Infectious disease · Meta-analysis
Sohani ZN, Zhong YJ, Afshar A, et al. Cefepime and mortality: a systematic review and bayesian meta-analysis. JAMA Netw Open. 2026;9(9):e2633017. doi:10.1001/jamanetworkopen.2026.33017
Cefepime vs other β-lactam antibiotics, for children and adults treated for infections: deaths
Fact box
How to read this page: this review reports its main result as a chance, not as a yes-or-no test. So each row answers two separate questions: how big the difference was (people, out of 100 or 1,000), and how sure the direction is (the chance, in %, that deaths were at all higher, or lower, with cefepime). A high chance does not mean a big difference.
Who this is for: children and adults treated in trials for infections, most often fever with low white blood cells (febrile neutropenia), pneumonia and severe bacterial infections. Average age was about 55 in the adult trials and between 5 and 6 in the children's trials. The numbers come from 110 randomised trials with 22,608 people counted in total.
What the treatment meant: cefepime, with or without a second drug, compared with another β-lactam antibiotic (most often ceftazidime, also carbapenems, piperacillin-tazobactam, cefotaxime and ceftriaxone), with the same second drug if one was used.
How long people were followed: deaths were counted at 30 days, or at the closest time each trial reported.
| Another β-lactam | Cefepime | How sure is the direction? |
|---|
Quality of the evidence: moderate (GRADE from the review), lowered for differences between trials and possible publication bias. The review's table lists two serious limitations but lowers the rating by one step (see Limitations and criticism).
Other things to consider: this page opens on the 1,000 view, because the difference is too small to show well out of 100. The answer depends on which trials you trust. Published trials and unpublished trials given to the FDA pointed in opposite directions, and the unpublished trials cannot be examined in detail. The review counts deaths from any cause: in people with cancer, death may reflect the disease, and the review does not show why any extra deaths would happen. Most trials were old: 68 of the 85 dated trials came out before 2010. ACORN, the largest and most recent trial, found no clear difference at 28 days. Follow-up times differed between trials and the all-trials range includes no difference, so no number needed to harm is shown here. Some people appear to have been counted twice (see Limitations and criticism).
Main outcome as a picture
Another β-lactam
Cefepime
Study question (PICO)
Population, intervention, comparison, outcomes, each trial's result, and who was includedclick to show
- Question
- What is the association of cefepime with all-cause mortality compared with other β-lactam antibiotics?
- Population
- Children and adults in 110 randomised trials: 22,608 patients (cefepime 11,726; comparator β-lactam 10,882), about 56% male. 73 trials were published and peer reviewed, 12 were abstracts or other unpublished reports, and 25 were unpublished trials provided by the FDA. Among published trials, febrile neutropenia was the most studied indication (31 trials). Weighted mean age was 54.8 years with cefepime in adult trials and 5.5 years in paediatric trials.
- Intervention
- Cefepime alone, or cefepime plus a second drug. Doses were grouped as high (at least 2 g every 8 hours in adults), intermediate (at least 2 g every 12 hours but less than 2 g every 8 hours) and low (less than 2 g every 12 hours). In children the same groups were at least 50 mg/kg every 8 hours, at least 50 mg/kg every 12 hours, and less than 50 mg/kg every 12 hours.
- Comparison
- Any other β-lactam alone, or a β-lactam plus the same second drug. Ceftazidime was the most common comparator.
- Outcomes
- Primary: all-cause mortality at 30 days, or the reported time point closest to 30 days. Clinical success and microbiological failure were prespecified but were not meta-analysed, because reporting was inconsistent. Adverse events are to be reported separately.
- Design of the review
- Systematic review and Bayesian random-effects meta-analysis, registered with PROSPERO (CRD42023455194). Cochrane CENTRAL, Medline, Embase, Web of Science, WHO ICTRP, ClinicalTrials.gov and LILACS searched from inception to May 25, 2026, with no language restriction. Two independent reviewers extracted data; risk of bias with the Cochrane tool, version 2.0; certainty with GRADE. Results are given as the probability that deaths were higher with cefepime.
- Each trial's result
-
The review's forest plot for its 10 largest trials, labelled in people. With 110 trials, the rest are grouped in one row. The review printed no weights, so the figure shows each trial's share of the people instead.
- Inclusion criteria (review)
- Randomised trials of children and adults comparing cefepime alone with any β-lactam alone, or cefepime plus a second drug with a β-lactam plus the same second drug. 25 of 41 FDA-provided trials met these criteria.
- Exclusion criteria (review)
- Studies of cefepime as prophylaxis, and cefepime with a β-lactamase inhibitor. Sixteen FDA-provided trials were left out because they used non–β-lactam comparators or concomitant antibiotics that were not balanced between arms. Of 118 trials found, 110 had mortality data. Each trial's own exclusion criteria were not checked.
Source: Sohani et al., 2026, Abstract, Methods, Results, Table 1, Figure 2, and Supplements 1 and 2.
Limitations and criticism
15 published and own critical points, plus sources not yet checkedclick to show
Scite's tally shows 1 citing publication and its citation graph shows 2, the invited commentary and Leibovici 2026; it classifies none of their statements as contrasting (checked 6 November 2026). No correction or retraction notice is listed.
“all-cause mortality is an imperfect safety outcome”
The authors say many trials were open label, so doctors could change doses, duration or other treatments after randomisation. Patient-level data were not available, the unpublished trials gave few details, and in cancer patients death may reflect the disease. They call the result a global safety signal, not evidence of a particular cause, and say it does not mean cefepime should no longer be used.
Sohani et al., 2026, Discussion and Conclusions (read in full, PDF supplied).
The review's main figure, 94.4%, is the chance that deaths were at all higher with cefepime, however small the increase. It is not the size of the increase. By our conversion of the review's result, the size was about 6 more deaths in every 1,000 people, with a range from 1 fewer to 14 more, and that range includes no difference. The review's own raw totals, which add up the trials without weighting them, give a difference of 0.4 points, about 4 in every 1,000. The authors conclude that cefepime was associated with more deaths; a yes-or-no test on the same result would have said no clear difference. Both descriptions fit the data. From the review's result and its range, the chance of an increase of at least 5 deaths in every 1,000 is about 58%, and of at least 10 in every 1,000 about 14% (our calculation: a bell-shaped curve fitted to the review's range and to its 94.4%; our separate run from the trial counts gave similar figures). For this reason the fact box shows the best estimate in every row with the chance beside it, does not use “about the same in each group”, and shows the chances as 50 equally likely answers under the picture.
Our reading of Sohani et al., 2026, Abstract, Results and Number Needed to Harm; Uslan and Smith, 2026.
The review's probability of any increase in deaths was 98.6% in the 73 published trials and 11.7% in the 25 unpublished FDA trials. By raw totals, deaths were 7.3 vs 6.4 in every 100 in published trials and 5.1 vs 6.2 in the FDA trials (cefepime first). The authors give a 98.8% probability that the unpublished studies differ from the published literature. They offer selective publication or different patients, doses or antibiotics as explanations, and could not tell which. The FDA trials had no risk-of-bias assessment, and funnel and Egger plots did not show significant publication bias. Adding the 12 abstracts and other reports to the published trials (85 trials) still gave a range above no difference (eFigure 9). The main result pools all these sets, so it mixes groups that disagree.
Our reading of Sohani et al., 2026, Figure 2, Publication Bias, Discussion and eTable 2.
“A high probability of a small effect is not evidence of a large one”
The commentary praises the probability approach but warns readers against confusing likelihood with size: the difference is 0.4 percentage points. It argues that old trials bear limited resemblance to current practice and that both underdosing and overexposure could explain the signal. It calls for release of the FDA's patient-level data and studies of optimised dosing, and advises 2 g every 8 hours with careful kidney adjustment. Our check: it gives 22,598 patients where the review gives 22,608, and quotes the FDA's patient-level mortality without its trial-level figures.
Uslan and Smith, 2026, JAMA Netw Open (read in full).
ACORN makes up 11.1% of the people in the review, and the review's counts (104 of 1,214 and 106 of 1,297) match the trial's day-28 deaths, which the trial called exploratory: 8.6% vs 8.2%, a difference of 0.4 points that could be from 1.9 fewer to 2.6 more. At day 14, the trial's deaths were 7.6% vs 6.0%, a difference that could be from 0.5 fewer to 3.6 more. Both ranges include no difference. Read the way this page reads the review, the chance of any increase was about 94% at day 14 and about 64% at day 28 (our calculation from the trial's ranges). Death was not the trial's main outcome: its main outcome, the worst stage of kidney injury or death by day 14, did not differ significantly between the groups. Follow-up stopped at hospital discharge, people received the assigned drug for a median of 3 days, and 18.8% and 17.2% received at least one dose of the other drug. Delirium or coma was more common with cefepime: 20.8% vs 17.3%.
Our check of Sohani et al., 2026, eTable 2; Qian et al., 2023, Table 2 (read in full); ClinicalTrials.gov NCT05094154.
Seo 2017 had 6 patients on cefepime, 2 of whom died. The review lists this arm twice, once against ertapenem and once against piperacillin-tazobactam (2 of 6 each time), as two trials. Nakane 2015 shows the same pattern, 4 deaths in 87 cefepime patients listed twice under one reference, but we could not read that paper. If both are shared arms, 93 cefepime patients and 6 deaths are counted twice in the totals of 11,726 and 778; Seo alone accounts for 6 patients and 2 deaths. The Seo trial stopped giving cefepime after 6 patients because of treatment failures. The numbers on this page are based on the review's totals and are not adjusted for any of this.
Our check of Sohani et al., 2026, eTable 2 and eFigure 9; Seo et al., 2017 (methods and results read via Scite). Nakane 2015 itself was not read.
Sáez-Llorens 2001 (174 and 171 children) is a review of two Latin American trials in 345 children, comparing cefepime with cefotaxime or ceftriaxone. The review counts it as one trial and also counts Sáez-Llorens 1995 (cefepime vs cefotaxime) and O'Ryan 1996 (cefepime vs ceftriaxone); together those two hold 9 and 9 deaths, against 10 and 9 in the 2001 report. The deaths fit closely and the people less well (298 against 345), and we could not confirm the overlap from the abstract. Gomez 2001, an abstract, and Gomez 2010 have the same regimens and overlapping authors, but their counts differ widely (13 of 86 and 5 of 100, against 15 of 152 and 17 of 165), so this is a question, not a finding. Gomez 2010's 152 and 165 are episodes of illness, not people (317 episodes in 190 patients). Unpublished trial AI411201 (5 of 56 and 2 of 55) closely matches the Willis 1998 abstract (5 of 54 and 2 of 55), both against cefotaxime. If that is one trial, the same deaths are counted twice among the unpublished entries.
Our check of Sohani et al., 2026, eTable 2 and references; Sáez-Llorens and O'Ryan, 2001, and Gómez et al., 2010 (abstracts). Willis 1998 and the FDA trial reports could not be read.
The review says higher doses had a higher probability of harm: 80.5% for low doses, 93.7% for intermediate and 94.6% for high doses. Converted to people, all three show about the same size of difference, 7 to 11 more deaths in every 1,000 by best estimate, with the low-dose group having the fewest trials (15, against 30 and 27). So the probabilities may reflect how much data each group had, not a dose effect. An earlier review, Andreatos 2017, reported that excess deaths in febrile neutropenia seemed to come from low-dose trials, but its low dose (2 g every 12 hours) is what this review calls intermediate, and it noted that the intervals overlapped. Doses were grouped by trial, and adjustments for kidney failure varied.
Our reading of Sohani et al., 2026, Figure 3D and Discussion; Andreatos et al., 2017 (read in full).
The GRADE table marks inconsistency as serious, downgrading 1 level, and publication bias as serious, consider downgrading 1 level, yet the final rating is moderate, which is one step down from high. It also says the range excluded no effect, which is true of the published trials but not of all 110. The sample size it calls sufficient (2,554 patients) assumed deaths of 8 and 6 in 100, a difference of 2 points. To detect the raw difference seen across all trials (778 of 11,726 and 674 of 10,882), a trial would need about 48,400 people in each group (our calculation, 80% power, 5% two-sided). The table says heterogeneity was significant, but the review prints no figure for it. A standard overall test on the trial counts in eTable 2 did not find more variation between trials than chance would give (our calculation: Cochran's Q of 72.6 on 88 degrees of freedom, for the 89 trials with at least one death). Such a test can miss a difference between two sets of trials, and the split between published and unpublished trials, described above, is a separate matter.
Our reading of Sohani et al., 2026, Table 2, Results and eTable 2.
The review reports a number needed to harm from raw totals added across trials, a difference of 0.4 points for all trials and 0.9 for published trials. Our protocol does not show it: deaths were counted at different times, the all-trials range includes no difference, and adding raw counts ignores that some trials gave more patients cefepime than the comparison drug.
Our reading of Sohani et al., 2026, Number Needed to Harm.
Of the 110 trials, 11 were judged at low risk of bias, 34 had some concerns, 25 were at high risk and 40 could not be assessed. The review's result did not change much without the high-risk trials (95.7% probability). Of the 85 trials with a date, 68 were published before 2010. The 25 trials at high risk of bias are not a row in this fact box; the review shows them in eFigure 2.
Our reading of Sohani et al., 2026, Risk of Bias in Studies and eTable 2.
The main text and the supplement print slightly different limits for several results, including the lower limit for published trials; both stay above no difference. The paper and the authors' model code give different values for the prior on the treatment effect. Some trial names and years differ between eTable 2 and the forest plots (Lee, Ramphal, Avilés-Robles, Naseem). Counts in eTable 2 matched eFigure 9 for all 110 trials. Using those counts and the authors' priors for the treatment effect and for differences between trials, our own approximate run gave the same best estimates and ranges for the two main results. Its probabilities differed by less than 1 point, and it did not reproduce the subgroup rows, which the review models differently.
Our check of Sohani et al., 2026, main text, Supplement 1, Supplement 2 and the authors' public model notes.
“cefepime should be avoided in patients with neutropenic fever or with skin structure infections”
This debate began with a 2007 review by Yahav and colleagues. The FDA's analysis found raw deaths of 6.21% vs 6.00% across 88 trials and 5.63% vs 5.68% in the 35 trials with patient-level data. As Mantel-Haenszel adjusted differences, both came to about 5 more deaths in every 1,000, which could be from 2 fewer to 12 more, and from 5 fewer to 14 more; it found no statistically significant increase. Kalil's Bayesian re-reading gave probabilities of harm in febrile neutropenia of 90.9% and 80.8% from the FDA's data and 99.2% from Yahav's. Andreatos 2017 found more deaths with cefepime in febrile neutropenia. These are older reviews with fewer trials, so this page uses Sohani.
Kim et al., 2010 (abstract); Kalil, 2011 (abstract); Andreatos et al., 2017 (read in full); Sohani et al., 2026, Introduction.
“The recent IV analysis does not support a mortality benefit for cefepime”
Chanderraj 2024 used a piperacillin-tazobactam shortage at one US health system to compare 7,569 patients with sepsis, and found more deaths by 90 days with piperacillin-tazobactam: a difference of 5.0%, according to the commentary. Hamilton and colleagues argue that adjusting for metronidazole introduced bias. Three of the Hamilton authors, Hamilton, Lee and Butler-Laporte, are also authors of this review.
Chanderraj et al., 2024 (methods read via Scite; result taken from the commentary); Hamilton et al., 2026 (abstract); Uslan and Smith, 2026.
Yahav and Paul, authors of the 2007 review that first raised the alarm, are also authors of this review. Leibovici, a co-author of that review, has written a commentary on the "cefepime story", of which we could read the opening only. Paul and Yahav report grants paid to their institutions from Shionogi. The authors state that ChatGPT was used to keep within word limits.
Our reading of Sohani et al., 2026, references and Article Information; Leibovici, 2026 (opening paragraph).
- Full text of Kim et al., 2010, the FDA analysis (doi 10.1086/655131), which may show whether unpublished FDA trials are also reported in published papers.
- Full text of Yahav et al., 2007 (doi 10.1016/S1473-3099(07)70109-3), and the 2010 letter by Leibovici, Yahav and Paul.
- Full text of Leibovici, 2026, Clin Microbiol Infect (doi 10.1016/j.cmi.2026.10.020).
- Full texts of Sáez-Llorens and O'Ryan, 2001, and Gómez et al., 2010, and the abstracts O'Ryan 1996, Gomez 2001 and Willis 1998, to confirm the possible double counting.
- The published letters about ACORN, and the 2026 Surviving Sepsis Campaign guideline.
- All other individual trial papers. Apart from ACORN and Seo 2017, trial counts were checked against the review's own tables, not against the trials.
References
15 references, with how much of each was readclick to show
- Sohani ZN, Zhong YJ, Afshar A, et al. Cefepime and mortality: a systematic review and bayesian meta-analysis. JAMA Netw Open. 2026;9(9):e2633017. doi:10.1001/jamanetworkopen.2026.33017 Read in full
Full text read (PDF supplied); Supplements 1 to 3 and the authors' public model notes read - Uslan DZ, Smith EA. Cefepime and mortality—a dosing problem, not a drug problem. JAMA Netw Open. 2026;9(9):e2632904. doi:10.1001/jamanetworkopen.2026.32904 Read in full
Read in full on the journal's site - Qian ET, Casey JD, Wright A, et al. Cefepime vs piperacillin-tazobactam in adults hospitalized with acute infection: the ACORN randomized clinical trial. JAMA. 2023;330(16):1557-1567. doi:10.1001/jama.2023.20583 Read in full
Read in full on the journal's site; supplement not read - Seo YB, Lee J, Kim YK, et al. Randomized controlled trial of piperacillin-tazobactam, cefepime and ertapenem for the treatment of urinary tract infection caused by extended-spectrum beta-lactamase-producing Escherichia coli. BMC Infect Dis. 2017;17(1):404. doi:10.1186/s12879-017-2502-x Partly read
Methods and results read via Scite - Kim PW, Wu YT, Cooper C, et al. Meta-analysis of a possible signal of increased mortality associated with cefepime use. Clin Infect Dis. 2010;51(4):381-389. doi:10.1086/655131 Partly read
Abstract only - Yahav D, Paul M, Fraser A, Sarid N, Leibovici L. Efficacy and safety of cefepime: a systematic review and meta-analysis. Lancet Infect Dis. 2007;7(5):338-348. doi:10.1016/S1473-3099(07)70109-3 Could not read
Could not be read; described from other papers - Kalil AC. Is cefepime safe for clinical use? A Bayesian viewpoint. J Antimicrob Chemother. 2011;66(6):1207-1209. doi:10.1093/jac/dkr138 Partly read
Abstract only - Andreatos N, Flokas ME, Apostolopoulou A, Alevizakos M, Mylonakis E. The dose-dependent efficacy of cefepime in the empiric management of febrile neutropenia: a systematic review and meta-analysis. Open Forum Infect Dis. 2017;4(3):ofx113. doi:10.1093/ofid/ofx113 Read in full
Read in full via Europe PMC - Chanderraj R, Admon AJ, He Y, et al. Mortality of patients with sepsis administered piperacillin-tazobactam vs cefepime. JAMA Intern Med. 2024;184(7):769-777. doi:10.1001/jamainternmed.2024.0581 Partly read
Methods read via Scite; main result taken from the commentary - Hamilton F, Lee TC, Davey Smith G, Hartwig FP, Butler-Laporte G. Does piperacillin-tazobactam increase mortality risk compared with cefepime? Collider bias and the importance of assumptions in instrumental variable analyses. Clin Infect Dis. 2026;82(1):e17-e23. doi:10.1093/cid/ciaf317 Partly read
Abstract only - Sáez-Llorens X, O'Ryan M. Cefepime in the empiric treatment of meningitis in children. Pediatr Infect Dis J. 2001;20(3):356-361. doi:10.1097/00006454-200103000-00035 Partly read
Abstract only - Gómez L, Estrada C, Gómez I, et al. Low-dose beta-lactam plus amikacin in febrile neutropenia: cefepime vs. piperacillin/tazobactam, a randomized trial. Eur J Clin Microbiol Infect Dis. 2010;29(4):417-427. doi:10.1007/s10096-010-0879-1 Partly read
Abstract only - Leibovici L. Lessons to be learned from the 'cefepime story'. Clin Microbiol Infect. Published online October 27, 2026. doi:10.1016/j.cmi.2026.10.020 Partly read
Opening paragraph only - Mastrandrea MD, Field CB, Stocker TF, et al. Guidance note for lead authors of the IPCC Fifth Assessment Report on consistent treatment of uncertainties. Intergovernmental Panel on Climate Change; 2010. ipcc.ch Read in full
Read in full; used only for the words given to the chances - EFSA Scientific Committee, Benford D, Halldorsson T, et al. Guidance on uncertainty analysis in scientific assessments. EFSA J. 2018;16(1):5123. doi:10.2903/j.efsa.2018.5123 Partly read
Probability scale and the text around it read; used only for the words given to the chances
How the numbers were checked
Check logclick to show
Independent fact check (6 November 2026), versions 1 and 2: agents A, B and C, working only from the page text and the sources, without the checks above. Agent C transcribed the forest plots before seeing the page; its images of Figures 2 and 3 were full pages, so that pass was not strictly blind to the paper's body text, and its 110 trial rows matched the text layer. First pass: agent A, 230 lines (219 match, 4 not found, 6 not checkable, 1 confirmed absent); agent B, 147 lines (131 match, 3 mismatched, 2 not found, 10 not checkable, 1 confirmed absent); agent C, 246 lines (242 match, 1 mismatched, 3 not checkable). All were resolved on the page and the changed lines rechecked: agent A, 238 lines (229 match, 6 not checkable, 3 confirmed absent); agent B, 48 lines then a final 12 lines, all 12 matching after 2 further corrections. Final state: no line mismatched or not found. Resolutions: sample size recalculated from exact counts; wording beyond the sources removed; the double-counting items softened to what the evidence shows; Andreatos's dose definition added; the claim about reproducing the results narrowed. Lines left not checkable are page metadata, the page's own process statements, drawing geometry (tested by the render checks above), and reading badges. Versions 3 and 4 (new display for a review that reports chances): the same three agents rechecked the redesigned page. Agent A covered its whole part three times: 359 lines (346 match, 1 mismatched, 1 not found, 8 not checkable, 3 confirmed absent), then 374 lines (361 match, 1 mismatched), and finally 385 lines (373 match, 8 not checkable, 4 confirmed absent). Agent B checked the criticism and references: 40 lines (39 match, 1 confirmed absent), then 23 and 19 lines on the changed sentences, all matching. Agent C recomputed every table row at both denominators, the pictures and the chances from its own transcription of the forest plots: 285 lines (279 match, 6 not checkable). Final state: no line mismatched or not found. Resolutions: the opening paragraph no longer says the review gives no yes-or-no answer; the last column gives the chance of the more likely direction in every row; the headline separates the direction from the best estimate and says “less than 1” where rounding would overstate; “In short” is given per 1,000; the largest trial is described the same way as the review; the test of variation between trials is worded as not finding a difference; the two scales behind the words are described exactly. Lines left not checkable are page metadata, the page’s own process statements, reading badges and the geometry of the drawings, which the render checks above test. Version 5 (new text above the dots; no number changed): a fresh agent A checked the new paragraph and the plot under it in all 10 states, in four passes as the wording was corrected: 99 lines (84 match, 3 mismatched, 2 not checkable, 10 confirmed absent), then 101 lines (87 match, 1 not found), then 105 lines (89 match, 1 not found), and finally 106 lines (90 match, 3 not checkable, 13 confirmed absent). Final state: no line mismatched or not found. Resolutions: “answer” now means a dot and nothing else, as in the heading; the 100% and the 2% are called a chance; no reason is given for why the trials cannot give the number exactly, because the reason first given was not the review’s; the dots are compared (more likely where they pile up than where they are thin) and no single pile is called the most likely; the comparison group is named as on the rest of the page. Agents B and C were not rerun: no criticism, reference, table row or trial figure changed. Left standing, rated low by the agent: the text speaks of the true difference as one number, though a difference in every 1,000 depends on how many died in the comparison group; “differential diagnosis” is a doctor’s term, kept on purpose. Lines left not checkable are the drawing itself and the collapsed check log, which the render checks above test.
Version history
v1, 6 November 2026: first version, draft.
v2, 6 November 2026: corrected after the independent fact check; draft banner removed.
v3, 6 November 2026: display changed for a review that reports chances. Every row now shows the best estimate with its range and the chance of any increase; “about the same in each group” is no longer used; the chance of larger differences was added; the page opens on the 1,000 view. Draft again until these changes are independently rechecked.
v4, 6 November 2026: after reading the literature on wording and display, and after the independent fact check: the words follow the IPCC cut-offs (very likely from 90%, likely from 66%) in place of our own stricter scale; 50 dots in place of 100; the headline separates the direction from the best estimate and never rounds a difference under 1 upward; rows where a decrease is more likely show the chance of a decrease; the largest trial is described the same way as the review; a test of variation between trials was added; a note says the lines at 5 and 10 are reference marks.
v5, 6 November 2026: the text above the dots was rewritten, because a first reader could not follow it. It now explains the dots with the example of a doctor weighing the differential diagnoses of chest pain. No number changed; the new text was independently rechecked.
For discussion
- Published trials say harm and unpublished FDA trials say no harm. Which do you believe, and what would change your mind?
- A 94% probability of harm sounds alarming; a difference of a few deaths per 1,000 does not. Which should a prescriber hold on to?
- ACORN found more delirium or coma with cefepime, but no clear difference in deaths at 28 days. Does that change your choice for a patient with poor kidney function?
- Does this change what we do on Monday?
Comments (0)
Spotted an error in the numbers? Say so here and the review will be corrected. Mock site: comments are saved in this browser only.