diabetic nephropathy, 17.8% were currently smoking, and systolic BP had a mean of 136.6. Participants used a wide variety of antidiabetic medications, and 50.2% of included patients were on insulin. 74.9% of patients were taking a statin.
Rates of the primary outcome were significantly lower in the canagliflozin group compared to placebo, with 26.9 vs. 31.5 participants per 1000 patient-years (HR 0.86; 95% CI 0.75 to 0.97; P=0.02 for superiority). Assuming treatment for the mean follow up of 3.1 years, this gives us a number needed to treat of 72. Results were statistically homogeneous in prespecified subgroup analysis, except for patients using diuretics at baseline (who did not see any benefit).
Because the first secondary outcome in the hypothesis testing sequence (all-cause mortality) was not significant (P = 0.24), investigators were unable to meaningfully evaluate further secondary outcomes for significance — you can’t meaningfully evaluate the possibility that three coin flips in a row will come up heads if the second one lands tails. However, for your viewing pleasure, confidence intervals are provided: all cause mortality HR 0.87 (0.74 to 1.01), death from cardiovascular cause HR 0.87 (0.72 to 1.06), progression of albuminuria HR 0.73 (0.67 to 0.79). A composite renal outcome of 40% GFR decrement, need for renal-replacement therapy, or death from renal causes had a hazard ratio of 0.60 (0.47 to 0.77). All of these hazard ratios were homogenous across subgroups.
In important but non-headline news, canagliflozin reduced HbA1c by 0.58% compared with placebo, and decreased systolic BP by 3.93 mmHg. It also generated a mean 1.6 kg weight loss, and increased HDL by 2.05 mg/dL but kept the LDL:HDL ratio the same by also bumping LDL by 4.68 mg/dL. All of these are consistent with known drug effects and have previously been shown to be significantly different from placebo.
With regard to safety and tolerability, rates of drug discontinuation were similar (29.2% of canagliflozin patients vs 29.9% of placebo). Serious events trended towards slightly more common in the placebo group, HR 0.93 (0.87 to 1.00). There was no observed difference in rates of diabetic ketoacidosis. Male genitourinary infections occurred in the intervention arm at a rate of 34.9 per 1000 patient years (P-Y) compared to 10.8 per 1000 P-Y with placebo, and mycotic genital infection in women had a whopping 68.8 cases/1000 P-Y compared to 17.5. The number needed to harm is 14 for male GU infection and 6.5 for female mycotic infection. Fractures were more common as well (a known risk of SGLT2s, possibly due to drug-related bone density loss), as was volume depletion, but the absolute rates of these events remained low, as did the absolute increase in risk (NNH 95 for fracture, 44 for volume depletion).
And now the real sticking point: amputations occurred more commonly in the canagliflozin group than placebo, 6.3 vs 3.4 events per 1000 P-Y, NNH 115. This is weird, and somewhat concerning given that it has not been described before. Unsurprisingly the risk was highest in patients with known vascular disease, but the relative risk was consistent across subgroups. There is no known mechanism by which canagliflozin might cause the need for amputation, but with an irreversible complication like amputation, not having a reasonable mechanism on which to blame this makes me feel worse and not better.
We are seeing some signals here that were not observed in EMPA-REG. For one thing, fracture risk in EMPA-REG was similar across groups. And while amputation was not a prespecified adverse event in that trial, there was no mention of increased frequency in the published data. You could posit that canagliflozin has a non-class effect causing these differences, and I couldn’t say much to refute you. But another possibility is that the slight differences in trial size (9734 in the final analysis for CANVAS vs 7020 in EMPA-REG OUTCOME) may have introduced the power to demonstrate significant differences in rarer events.
What we don’t see is a reduction in cardiovascular or all-cause mortality, both of which we saw in EMPA-REG. What’s up with that? Well, for one thing, there’s a difference in the study population. EMPA-REG recruited only patients with established cardiovascular disease, whereas CANVAS included patients at elevated risk but without known disease. If the baseline event rate was lower, CANVAS could have lost some power to detect a difference in mortality.
These numbers taken in conjunction tell the story of a drug with a real benefit, a reduction in MACE and renal protective effects that are insignificant only because of sequential hypothesis testing rather than its observed size. But also one with adverse effects that are significantly more common than the events you’re seeking to prevent, and troubling signals that it and its sister drug may or may not be truly equivalent. At the end of the day, if you have the choice of both drugs (i.e. you aren’t boxed into picking one based on the patient’s insurance), it’s a simple decision. Empagliflozin has the same benefit as canagliflozin without the concerning signals towards irreversible adverse effects. Sure, there’s no reasonable mechanism we can propose for how it might cause these things, but why take the chance?
And as for how SGLT2s are fitting into my practice — well, metformin should be the first line medicine for diabetes. It has its own all-cause mortality benefit, is cheap, and is generally well-tolerated. The evidence is good and the guidelines are clear. But after that? Well, historically that’s varied a lot from one doctor to the next. A plurality of my colleagues tend to reach for sitagliptin — like metformin it has few major adverse effects (especially the hypoglycemic effects of sulfonylureas), and it’s weight-neutral. But it only produces about a 0.5% reduction in HbA1c, similar to SGLT2s. And you know what it doesn’t have? A cardiovascular or all-cause mortality benefit. It has had the same FDA mandated studies, and only reached non-inferiority. So for me, the SGLT2 inhibitors now beat out the DPP4 inhibitors in my generic diabetes care algorithm.
Choosing between SGLT2 inhibitors and GLP-1 agonists is harder. Liraglutide is safe, has a much more impressive weight benefit that SGLT2 inhibitors, and has a similar degree of demonstrated cardiovascular benefit (HR 0.87; P=0.01 for superiority) including a hazard ratio of 0.78 for cardiovascular death (P = 0.007). I think in the generic patient this would be the first drug I would add to metformin. However, the adverse effects (nausea and vomiting, sometimes severe) are real, with 9.5% of patients self-discontinuing liraglutide in that same study. And it’s an injection. If you don’t think needles are a barrier to patient acceptance of medications, you haven’t treated a diabetic patient recently.
The cardiovascular benefits of SGLT2s have put them back in the conversation about the first drug to add for diabetic patients, a conversation from which I and many physicians had left them out previously. I’m encouraged by their observed effects on the kidney despite our lack of assurance about their statistical significance, and the fact that they can be used safely along with insulin doesn’t hurt either. But they still have the same concerning adverse effect profile that made it difficult for them to become widely adopted in the first place, and CANVAS does more to underline those concerns than to assuage them. Still, for the right patient, perhaps a patient at elevated cardiovascular risk, with some low-grade diabetic neprhopathy, without a history of vascular disease, and who is willing to chance a UTI or two? SGLT2s as a class now offer a proven reduction in cardiovascular events. As doctors fighting diabetes, having this tool can only be a good thing — as long as we’re smart in how we use it.
Read the original study here.