Does Piperacillin-Tazobactam Increase Mortality?

---

AI Survey Results

Thanks to the 124 people who took the AI survey. Results are at the end of this post.

---

Written by Clay Smith

Spoon Feed
In adult patients with sepsis, empiric use of vancomycin and piperacillin/tazobactam compared to vancomycin and cefepime, when anti-anaerobic coverage was not clinically indicated, was associated with a 5% increase in mortality.

Broader may not be better…
This was a retrospective cohort study that examined a 15-month period of piperacillin/tazobactam shortage to compare 90-day mortality in adult patients with suspected sepsis who received vancomycin and piperacillin/tazobactam or vancomycin and cefepime. To create greater equipoise between groups, they excluded patients with indications for anti-anaerobic coverage (i.e. necrotizing, head and neck, or intra-abdominal sources of infection) and patients with central nervous system infection. They adjusted for known confounders and used the shortage period as an instrumental variable – a special form a regression analysis, in which an an exogenous factor that influences treatment but is not associated with other patient characteristics or the outcome, is used to adjust for residual, unknown confounders. In this way, it can emulate a randomized clinical trial. The authors included 7,569 patients in the analysis, including 3,046 and 4,523 who received cefepime and piperacillin/tazobactam, respectively. In the unadjusted analysis, vancomycin and piperacillin/tazobactam, compared to vancomycin and cefepime, was associated with a 2.6% (95%CI 0.7-4.4%) increase in 90-day mortality. With instrumental variable analysis, there was 22.5% versus 17.5% mortality at 90 days, an increase of 5% (95%CI 1.9%-8.1%; P = 0.002) among patients who received vancomycin and piperacillin/tazobactam compared to vancomycin and cefepime. Additionally, the piperacillin/tazobactam cohort had fewer organ failure-free days, ventilator-free days, and vasopressor-free days than the cefepime cohort. Assuming the association with a 5% difference in mortality was correct, this would have been a number needed to harm of 20.

How will this change my practice?
For me, this study is practice changing. After ACORN, I had largely switched to piperacillin/tazobactam over cefepime, since ACORN demonstrated there was no increased risk of acute kidney injury with piperacillin/tazobactam, and cefepime was associated with increased neurotoxicity.  However, this study gives me pause. It seems that broader anaerobic coverage may not only be unnecessary in many patients, it is very likely harmful.  As we learn more about the importance of healthy gut flora, this is biologically plausible.

Source
Mortality of Patients With Sepsis Administered Piperacillin-Tazobactam vs Cefepime. JAMA Intern Med. 2024 Jul 1;184(7):769-777. doi: 10.1001/jamainternmed.2024.0581. PMID: 38739397.

---

AI Survey Results

As promised, here are the full results of the AI Survey for JournalFeed. Thanks to the 124 people who responded.

Here are my thoughts.

• Question 1: Overall, most readers seemed to favor the hybrid AI and human summary, but 17% viewed it unfavorably.
• Question 2: The first summary was written 100% by AI, even the tag line “Choose your weapons wisely: cefepime vs. piperacillin-tazobactam in sepsis.” Most (65%) of you were able to spot it was AI. The descriptions of what gave it away were highly varied.
• Question 3: The second summary was written 100% by me (Clay Smith). Again, most (63%) were able to tell it was written by a human. It was humbling that some thought the AI summary was better and clearer than the one I wrote. One commenter thought it was AI-generated and had, “some difficulty parsing what the key pearls from the study are for the emergency physician.” ?
• I have also written a summary of this article for Annals of Emergency Medicine Journal Club that will post in December 2024, which is 100% written by me with no portion written by AI.

Summary: The current base version of Chat GPT 4o can summarize a PDF academic article incredibly well. With proper prompts, it can do it in an academic tone for an emergency physician audience and can even write in the first person as if it were an emergency physician. It was difficult for one third of readers to tell which summary was AI-generated vs human-written, which raises an important ethical issue. I think it is unethical to use AI to do all the writing and to not disclose that. I find this particularly troubling in the section on “How will this change my practice?” AI doesn’t practice medicine. Only humans practice medicine. With the right prompt, we can make AI sound like an ER doc, speak in first person, and write things that are factually correct. However, when it comes to patient care and changing practice (or not…), this should come from a human who cares for real patients at the bedside…in my opinion. AI is a tool, and JournalFeed may begin to use it for some tasks.

However, our commitment regarding AI is this:

1. Transparency – You deserve to know who (or what) is writing. We will make this clear.
2. Oversight – AI is like a good intern…helpful most of the time, but you gotta watch or it might do something crazy.
3. Context, critique, commentary – AI is very good at summarizing what’s written down. It is not so good at understanding how an article fits into the broader context of medical literature. It is also not good at understanding what’s not written, that is – reading between the lines. What did the authors not say that they should have? What was the fatal flaw of the study? How should we view this critically? Finally, only a physician (clinician) who sees real patients should comment on how an article will change (or not change) your practice. This will always be written by a human.

Clay Smith

---