WISCA has been previously used to study antibiotic coverages in different adult and paediatric infectious syndromes [12, 13, 15,16,17]. To our knowledge, this is the first study developing a WISCA model to guide the empirical choice of the most suitable antibiotic empiric therapy in onco-haematological paediatric patients. The secondary immunodeficiency and the high exposure of patients to previous antibiotic treatments pose a challenge for severe infections, possibly due to MDR organisms, making this population a unique epidemiological setting where empirical therapies need to be optimized [18, 19].

Genoa collected many more blood cultures compared to Padua, this is probably due to the different diagnostic strategies for children presenting with FN. In particular, Genoa provided more solid tumours in its population. As the epidemiology of tumours does not change in the same region, we believe this is the consequence of an increased diagnostic capacity, especially for catheter-related infections, caused by CoNS, which are the majority of episodes in solid tumours children in Genoa centre.

The primary objective of this study was to assess the ability of a Bayesian WISCA model to estimate antibiotic treatment appropriateness, which is not intended as 100% coverage, but as an acceptable compromise between pathogen coverage and responsible antimicrobial prescribing practice. Our study showed that, despite recommendations [1, 2], none of the monotherapies offered an adequate coverage rate for the identified pathogens; indeed, both centres are not currently using monotherapies to manage FN. However, combination therapies considerably increased the median coverage rates. According to the principle that led to the strategy of empirical therapy (early treatment of Gram-negative bacteraemia to reduce mortality [4]) the key to reaching the optimal coverage rate was the association of an anti-pseudomonas molecule with a second gram-negative agent, as amikacin. The association with a glycopeptide further increased the coverage rate. In our settings this may have a limited clinical relevance, but it is probably the result of the high incidence of methicillin-resistant CoNS (Table 1), while methicillin-resistant S. aureus was found only once. However, the glycopeptide-containing combination could be useful in patients colonized by methicillin-resistant S. aureus that present a significantly higher risk of S. aureus bacteraemia that would not be adequately treated by anti-Gram-negative antibiotics (such as PI-TZ) [3]. Teicoplanin (in all combinations) performed worse than vancomycin. This is due to the limited number of CoNS tested for teicoplanin, which is not routinary tested.

Our study confirms the reliability of the combination of PI–TZ–amikacin, resulting in a median coverage of 98% when focusing exclusively on gram-negative bacteria. Considering the entire pool of isolated bacteria, including both gram-positive and gram-negative pathogens, PI–TZ–amikacin provides a coverage of 78%, which increases to 97% after the addition of vancomycin.

Uncertainty intervals, overall, were quite large, due to the small sample size.

In certain situations, such as certain combinations (e.g., ciprofloxacin-amikacin), they appear to perform worse than monotherapy (ciprofloxacin alone). This results in a decrease in sample size when considering cultures with both drugs tested.

Paediatric BSI is uncommon in onco-haematological patients [20], and data are limited even in a six-year, bicentric study. The small-sample limitation is a known issue when applying WISCA: in a previous study, Bielicki et al. in 2016 used WISCA to evaluate five empirical antibiotic regimens for paediatric BSI, using pooled data from 19 hospitals to overcome sample-narrowness. Although statistical significance was achieved, the results could not be generalized as the epidemiology of BSI was not overlapping between centres [12]. This highlights the pathogen temporal and geographical variability as intrinsic characteristics of infectious syndrome epidemiology, limiting the utility of applying data gathered from heterogeneous cohorts, but confirms the recommendation of antibiotic choices based on local epidemiological data [1, 2].

Another way to reach significance when evaluating antibiotic coverage is to limit the analysis to a few regimens and a few, most prevalent, pathogen isolates. This approach has been successfully used to study coverage of third-generation cephalosporins and meropenem toward causative pathogens of paediatric BSI from 23 countries [17]; however, these results cannot be applied to a local level to drive empiric prescribing, with limited clinical usefulness. Indeed, for local adaptation, it would be necessary to identify that local pathogen epidemiology and susceptibility patterns are homogeneous to the pooled data from where WISCA was calculated.

The analysis in this study was not restricted, including even once-found bacteria (which will, however, “weight less”), as onco-haematological patients are a high-risk population, and empiric coverage must consider even uncommon aetiologies, as sepsis in neutropenic patients may be a life-threatening event. The same global approach has been used to study critical care infections in adults by Randhawa et al. [15]. However, in neutropenic patients, infections with Gram-negative bacteria have a poorer outcome and are associated with increased adverse events [21]. Thus, this study developed a second WISCA model including only Gram-negative bacteria, considering the possibility of targeting Gram-positive bacteria (as CoNS) in clinically stable patients when cultures turn back positive. This strategy is currently adopted in Genoa Hospital. The restriction of glycopeptide use in paediatric cancer patients with FN resulted safe in an observational study [22].

Further, we decided not to restrict the analysis to a few antimicrobial regimens, as drug prescribing in onco-haematological patients is often challenging due to possible organ dysfunction, allergies, drug-to-drug interactions, fluid overload, and venous catheter incompatibilities. We then decided to analyse all monotherapies and combinations potentially used, considering even those antibiotics that are not considered first-line molecules in children, such as ciprofloxacin.

Another limitation of the WISCA application in our study is the lack of correlation between infective pathogens and infection outcomes (e.g., mortality, PICU admission). Those data would allow a WISCA model to estimate regimens with expected maximum clinical concordance and, therefore, the most significant potential impact. This strategy has not yet been adopted in the reported WISCA study.

Lastly, when applying empiric antibiotic therapy, an “acceptable” cut-off of coverage is usually self-estimated according to clinicians’ and microbiologists’ experience. We used an 85% estimated median coverage rate to define a regimen “appropriate”, which overlaps with the study by Randhawa et al. on critical care infections. However, many clinicians could have a preference for antibiotic regimens perceived to have a coverage of 90% or more [23].