Is procalcitonin "safe" to guide antibiotic use in patients with sepsis?
ADAPT-Sepsis tests the strategy in the U.K., with global ambitions
Many randomized trials have tested the biomarker procalcitonin as a guide to de-escalate or stop antibiotic therapy in patients with known or suspected infection. A large proportion have concluded PCT is a safe and effective method to shorten antibiotic courses, including in patients with sepsis.
But most such trials have been small, often single-center, and vulnerable to bias due to the unblinded nature of the intervention. How reliable is this body of evidence?
The authors of a 2017 Cochrane review of 10 randomized trials with 1215 participants concluded, not very:
“Up‐to‐date evidence of very low to moderate quality, with insufficient sample power per outcome, does not clearly support the use of procalcitonin‐guided antimicrobial therapy to minimize mortality, mechanical ventilation, clinical severity, reinfection or duration of antimicrobial therapy of patients with septic conditions.”
The critically ill patients in those trials randomized to procalcitonin received about 1.3 fewer days of antibiotics. Overall, there was a trend toward safety (or even reduced mortality) with the use of PCT, but the confidence intervals in some large data slices included significant harm (30-100% increases in mortality).
This wasn’t strongly encouraging for relying on PCT in the ICU, except as an adjunct to clinical judgment in determining the total course of antibiotics in critically ill patients.
Then, in late 2024, a single randomized trial more than doubled the size of the evidence base for that equivocal Cochrane review.
What can we conclude?
The ADAPT-Sepsis Trial
Between 2018 and 2024 at 41 centers in the U.K., 2760 adults with suspected sepsis (defined as “acute organ dysfunction associated with suspected infection”, virtually all with SOFA scores ≥2, half with shock) were randomized to one of three strategies to generate automated advice for antibiotic discontinuation:
Daily procalcitonin testing, with protocolized automated advice (918 patients);
Daily C-reactive protein with protocolized automated advice (924 patients);
Standard care (no testing and a generic daily EMR prompt, 918 patients).
The test thresholds and guidance used in their protocol were as follows:
After 28 days, patients randomized to the PCT arm received 0.88 fewer days of antibiotics (9.8 days vs 10.7 days for standard care; 95% CI, 0.19 to 1.58; P = .01).
There was no difference in antibiotic exposure between the CRP-guided (10.6 days) and standard care arms (10.7 days).
The CRP-guided group generated as many prompts to stop antibiotics as the PCT group, but the PCT arm produced more of the “strongly supports” prompts.
Clinicians were permitted to check CRP off-protocol, but not PCT. Authors speculated this could have contaminated the CRP arm results.
Although the trial was conducted through the Covid pandemic, authors report there were only 19 patients included with SARS-CoV2 (possibly because eligibility for monoclonal antibody treatment was an exclusion criterion).
“Noninferior”? Statistically, but not clinically
All-cause mortality at 28 days was numerically higher in the biomarker-guided groups:
For procalcitonin, mortality was 20.9% vs 19.4% with standard care, an absolute difference of +1.57% (95% CI, −2.18 to 5.32; P = .02).
In the CRP-guided arm, mortality was 21.1% vs 19.4%, an absolute difference of +1.69% (95% CI, −2.07 to 5.45; P = .03).
Because ADAPT-Sepsis was designed as a noninferiority trial, and the authors chose (arbitrarily) a 5.4% noninferiority margin, the trial was technically positive for a PCT-guided strategy (i.e., PCT-guided antibiotic discontinuation was “noninferior” to standard care, safety-wise), and “equivocal” (the authors’ term) for a CRP-guided strategy, which just missed statistical significance for noninferiority (note the upper limit of the CI barely crossing 5.4%).
As in any noninferiority trial, the thresholds chosen deserve close scrutiny.
The higher the noninferiority threshold selected, the easier it is to “prove” noninferiority, with fewer patients enrolled and lower trial costs (in this case paid by the U.K.’s National Health Service).
In ADAPT-Sepsis, the design’s threshold of 5.4% implied an acceptance of the possibility that one in 20 patients could die from premature antibiotic discontinuation guided by PCT. This would be considered “noninferior.”
Empirically, the confidence interval’s observed upper limit of a 5.3% difference between PCT and standard care includes a ~2.5% possibility that one septic patient of every 19 treated died due to PCT guidance, with a significantly higher potential for smaller degrees of harm.
In other words, as in many similarly-designed trials, statistical noninferiority did not translate to clinical equivalence in ADAPT-Sepsis.
That didn’t stop the authors from concluding that procalcitonin-guided antibiotic protocols
“resulted in safe [emphasis added] reductions in total antibiotic duration when compared with standard care. Noninferiority for 28-day all-cause mortality, the primary safety outcome, was met for the daily PCT-guided protocol.”
“Safety,” or Statistical Feasibility?
No doubt the investigators would have liked to demonstrate a higher certainty of safety, but limited time and funding constrained them from achieving that. A trial using a clinically meaningful (i.e., much smaller) noninferiority margin would have required enrollment of many more patients and been deemed infeasible by the NHS.
Some would say we’re left with a trial that proves nothing—or at least, with a potential margin for excess death of up to 5%, nothing important.
But in the world of clinical trials, this passes as a major achievement, because it’s better than what came before.
The trial designers explain their decisions on the noninferiority margin (and the sample size and expenses it would require) in the trial protocol. Paraphrasing:
A highly-cited 2009 paper by FDA-affiliated authors argued that a loss of efficacy over placebo by up to 50% and an absolute increase in mortality of 7% are acceptable to conclude a new treatment is “noninferior” when testing antibiotics for nosocomial pneumonia.
Previous researchers testing antibiotic durations and procalcitonin use specified even higher noninferiority margins of 10%, so 5.4% is a significant improvement.
Procalcitonin-guided strategies are already considered safe in evidence assessments by U.K. health authorities (NICE), so proving such protocols to be safe with high certainty is not necessary.
ADAPT-Sepsis was designed to be positive, and was—just barely—but incorporated a margin for potential harm that will be unacceptable to most practicing clinicians.
ADAPT-Sepsis’s Tricky Head Game
Clinicians were blinded to patient allocation, and this was a significant achievement. But there was an interesting design choice: the trial employed patient-level randomization, so any given patient with suspected sepsis in a participating ICU could have been in the PCT arm, the CRP arm, or standard care.
Here’s why that could have biased toward a positive result.
Consider a clinician caring for patients with suspected sepsis in the same ICU, often several simultaneously, some of whom were randomized to biomarker guidance, and others not, over a period of five years. Think of all the alerts.
For those randomized to biomarker guidance, the clinician most frequently received repeated prompts to pursue “standard care,” followed by repeated prompts to stop antibiotics as they improved.
For those randomized to standard care, the clinician would receive repeated, identical prompts to continue standard care, with no reassuring directive to stop antibiotics.
This is not neutral. Besides creating an incentive to stop antibiotics in the biomarker arms (the intended intervention, which is fine), it created a subtle disincentive to stop antibiotics in the standard care group while a part of the clinician’s lizard brain awaited the reassuring stop advice that would never come.
Cluster randomization of ICUs (to either biomarker-guided care or standard care) would have mitigated this, but at the cost of “a much larger sample size … making the study infeasible; therefore, this was not adopted.”
But all trials testing biomarker-driven protocols are subject to similar “strange loop,” Hawthorne-like effects. That’s because the intervention acts on the clinician’s psychology, not the patient’s body directly. The protocol is intended to counteract the clinician’s (ostensibly unjustified) anxiety at stopping antibiotics with (ostensibly authoritative) advice. Human factors are inextricably baked into the intervention—because they are the intervention.
The Stakes Are Raised
These were some of the sickest patients in whom a biomarker-driven protocol has been tested, and certainly on the largest scale. Patients received an average of ~10-11 days of antibiotics in all arms, and the most aggressive prompting could only persuade the treating physicians to shave less than a day from those courses.
The authors intended for ADAPT-Sepsis to potentially change practice worldwide, by convincing the Surviving Sepsis committee members that automated procalcitonin-prompted advice on antibiotic discontinuation is ready for widespread adoption:
“This multicenter trial was designed to inform international guidance [hyperlinked to Surviving Sepsis guidelines] for both daily PCT- and CRP-guided antibiotic discontinuation protocols for sepsis.
And it sounds like they believe they may have succeeded:
“The ADAPT-Sepsis trial strengthens substantially international recommendations for the routine use of protocolized daily PCT-guided antibiotic discontinuation in critically ill adults with sepsis.”
Do you agree? Are you ready to greet a new daily pop-up in Epic that SUPPORTS you stopping antibiotics when procalcitonin falls from 40.1 to 8 μg/L, and STRONGLY SUPPORTS stopping antibiotics when PCT normalizes?
(In between answering your sepsis alerts, I mean.)
Conclusions
Numerous randomized trials, most of them small and at risk for bias, previously suggested that procalcitonin-guided antibiotic discontinuation resulted not only in reduced antibiotic exposure in severe illness, but a tendency toward lower mortality as well.
In a stroke, the much larger ADAPT-Sepsis trial cast those provisional conclusions into doubt.
The trial showed that procalcitonin-guided prompting can reduce antibiotic exposure, concordant with the strong signal from many previous trials.
But ADAPT-Sepsis did not show a procalcitonin-guided protocol to stop antibiotics is safe. Patients randomized to the strategy had 1.5% increased mortality (non-significant), and the confidence interval for harm included one excess death for every 19 patients treated with the protocol.
In patients with severe infections, procalcitonin thresholds have previously been shown to have false negative rates that are unacceptably high for their use as a guiding light in antibiotic decisions.
Yet according to the authors, ADAPT-Sepsis was designed to provide a basis for protocolizing antibiotic stewardship in sepsis care worldwide. Secondary goals were cost savings to the NHS for the 200,000 patients treated for sepsis annually in the U.K., and hypothetical reductions in antibiotic resistance.
With such grand ambitions and political entanglements, the authors may have felt duty-bound to certify centralized, procalcitonin-guided algorithmic protocols as safe and effective. In truth, the previous weak safety signal is now in question, and an even larger and more expensive trial would be needed to justify the large-scale deployment of any similar automated, protocolized strategy.
None of this increase in uncertainty disproves the potential benefits of procalcitonin as a supporting data point in the decision to stop antibiotics.
Rather, ADAPT-Sepsis resoundingly affirmed the strongest, most consistent signal in the procalcitonin literature: when treating severe infections, clinician judgment—with or without the support of biomarkers—remains the undisputed care standard.
Statistical analysis can be a "manipulative", common sense cannot
Not only is it unsafe but also it cannot be "cost effective" (the next analysis planned)! Especially since it is not effective! Nevertheless, the trumpets were blaring in Melbourne last month!