The SNAP Trial Exemplifies What EBM Should Look Like

It's always worth showing properly designed and conducted trials--though it's bad for pageviews. Nonetheless, SNAP authors show the way to answering questions in biomedicine

Staphylococcus aureus bacteremia is one of the most feared infections. It’s the leading cause of bacteria-related death worldwide, and about one in four affected patients dies within 3 months.

S. aureus comes in tho varieties—methicillin sensitive (MSSA) and methicillin resistant (MRSA). The Canadian-led SNAP trial studied MSSA.

It turns out that experts disagree on the choice of antibiotic in MSSA. I had thought it was an antistaphylococcal antibiotic like nafcillin—the Canadian equivalent would be cloxacillin, or flucloxacillin. The other choice is good old cefazolin (we used to call it Kefzol.).

A 2023 survey study found that cefazolin was the first-choice antibiotic treatment in North America (78% of respondents), whereas antistaphylococcal penicillins were preferred in all other continents (51%–82%; P < .01).

Proponents of the anti-staphylococcal meds cite a concern that in high doses of bacteria, some S. aureus strains produce beta-lactamases that can hydrolyze cefazolin in vitro, but not the antistaphylococcal penicillins. Yet a meta-analysis of observational studies suggest cefazolin may be superior.

The SNAP Trial

At 91 sites in 8 countries, hospitalized patients with MSSA bacteremia were randomized in open-label fashion to cefazolin vs an antistaphylococcal penicillin (cloxacillin or flucloxacillin) depending on the country.

The primary endpoint was excellent—death, and it was analyzed for non-inferiority, with one of the tightest non-inferiority margins I have seen—just 1.2.

The statistical plan used a Bayesian analyses—which output the probability of benefit given the data. Recall that the typical frequentist (P-value) analyses outputs the surprise factor of the data given the made up hypothesis that there is no difference in effects.

The Result

Efficacy: Mortality at 90 days among adults who could be evaluated was 15.0% (97 deaths among 645 patients) in the cefazolin group and 17.0% (109 deaths among 642 patients) in the antistaphylococcal-penicillin group (adjusted odds ratio, 0.81; 95% credible interval, 0.59 to 1.12; probability of noninferiority, 99.2%; probability of superiority, 89.8%).

Safety: Acute kidney injury occurred in 92 of 660 patients (13.9%) in the cefazolin group, as compared with 127 of 648 (19.6%) in the antistaphylococcal-penicillin group (adjusted odds ratio, 0.67; 95% credible interval, 0.50 to 0.89; probability of superiority, 99.7%).

Comments:

This looks to be a clear win for cefazolin. A 19% reduction in death and a 90% probability of superiority. Plus, there was a 33% lower chance of kidney injury. Give me cefazolin for my MSSA, please.

The strengths of the trial were numerous.

I like the adaptive Bayesian design. Not only does it output results that doctors want, ie, the probability of noninferiority or superiority, but the adaptive design increases efficiency of the trial. The authors stopped enrollment with 1,341 patients after the noninferiority threshold was met. The maximum planned sample size was 6,000. So they answered the question using roughly one-fifth of the planned maximum enrollment, sparing thousands of patients from being randomized to an arm that turned out to be inferior on safety outcomes.

The noninferiority margin was narrow at 1.2. This corresponds to an absolute risk difference of 2.5 if the control mortality was 15%. (It was 17%.)

The primary endpoint was mortality—the only bias-free endpoint.

The trial had a pragmatic design which allowed doctors to use judgement after randomization. This, plus the multi-site multi-country enrollment, increases external validity of the results.

Trial conduct was good. Missing data was minimal, and sensitivity analyses assuming all missing patients were alive (or dead) did not change the results.

Limitations:

There are always limitations. The open-label design is not likely to affect the mortality endpoint, but could have affected the safety endpoint. For instance, doctors may have been more likely to search for and diagnosis kidney issues in the antistaphylococcal arm.

The endocarditis subgroup—where cefazolin inoculum effect might be greatest—was underpowered. The point estimate in this subgroup favored cefazolin (0.54). But the credible intervals were wide (0.20 to 1.43). The probability of non-inferiority drops to 94%.

My greatest concern, and it is a technicality, was the treatment-adherent analysis. If a treatment is truly effective, you’d expect the treatment adherent patients (the ones who actually got the treatment) to show a stronger effect. In a non-inferiority test, this is actually the most conservative assessment. (As opposed to a superiority trial, where intention-to-treat is the most conservative.)

Here are the numbers for the protocol-adherent: OR 0.88, 95% credible interval 0.61 to 1.26. So, technically, the upper-bound of the 95% confidence interval of 1.26 is greater than the margin of 1.20. And the posterior probability of noninferiority in this population drops to 95.4%, well below the 99% stopping threshold that was used for the primary analysis.

One reason to be less concerned about this finding is that there was selective dropout—likely due to more toxicity in the penicillin arm (9.1% vs 1.6% changed antibiotics due to adverse events). Selectively removing sicker/more-drug-intolerant patients from the penicillin arm would biases against cefazolin in the per-protocol analysis.

Conclusion

SNAP shows the way to answering questions in Medicine. Governments funded the pragmatic well-conducted trial. Now we know that old-school cefazolin should be preferred in a deadly common infection.

And, of course, we also know that such studies are possible. Good on the SNAP team.