Persistent or recurrent fever after 4 to 7 days of broad-spectrum antibacterials, with no documented infection, is the classic trigger for a mold-active antifungal. In 1095 evaluable patients, caspofungin met non-inferiority to liposomal amphotericin B (overall success 33.9% vs 33.7%; 95.2% CI for the difference -5.6 to 6.0), was better tolerated (nephrotoxicity 2.6% vs 11.5%), did better in patients with a baseline fungal infection (51.9% vs 25.9%; P=0.04), and left more patients alive at 7 days after therapy (92.6% vs 89.2%; P=0.05).1
Voriconazole versus liposomal amphotericin B did not settle as cleanly. In 837 evaluable patients, overall success was numerically lower with voriconazole (26.0% vs 30.6%; 95% CI for the difference -10.6 to 1.6), so the trial did not establish non-inferiority on its primary composite, even though voriconazole produced fewer documented breakthrough infections (1.9% vs 5.0%; P=0.02). The negative lower bound of that interval is why voriconazole is not a standard empiric agent here, and why caspofungin and liposomal amphotericin B remain the reference options.2
These empiric-antifungal trials score a favorable response only when all five components are met: (1) successful treatment of any baseline fungal infection, (2) no breakthrough fungal infection, (3) survival at least 7 days after therapy, (4) no premature discontinuation for toxicity or lack of efficacy, and (5) resolution of fever during neutropenia. Component 5 is hard - fever often persists while the marrow is empty - so headline success rates sit near 33% even for effective drugs. Read these as a stringent composite, not as 2 in 3 patients failing.1
The diagnostic-driven (pre-emptive) strategy withholds antifungals until galactomannan, beta-D-glucan, or chest CT signals disease. The trials do not converge on a single answer, and the common belief that pre-emptive management is unsafe is not what the larger studies show.
The first randomized test (PREVERT, N=293) found survival non-inferior (97.3% empiric vs 95.1% pre-emptive; lower 95% confidence limit -5.9%, within the pre-specified -8% margin), but proven or probable fungal infection was more common in the pre-emptive arm (13 of 143 vs 4 of 150), concentrated during induction chemotherapy (12 of 73 vs 3 of 78). Pre-emptive care cut antifungal drug cost by about 35%.3
The larger EORTC trial (549 evaluable of 556 randomized) put both strategies on a fluconazole-prophylaxis backbone and triggered the pre-emptive arm on twice-weekly galactomannan plus CT on demand. Day-42 overall survival was non-inferior for the pre-emptive arm (96.7% vs 93.1% empiric; the protocol accepted a maximum relative risk for death of 1.62), invasive fungal disease at day 84 did not differ (7.7% vs 6.6%), and the pre-emptive strategy halved antifungal exposure (27% vs 63% received caspofungin).4
A third approach triggers on the D-index, the running area under the neutrophil-count-below-500 curve. In CEDMIC (N=413), D-index-guided micafungin was non-inferior to empiric micafungin for proven or probable fungal infection (0.5% vs 2.5%), with identical survival (98.6% vs 98.0% at day 42) and micafungin use cut from 60.2% to 32.5%; the micafungin subgroup analysis put composite efficacy at 68.7% (D-index) versus 79.6% (empiric), without a significant difference in the high-risk stratum (69.1% vs 78.0%, P=0.30).5,6
Pooled data hold this line without resolving it. A 2022 Cochrane review (7 RCTs, 1480 high-risk patients) found little-to-no mortality difference (RR 0.97, 95% CI 0.72-1.30) and a shorter antifungal course with pre-emptive care (mean difference -3.52 days), at low to very-low certainty. A 2015 meta-analysis with a cost model found less antifungal exposure (RR 0.48) and no mortality signal, at roughly break-even cost. The economic case for pre-emptive management is real but fragile - it swings on local drug and test prices.7,8
| Strategy | Trigger to start antifungal | Antifungal exposure | IFI / survival signal |
|---|---|---|---|
| Empiric | Persistent fever 4-7 days | Highest (EORTC 63%) | Reference; fewest missed IFI |
| Pre-emptive (GM/CT) | Galactomannan and/or CT positive | ~27% (EORTC) | Non-inferior survival; more IFI in PREVERT induction arm |
| D-index-guided | D-index at least 5500 or positive test | ~33% (CEDMIC) | Non-inferior IFI; survival equal |
A non-inferiority trial asks whether a new strategy is not worse than the standard by more than a pre-specified margin. The margin can be an absolute risk difference (PREVERT: -8% survival) or a maximum acceptable relative risk (EORTC: 1.62 for death). Diagnostic-driven means a biomarker or image - galactomannan, CT, or the D-index - triggers therapy instead of fever alone. The strategy only pays off when those tests are reliable and timely.3,4
Galactomannan detects an Aspergillus cell-wall antigen: across 27 studies, pooled sensitivity 0.71 (95% CI 0.68-0.74) and specificity 0.89 (0.88-0.90) for proven invasive aspergillosis, and it performs better in hematology and transplant patients than in solid-organ transplant recipients. Its structural weakness: mold-active prophylaxis suppresses antigen release and lowers sensitivity - so the screen used to justify a pre-emptive strategy is least reliable in exactly the prophylaxed patients who are now the norm.9,10
(1,3)-beta-D-glucan is pan-fungal, not Aspergillus-specific: pooled sensitivity 76.8% (95% CI 67.1-84.3), specificity 85.3% (79.6-89.7), summary AUC 0.89. Treat it as a sensitive screen, not a rule-in. False positives are common with cellulose hemodialysis membranes, intravenous immunoglobulin, albumin or blood products, glucan-containing gauze, and some intravenous antimicrobials.11,12
On CT, the halo sign is early and transient: present in 24 of 25 neutropenic patients with proven invasive pulmonary aspergillosis on the day of diagnosis, but in only 68%, 22%, and 19% of cases at days 3, 7, and 14; the air-crescent sign is late (8%, 28%, 63% on the same days). Scan the neutropenic chest early - waiting loses the most specific radiologic clue.13
A newer molecular test may repair the galactomannan gap. In 238 patients, Aspergillus plasma cell-free DNA PCR was more sensitive than serum galactomannan for proven or probable aspergillosis in hematology and transplant hosts (92.0% vs 67.9%; P=0.04) at similar specificity (91.8% vs 89.8%), and - the point that matters - it held its sensitivity under mold-active prophylaxis (roughly 83 to 86%) where galactomannan sensitivity fell to about 56%. The data are single-center and retrospective, so this is promising rather than established, but it targets precisely the population in which galactomannan fails.14
| Test | Pooled sensitivity / specificity | Best use | Key caveat |
|---|---|---|---|
| Galactomannan | 0.71 / 0.89 (proven IA) | Heme/HSCT; supports pre-emptive screening | Mold-active prophylaxis blunts sensitivity |
| beta-D-glucan | 0.77 / 0.85 (any IFI) | Sensitive pan-fungal screen / rule-out | Many false positives; not mold-specific |
| CT halo sign | Early, transient | Prompt imaging when neutropenic + pulmonary signs | Largely gone by day 7 |
| Plasma cfDNA PCR | 0.92 / 0.92 (heme/HSCT, proven/probable) | Breakthrough diagnosis under prophylaxis | Single-center data; limited availability |
A negative galactomannan in a patient on posaconazole or voriconazole does not exclude aspergillosis: mold-active drug suppresses antigen release, dropping an already-imperfect 0.71 sensitivity further. This is the structural flaw of galactomannan screening today - the population most likely to be screened is the one in which the screen is least reliable. A molecular assay that keeps its sensitivity under prophylaxis, such as plasma cell-free DNA PCR, is the plausible fix.9,14
PROVEN = organism from a normally sterile site by culture or histopathology (any host). PROBABLE (immunocompromised only) = a host factor PLUS a clinical feature PLUS mycologic evidence (for aspergillosis, e.g. serum or plasma galactomannan at least 1.0, or a positive culture/PCR). POSSIBLE = host factor plus clinical feature with NO mycology. Beta-D-glucan counts as mycologic evidence for IFD in general but is explicitly NOT accepted as evidence of a specific invasive mold. These are research-enrollment definitions - the authors state they are not meant to direct patient care.10
Even with modern antifungals, invasive aspergillosis in hematologic malignancy stays lethal, and the gains may be plateauing. In a 10-year single-center cohort of 111 culture-positive cases, attributable mortality was 18.2% at 6 weeks and 22.7% at 12 weeks - well below the 30 to 40% of the pre-triazole era, but flat across the decade (no difference between 2016-2020 and 2021-2025). Contributable mortality kept climbing (40.0% then 48.2%), and persistent neutropenia was the only independent predictor of early attributable death (subdistribution HR 2.59). The authors' reading: better drugs have stopped moving the needle, and further gains depend on host recovery, not drug choice.15
The strategy question stays formally open. IDSA frames the optimal empiric-versus-pre-emptive choice as still evolving; ECIL declined to prefer pre-emptive over empiric because the first RCT showed excess fungal disease in the pre-emptive arm. Practically: pre-emptive therapy suits lower-risk patients with reliable diagnostics and a stewardship goal, while empiric therapy remains the safer default for the highest-risk induction patient, where galactomannan is least dependable and the newer molecular tests are not yet everywhere.16,17
Meta-analysis reframed fluoroquinolone prophylaxis from fewer fevers to fewer deaths. Across 95 trials, prophylaxis reduced all-cause mortality versus placebo (RR 0.67, 95% CI 0.55-0.81), with the fluoroquinolone subgroup at RR 0.52 (0.35-0.77). The 2012 Cochrane update (109 trials, 13,579 patients) preserved the all-cause mortality benefit (RR 0.66, 0.55-0.79), number-needed-to-treat 34, and found quinolones caused fewer discontinuations and less post-treatment resistance than trimethoprim-sulfamethoxazole.18,19
The individual RCTs did not show a survival benefit. In high-risk patients (GIMEMA, N=760), levofloxacin cut fever lasting the duration of neutropenia (65% vs 85%) but mortality was unchanged. In lower-risk solid-tumor and lymphoma patients (SIGNIFICANT, N=1565), levofloxacin reduced first-cycle febrile episodes (3.5% vs 7.9%), but severe infection (1.0% vs 2.0%, P=0.15) and deaths (4 vs 4) were unchanged. That is the basis for restricting prophylaxis to high-risk, prolonged (over 7-day) profound neutropenia rather than every chemotherapy patient.20,21
Modern data test the other direction - what happens when you stop. A 2026 meta-analysis (10 before-after studies, 2363 HSCT recipients) found that discontinuing quinolone prophylaxis raised bloodstream infections (the pooled risk ratio for prophylaxis versus none was 0.69, 95% CI 0.54-0.87) and gram-negative bloodstream infections in particular (RR 0.49, 0.33-0.74), but did NOT significantly raise infection-related mortality (RR 1.26, 0.71-2.23). It also priced the cost of continuing: prophylaxis selected for resistance, with more quinolone-resistant (RR 2.35, 1.68-3.29) and carbapenem-resistant (RR 5.32, 1.08-26.16) gram-negative bloodstream infections. Stopping trades more infections for fewer resistant ones, with no measurable mortality penalty in these observational cohorts.22
Each RCT is powered for febrile episodes, not death; deaths are too few per trial to move a mortality endpoint. Pooling about 100 trials aggregates enough events to detect a difference. The catch: that benefit is an average across heterogeneous, mostly pre-2005 trials with different resistance ecologies. It need not transfer to a unit with high fluoroquinolone resistance now - which is precisely the stewardship objection, and what the 2026 discontinuation data quantify.19,22
Fluconazole established azole prophylaxis (bone-marrow transplant, N=356): systemic fungal infection fell to 2.8% from 15.8% and fungal deaths dropped, but overall mortality was unchanged - and fluconazole covers no molds. Posaconazole closed the mold gap. In neutropenic AML/MDS (N=602), posaconazole cut proven or probable fungal infection to 2% from 8% and invasive aspergillosis to 1% from 7%, and uniquely improved overall survival (P=0.04), at the cost of more treatment-related serious adverse events (6% vs 2%).23,24
The graft-versus-host disease trial is the teaching case on endpoints. Posaconazole versus fluconazole (N=600) did not meet its primary endpoint of all invasive fungal infections (5.3% vs 9.0%; OR 0.56, 95% CI 0.30-1.07; P=0.07). It won on a secondary endpoint - invasive aspergillosis (2.3% vs 7.0%; OR 0.31, 0.13-0.75; P=0.006) - and on fewer fungal deaths (1% vs 4%; P=0.046), with overall mortality similar.25
Guidelines convert these trials into graded advice. ECIL-6 makes posaconazole the drug of choice for AML or MDS remission-induction (grade A-I) and for high-risk graft-versus-host disease post-engraftment (A-I); fluconazole keeps a role (B-I) only where mold incidence is low and only alongside a mold-directed diagnostic strategy. Two caveats matter at the bedside: graft-versus-host disease by itself is not an indication for mold-active prophylaxis, and in acute lymphoblastic leukemia the mold-active azoles are avoided because they block vincristine metabolism and cause severe neurotoxicity.26
| Agent / setting | Fungal-infection result | Mortality | Note |
|---|---|---|---|
| Fluconazole / BMT (Goodman) | Systemic FI 2.8% vs 15.8% | No overall benefit | No mold coverage |
| Posaconazole / AML-MDS (Cornely) | IFI 2% vs 8%; IA 1% vs 7% | Overall survival benefit (P=0.04) | More serious drug-related AEs |
| Posaconazole / GVHD (Ullmann) | All-IFI 5.3% vs 9.0% (P=0.07); IA 2.3% vs 7% | Overall similar; fewer fungal deaths | Primary endpoint not met |
The GVHD trial's pre-specified primary endpoint (all IFI) was not significant at P=0.07. The practice-changing signal came from a secondary endpoint (aspergillosis) and a cause-specific mortality endpoint. A drug can be worth using on a secondary endpoint, but the strength of evidence is graded by the endpoint pre-specified as primary - calling a trial positive when its primary missed is a common misread.25
Trimethoprim-sulfamethoxazole reduces Pneumocystis pneumonia by about 85% in non-HIV immunocompromised patients (RR 0.15, 95% CI 0.04-0.62) and PCP-related death (RR 0.17), number-needed-to-treat 19 - but with no all-cause mortality change versus placebo. Because relative efficacy is enormous, the decision is a risk-threshold decision: the trials' benefit was anchored to roughly a 6.2% control-arm PCP incidence. Prophylaxis is warranted when a patient's baseline risk approaches that range (allogeneic HCT, prolonged high-dose corticosteroids, certain T-cell-depleting or targeted agents, acute lymphoblastic leukemia), not as a blanket.27
With a fixed ~85% relative reduction, the number-needed-to-treat scales inversely with baseline risk. At the trials' ~6% incidence, NNT is about 19; at 1% incidence it climbs past 100 and the toxicity and cost balance tips against prophylaxis. So PJP prophylaxis is decided by whether a given patient clears an incidence threshold, not by whether the drug works - it clearly works.27
In CMV-seropositive allogeneic HCT recipients (N=565), letermovir prophylaxis cut clinically significant CMV infection through week 24 to 37.5% from 60.6% (P<0.001). All-cause mortality at week 48 was 20.9% with letermovir versus 25.5% with placebo - reported as two percentages with no P value or confidence interval in the primary trial, so a signal rather than a demonstrated survival benefit. Letermovir has no HSV or VZV activity, so acyclovir or valacyclovir must continue alongside it.28
The reflex to continue broad-spectrum antibiotics until neutrophil recovery is giving way to response-adapted discontinuation, and the guidelines sit on a fault line. IDSA continues empiric therapy to marrow recovery; the European (ECIL) and AGIHO fever-of-unknown-origin guidance endorses stopping on clinical criteria irrespective of the neutrophil count.16,29
In the How Long RCT (N=157 high-risk hematology patients with fever of unknown origin), stopping empiric therapy after at least 72 hours afebrile plus clinical recovery, irrespective of neutrophil count, produced more antibiotic-free days (mean 16.1 vs 13.6; difference -2.4, 95% CI -4.6 to -0.3; P=0.026) and was safe (1 death vs 3).30
Retrospective cohorts point the same way. SAFE (N=575 AML/MDS induction) compared a 3-day stop with stop-at-recovery: median broad-spectrum days fell to 9 from 19, and a serious complication (death or ICU by day 30) occurred in 12.5% versus 8.9% (HR 1.36, 95% CI 0.77-2.41 - not significantly worse). An allogeneic-HSCT cohort that de-escalated after at least 5 days and defervescence, versus continuing to engraftment, saw recurrent fever in 15% versus 19% with no difference in ICU stay, length of stay, or mortality. A 2025 meta-analysis (10 studies, 8 reporting mortality) found early de-escalation associated with lower mortality (OR 0.20, 95% CI 0.06-0.69).31–33
Two competing stop rules: discontinue when the marrow recovers (ANC at least 500) versus discontinue when the patient is clinically well (afebrile at least 48 to 72 hours, hemodynamically stable, cultures negative) regardless of count. The randomized evidence shows the clinical rule is safe and saves antibiotic-days; the count rule prolongs exposure with no demonstrated survival gain. The genuine residual debate is the still-profoundly-neutropenic patient with fever of unknown origin, where guidelines still differ on stopping before recovery.30,31
Defervescence shows the current fire is out. Stability throughout the episode is a proxy for the absence of an occult severe infection that a brief afebrile window can mask. The discontinuation trials selected patients who were never unstable; applying their result to someone who was transiently in shock extrapolates beyond the evidence. The eligibility phrase that carries the weight is hemodynamically stable since presentation, not the afebrile window alone.30
Many neutropenia pathways assume extended-infusion beta-lactams. The pharmacology is sound - beta-lactam killing is time-dependent, so a longer infusion raises the fraction of the dosing interval above the MIC - but the one febrile-neutropenia RCT is a cautionary tale about surrogate endpoints.
In BEATLE (N=150), extended versus intermittent infusion of an anti-pseudomonal beta-lactam (agent at physician discretion) for empiric therapy gave treatment success at day 5 (defervescence without a change of therapy) of 50.6% with extended versus 63.0% with intermittent infusion (risk difference -12.4%, 95% CI -29.4 to +4.7; P=0.17) - no benefit, and numerically favoring intermittent. Extended infusion did hit the 75% and 100% fT>MIC targets more often, with no difference in adverse events or 30-day mortality. The authors do not support routine empiric extended infusion in neutropenia.34
The disconnect is the teaching point: extended infusion improved the pharmacokinetic surrogate but not the clinical endpoint. In a mostly fever-of-unknown-origin population where a susceptible target organism often does not exist, hitting fT>MIC need not move outcomes. BEATLE was open-label, used a day-5 endpoint, and pooled different beta-lactams; the authors themselves point to the sicker, microbiologically-documented, higher-MIC subgroup as where extended infusion should be studied next.34
fT>MIC, the time free drug stays above the MIC, predicts bacterial kill in pharmacokinetic models. It is a means, not an end: when most empiric courses treat fever without a proven pathogen, a regimen can win on target attainment and still tie on defervescence - which is what BEATLE observed.34
Most neutropenia pathways keep cefepime first-line and route resistant organisms to Infectious Diseases. The case for reaching further empirically is a local-epidemiology judgment, not a universal one.
Bloodstream epidemiology frames the empiric target. A 2025 meta-analysis of 8665 positive cultures in chemotherapy-induced neutropenia found gram-negatives predominate over gram-positives, with fungi low (random-effects pooled proportions 59% / 40% / 2.5%; the same review's crude tally runs closer to 46% / 42% / 3%). Among gram-negatives, Escherichia coli (about 46%), Pseudomonas aeruginosa (about 19%), and Klebsiella pneumoniae (about 18%) lead - which is why the empiric backbone is an anti-pseudomonal gram-negative agent.35
Whether to reach past cefepime has been tested. Against standard of care (86% cefepime), imipenem/cilastatin/relebactam gave a higher favorable clinical response at end-of-IV therapy (90% vs 74%, P=0.04), but the arms converged at test-of-cure and late follow-up, with comparable eradication and safety. The separation was a small, single-center, early-endpoint effect that did not persist - not a mandate to broaden coverage for everyone.36
Colonization data sharpen the local-epidemiology point. In a 2026 cohort (87 patients, 170 neutropenic-fever episodes) where 21% carried OXA-48-producing Enterobacterales, 31% of carriers developed an OXA-48 infection during a febrile episode versus none of the non-carriers, and a recent negative rectal swab had a 100% negative predictive value. OXA-48 infection carried 27% versus 3% 15-day mortality, and only 47% of these infections received active therapy within 24 hours. The lesson is targeted, not blanket: escalate empirically to an OXA-48-active agent in a colonized patient who presents with sepsis, and de-escalate on a recent negative swab - colonization alone did not predict death, and two-thirds of carriers never developed infection.37
So reaching past cefepime turns on local ESBL, carbapenem-resistant Enterobacterales, and resistant-Pseudomonas prevalence and the individual patient's colonization history - verify against your own antibiogram and surveillance, not a headline trial.
Convergence at test-of-cure means the comparator caught up: an early separation can reflect faster defervescence without a durable outcome difference. Judge empiric agents on the end-of-treatment and follow-up endpoints, not the first read.36
Outpatient oral therapy is a high-stakes disposition, gated on MASCC or CISNE plus clinical judgment and documented Hematology-Oncology sign-off. Both the scores and the oral-switch evidence deserve a fellow's skepticism.
Every febrile-neutropenia prediction model has a weak evidence base. A 2025 systematic review of 90 studies graded all of them high risk of bias by PROBAST, with limited external validation. The discrimination is modest in concrete terms: pooled sensitivity and specificity are about 55.6% and 86.0% for MASCC and 78.9% and 64.9% for CISNE. The scores support clinical judgment; they do not replace it.38
The oral-switch evidence is thinner than its adoption suggests. EASI-SWITCH randomized low-risk patients (MASCC at least 21) to early oral ciprofloxacin plus co-amoxiclav versus continued IV therapy, but closed early for under-recruitment (N=129), so non-inferiority was not established against its 15% margin (itself widened from 10% mid-trial). Treatment failure was 24.6% versus 14.1% (intention-to-treat) and 17.7% versus 13.3% (per-protocol), driven by fever recurrence and physician-directed escalation - with no ICU admissions or deaths, a shorter length of stay, and a clear patient preference for the oral pathway.39
For agent choice, a pediatric cost-effectiveness RCT found oral levofloxacin dominant (lower cost, marginally higher quality-adjusted life-years) over amoxicillin-clavulanate plus ciprofloxacin in low-risk disease - hypothesis-generating for adults, given the single-country pediatric setting. The ASCO/IDSA guideline anchors the disposition: candidates are low-risk by MASCC or CISNE plus clinical criteria, and the recommended oral regimen is a fluoroquinolone plus amoxicillin-clavulanate (or clindamycin for penicillin allergy).40,41
The outpatient decision rests on tools that are, by their own evidence base, weakly discriminating - which is why guidelines pair them with clinical judgment and a documented sign-off rather than a score alone. Early oral switch trades a shorter admission for a higher, but low-acuity, failure and re-admission rate: a values-sensitive choice for a selected, adherent, well-supported patient.38
EASI-SWITCH's treatment failure bundled common, soft events (fever recurrence, physician-directed escalation) with rare, hard ones (ICU, death). Because the hard events were absent, a double-digit failure rate can look alarming while nothing serious happened. Unpack a composite into its components before believing the headline number.39