Prediction of candidemia with machine learning techniques: state of the art

Abstract

In this narrative review, we discuss studies assessing the use of machine learning (ML) models for the early diagnosis of candidemia, focusing on employed models and the related implications. There are currently few studies evaluating ML techniques for the early diagnosis of candidemia as a prediction task based on clinical and laboratory features. The use of ML tools holds promise to provide highly accurate and real-time support to clinicians for relevant therapeutic decisions at the bedside of patients with suspected candidemia. However, further research is needed in terms of sample size, data quality, recognition of biases and interpretation of model outputs by clinicians to better understand if and how these techniques could be safely adopted in daily clinical practice.

Plain language summary

Candida is a type of fungus that can cause fatal infections. To confirm the presence of the infection, doctors may search for the fungus in the blood. Here, we discuss if computer systems can help to identify infection more easily and more rapidly.

Plain language summary

Executive summary.

• Bloodstream infections (BSI) by Candida spp. are among the most frequent healthcare-associated BSI, and are associated with high mortality rates.

• In the past few years some authors have started to explore the potential role of machine learning (ML) models for the prediction of candidemia, as early diagnostic tools in the form of ML-guided clinical decision support systems.

Brief summary of main results of studies assessing the use of ML techniques for the early diagnosis of candidemia

• There are currently only a few studies evaluating the approach through ML techniques for the early diagnosis of candidemia in clinical practice.

Study design & populations

• All the four studies retrieved by our literature search were retrospective and investigated the diagnostic performance of various ML algorithms for the prediction of candidemia in adult patients.

Sample sizes & choice & availability of input features

• Various factors affect the pre-test probability and post-test probability of candidemia (where the test is the evaluated ML-based classifier) and may lead to very different diagnostic performances when extrapolated to settings different from those registered in the original studies.

Feature selection, model selection & performance evaluation

• Exploration of the performance of complex ML models in predicting candidemia on very large samples of automatically extracted laboratory and microbiological data could help, in our opinion, to better understand the potential of such models in identifying hidden patterns of associations that are difficult to be recognized by clinicians due to their apparently nonspecific nature.

• The best performances for the early diagnosis of candidemia (as a prediction/classification task) within each retrieved study were from models in the category of ensemble learning methods.

The future of ML techniques for diagnosis of candidemia

• The use of ML tools hold promise to provide highly accurate and real-time support to clinicians for relevant therapeutic decisions at the bedside of patients with suspected candidemia, although due caution is still necessary in terms of extrapolation and generalization.

• An accurate, transparent, and large dataset is at least as important as the choice of the best suited ML model for training.

• Further research is needed in terms of sample size, data quality, recognition of biases and interpretation of model outputs by clinicians to better understand if and how these techniques could be safely adopted in the daily clinical practice.

Bloodstream infections (BSI) by Candida spp. are among the most frequent healthcare-associated BSI, and are associated with mortality rates possibly surpassing 50% when they develop in critically ill patients in intensive care units (ICUs) [^1-4]. The development of candidemia has also been reported to unfavorably impact length of hospital stay, and to increase costs in healthcare [1].

Diagnosis of candidemia is usually made through collection of blood for culture, the results of which become available only 48–72 h after blood draw. Therefore, while blood cultures currently remain the gold standard for diagnosis (and also allow for species identification and susceptibility testing), more rapid antigen-based or molecular tests are also nowadays commonly used in clinical practice for helping physicians with early therapeutic decision, in other words, whether or not to add an antifungal, or to discontinue empirical antifungals, in patients with signs and symptoms of BSI and severe clinical presentation, while waiting blood culture results [^5-14].

In the past few years, some authors have also started to explore the potential role of machine learning (ML) models for the prediction of candidemia, as early diagnostic tools in the form of ML-guided clinical decision support systems. In the present narrative review, we discuss currently available studies assessing the use of ML models for the early diagnosis of candidemia, focusing on employed models and the related implications. Finally, we provide a personal view on some crucial points that should be considered for the future inclusion of these techniques in routine diagnostic algorithms and clinical reasoning at the bedside of patients with suspected candidemia.

To identify available studies dealing with the use of ML techniques to support the diagnosis of candidemia in clinical practice, we performed a literature search. On 2 August 2023, we searched the PubMed and Embase databases using the following combination of keywords: ((candidemia) OR (candidaemia)) AND (machine learning). We did not apply any filter on time, type of study and text availability. An update of the search, employing the same combinations of key words, was performed on 28 November 2023. Overall, four original studies were eventually selected for inclusion in the present review [^15-18]. The main characteristics of included studies are reported in Table 1. Finally, the text of the present review was structured in the following sections, reflecting crucial aspects of the use of ML techniques for the early diagnosis of candidemia: brief summary of main results of studies assessing the use of ML techniques for the early diagnosis of candidemia; considerations on study design and population; considerations on sample sizes and on choice and availability of input features; considerations on feature selection, model selection and performance evaluation. The manuscript was then completed with the following two sections: conclusion; future perspective.

Table 1.

Summary of the principal characteristics of included studies.

Study (year)	Objectives	End point	Study period	Strengths	Limitations	Ref.
Giacobbe et al. (2023)	Differential diagnosis of candidemia vs bacteremia	Candidemia, defined as presence of one or more positive blood cultures for a species of Candida spp. within a 30 days window	2011–2019	Automation in features extraction from hospital LIS; identification of nonspecific markers	Based on laboratory tests and microbiological cultures only; single center study; large proportion of missing values	[18]
Ripoli et al., (2020)	Detection of candidemia in patients hospitalized in IMW and experiencing fever and/or other clinical signs of infection (SIRS)	Candidemia, defined as a patient with at least one blood culture yielding Candida spp.	2012–2015	Attempt to explain model outputs to clinicians	Limited sample; case-control study; only one ML-model considered	[17]
Yuan et al. (2021)	Prediction of candidemia in ICU patients with new onset of SIRS	Candidemia, defined as a positive blood culture for Candida spp. obtained during SIRS only in absence of previous SIRS within 24 h	2013–2017	Limited number of positive cases managed with SMOTE; comparison of five different algorithms (four ML models + LR)	Results not generalizable to non-ICU patients; case-control study	[15]
Yoo et al. (2021)	Diagnosis of candidemia in patients with cancer	Candidemia, defined as positive culture for any Candida spp. from more than one blood sample within a 7 days window	2010–2018	Specific for patients with malignancy (which nonetheless could also be viewed as a limitation for generalization)	Single center study; non-standard definition of candidemia; the developed algorithm only works with none or few missing values	[16]

ICU: Intensive care unit; IMW: Internal medical ward; LIS: Laboratory information system; LR: Logistic regression; ML: Machine learning; SIRS: Systemic inflammatory response syndrome; SMOTE: Synthetic minority over-sampling technique.

Brief summary of main results of studies assessing the use of ML techniques for the early diagnosis of candidemia

In 2020, Ripoli and colleagues reported the results of a retrospective, multicenter study assessing the performance of a random forest (RF) algorithm for the early differential diagnosis between candidemia and bacteremia in internal medicine wards [17]. The study population consisted of 295 patients with candidemia and 138 controls with bacteremia. Eventually, the RF algorithm showed 84% sensitivity and 91% specificity, which were higher than those obtained through a classical multivariable logistic regression model (80% sensitivity and 85% specificity) [17].

In 2021, Yuan and colleagues also assessed the performance of four different ML algorithms (XGBoost, support vector machine, RF and ExtraTrees) and logistic regression for the early diagnosis of candidemia in ICUs patients with systemic inflammatory response syndrome (SIRS) [15]. The study was multicenter and retrospective, and included a total of 137 patients with SIRS and candidemia and 7795 patients with SIRS and blood cultures negative or positive for a pathogen other than Candida species. The ML algorithm showing the best predictive performance in the study population was XGBoost, with 84% sensitivity and 89% specificity. The model also showed a very high negative predictive value (NPV) of 99.6%, partly connected to the low prevalence of candidemia in the study population (1.7%) [15].

In the same year, Yoo and colleagues reported the results of a retrospective, single center study assessing the performance of various ML algorithms (including, among others, deep neural network, gradient boosting and RF) and logistic regression for the early diagnosis of candidemia in patients with malignancies [16]. The study sample consisted of 501 candidemia episodes and 2000 control episodes (either bacteremia episodes or negative blood cultures). The ML algorithm showing the best predictive performance in the study population was RF, with 89% sensitivity and 90% specificity. The model also showed 96.7% NPV, in a study population with a prevalence of candidemia of 20% [16].

Finally, we recently assessed the performance of an RF algorithm for the early differential diagnosis between candidemia and bacteremia on an automatically extracted dataset of 1275 episodes of candidemia and 11,208 episodes of bacteremia over 9 years (10% prevalence of candidemia) [18]. The study was retrospective and exploited automated data collection of laboratory and microbiological features through a previously validated automated extraction system [19]. In the overall study population, the performance of the RF algorithm in predicting candidemia on the basis on microbiological features and nonspecific laboratory markers resulted in 98% sensitivity and 65% specificity in the training set, and in 74% sensitivity and 57% specificity in the test set (outperforming predictions of penalized logistic regression models using the same training and test sets). When these performances were used as pre-training for selecting the most influential features to then predict candidemia in a subset of episodes for which the results of more specific laboratory markers for the diagnosis of candidemia (serum β-D-glucan [BDG] and serum procalcitonin [PCT]) were available, the diagnostic performance of an RF model based on the selected most influential features plus BDG and PCT numerically outperformed the performance of a RF model based on BDG and PCT only, although the difference was eventually not statistically significant [18].

Study design & populations

All the four studies retrieved by our literature search were retrospective and investigated the diagnostic performance of various ML algorithms for the prediction of candidemia in adult patients [^15-18]. As shown in Table 2, three studies were restricted to specific wards or baseline conditions (internal medicine wards, ICUs and patients with malignancies, respectively), whereas in our previous study we considered all hospitalized patients independent of wards and conditions [^15-18]. It is noteworthy that the population selection has implications for the baseline prevalence of candidemia, which is generally low, but with notable fluctuations according to the baseline risk of the given population, in turn influencing both NPV and positive predictive value (PPV) of explored models. Other factors affecting the original baseline prevalence of candidemia (in a way that may lead to artificial increases) are whether all patients without candidemia meeting the study entry criteria are included or only a subset of them is included as a control group (like in the studies by Ripoli and colleagues and by Yoo and colleagues [16,17]), and whether baseline prevalences were intentionally changed to avoid difficulties in model training (like in the study by Yoo and colleagues [16]). While similar adjustments are not incorrect and may help to evaluate the model performance more adequately (as well as to reduce the time of data collection when performed manually), readers should always pay caution when extrapolating NPV and PPV to their own settings, as their baseline prevalence of candidemia is likely different from that of samples adjusted for model training. Even more important clinically, when extrapolating results to their own setting, readers should pay careful attention to the type of population on which the model was trained. Indeed, both the baseline prevalence of candidemia and the diagnostic performance of markers (for example, that of inflammatory markers in populations frequently treated with potent anti-inflammatory drugs versus populations usually not receiving anti-inflammatory drugs) may vary greatly across, for example, ICUs, internal medicine wards, patients with malignances and whole hospital populations. All these factors variously affect the pre-test probability and post-test probability of candidemia (where the ‘test’ is the evaluated ML-based classifier) and may lead to very different diagnostic performances locally from those registered in original studies. Notably, these considerations do not only apply only to ML algorithms but also to classical diagnostic studies, and highlight the importance and need for external validation of trained models (both in different centers and in different populations). With regard to specific considerations on study design of future multicenter studies evaluating the performance of ML-based classifiers for the early diagnosis of candidemia, it is of note that also the type and characteristics of the information used for training the model could possibly and unintentionally influence and confound the assessment of the performance of the model. For example, if also all (or more likely, part) of data is collected from the text of clinical notes through ML techniques (for example, natural language processing [NLP]), between-center heterogeneity in how and how much information is presented in the text of clinical charts may lead to different accuracy in the extraction of the information, in turn influencing the performance of the model assessing the impact of extracted features on the endpoint of interest (early diagnosis of candidemia). In our opinion, measurement of such a potential heterogeneity in the ability of information extraction across center should also be defined at the time of study design, and may represent an important strategy to understand both how much it could influence model results and how to reduce it to improve generalization.

Table 2.

Study populations of included studies.

Study (year)	Population	Candidemia (no.)	No candidemia (no.)	Case-control matching	Ref.
Giacobbe et al. (2023)	Mixed-ward patients	1275	11,208†	-	[18]
Ripoli et al. (2020)	Internal medical wards patients	157	138†	1:1	[17]
Yuan et al. (2021)	Intensive care units patients	137	7865‡	-	[15]
Yoo et al. (2021)	Patients with malignancy	501	2000‡	1:4 in four subsets	[16]

^†Bacteremia episodes.

^‡Either bacteremia or negative blood cultures.

Sample sizes & choice & availability of input features

Clinical data included in observational studies mainly derive from procedures concerning patient care, that are not primarily aimed at research. This is classically defined as real world experience and comprehends all data collected during daily clinical practice. Nowadays, the richest sources of patient's data are laboratory information systems (LISs) and electronic medical records (EMRs), that contain multidimensional and heterogeneous data, usually collected manually by investigators to be included in specific datasets specifically designed for a given research project. Manual collection is extremely time consuming, so samples sizes are usually limited to hundreds or sometime a few thousands of patients. This is in line with the sample sizes of three of the retrieved studies (Table 2), whereas in our previous study, in which the data was not collected manually but through an automated extraction system, the final number of included episodes (either of candidemia or bacteremia) was far higher, reaching 12,483 (a number very difficult to achieve in general, and likely impossible to achieve for a single-center study by means of manual collection, at least in a short time window) [^15-18]. On the other hand, in our study, only laboratory and microbiological features were collected by the automated extraction system, whereas in the other studies also clinical features that are well recognized as predictors of candidemia (for example, presence of central venous catheters, total parenteral nutrition, receipt of broad spectrum antibiotics) were collected manually [^15-18]. This reflects the current difficulties in automatically and reliably extracting high-detail and accurate clinical features from unstructured data (for example, free text from clinical notes in EMRs), although advancements in the field are likely to be achieved in the near futures by exploiting progresses in the application of NLP techniques to the extraction of clinical features from EMRs [^20-23]. In the meantime, in our opinion, exploration of the performance of complex ML models in predicting candidemia on very large samples of automatically extracted laboratory and microbiological data can help to better understand the potential of such models in identifying hidden patterns of associations that are difficult to be recognized by clinicians at the bedside due to their apparently nonspecific nature. In turn, such results could be exploited for improving the performance of real-time ML-based clinical decision support systems for the early diagnosis of candidemia, provided interpretability or sufficient explainability of employed models is guaranteed (see future perspective below).

Feature selection, model selection & performance evaluation

Before proceeding with training and evaluation of the performance of ML models, descriptive univariable comparisons, for example, through chi-square test and Kruskal–Wallis test, were performed in all the included studies in order to spot possible associations between each variable and the outcome measure of interest (diagnosis of candidemia), eventually leading to feature selection for inclusion in ML models [^15-18]. The number of features screened, the methods and processes used for feature selection, and the number and type of features eventually included in the best performing models are summarized graphically in Figure 1. In line with the design of the study, in our previous experience we selected only features derived from laboratory tests and microbiological cultures results, since the employed dataset included automatically collected data from the LIS and not from EMRs [18]. Conversely, in the other included studies, feature selection was based also on demographics and clinical data (for example, comorbidities, previous antibiotic treatments, total parenteral nutrition) [^15-17]. In our opinion, such different designs in these early studies attempting to explore how ML-based classifiers could help in the early diagnosis of candidemia reflect both the current lack of standardized approaches and the need to adhere to local possibility of extraction/collection in terms of type and quantity of data. Nonetheless both these considerations should not be necessarily viewed as limitations. Indeed, in our opinion, they could also help to discover which approach (or part of it) could work better in a hypothesis-generating way, prompting further study of specific aspects about how to improve both training and performance of ML models for this specific task. Certainly, caution is still necessary in extrapolating preliminary evidences, especially considering the limited number of studies (and consequently models and choices regarding how to train them) currently available as retrieved by our literature search restricted to diagnosis (while other studies are available that explore the use of ML-based classifiers for the prognosis of candidemia, which nonetheless represent a different endpoint [^24-26]).

Figure 1.

Feature selection process in the different studies. 

*Comprehensive of higher variable importance, higher clinical importance, better extractability and less missing values.

BDG: Serum β-D-glucan; PCT: Serum procalcitonin; PFI: Permutation feature importance. 

Details regarding model selection and performance evaluation of ML algorithms in the different studies are reported in Table 3. Notably, all included studies basically compared the performances of several ML models (plus logistic regression in most cases) for the classification task ‘candidemia episode yes/no’ [^15-18]. For each of the selected models, a first subset of the study sample (the training set) was used to identify the best set of parameters for each model. Then, the predictive performance of models trained on the training set was evaluated on the remaining subsample (the test set), using several performance metrics. In the study by Yuan and colleagues, the authors employed the synthetic minority over-sampling technique (SMOTE) algorithm to deal with the imbalance of sample categories [15]. In our previous study, no augmentation, oversampling or undersampling techniques were adopted but rather the training and test sets had a ground truth distribution similar to that of the available dataset, which is a relevant point to be considered for actually evaluating and creating operational ML models [18]. Finally, it should be highlighted that all ML models eventually showing the best performances for the early diagnosis of candidemia (as a prediction/classification task) within each study belong to the category of ensemble learning methods. Specifically, they are a subclass of supervised learning algorithms whose final prediction is obtained as a weighted average of the predictions returned by a set of devised classifiers (for example, RF is composed by K decision trees) [27]. The impact of biases on model performances specific to ML-based classifiers requires particular attention. In particular, the frequent need to fine-tune hyperparameters (in order to achieve better model performances) once first results are available is apparently in contrast with the classical approach of developing a statistical analysis plan before study conduction. This is also not necessarily a disadvantage, in our opinion, since fine-tuning in this case can also be conceived a priori, as a predefined measure with a certain degree of flexibility. What we think would become fundamental is, for each specific topic to achieve standardization of the desired bias-variance trade-off (for example, for the early diagnosis of candidemia, we may propose as a possible solution to focus the fine-tuning of hyperparameters on selecting a non-overfitted model prioritizing sensitivity over specificity in order not to miss true cases of candidemia, that otherwise would encounter perilous delays in antifungal treatment, still improving specificity compared to standard of care or in addition to standard of care). Of course, other and possibly better solutions may exist and deserve investigation, nonetheless still conceptually highlighting the need to tailoring each specific approach to the pertinent clinical research question.

Table 3.

Summary of the information related to ML models extracted from the included studies.

Study (year)	Technical details	ML models	Output metrics	Best performances	Software (packages) used	Ref.
Giacobbe et al. (2023)	(Stratified) train test split: 70:30; (stratified) tenfold cross-validation; parameters tuning with GridSearch	RF	Sensitivity, specificity, PPV, NPV, Accuracy, F1 score, TSS	RF: sensitivity† = 0.566, Specificity† = 0.739	Python	[18]
Ripoli et al. (2020)	Train test split: 2/3:1/3; tenfold cross-validation	RF	Sensitivity, specificity, AUROC, HLT statistics	RF: sensitivity† = 0.842, Specificity† = 0.91	R (Boruta)	[17]
Yuan et al. (2021)	(Stratified) train test split: 80:20; SMOTE algorithm to balance sample categories	ET, RF, SVM, XGBoost	Sensitivity, specificity, PPV, NPV, AUROC	XGBoost: sensitivity† = 0.81, Specificity† = 0.89	Python 3.7.0, (SMOTE)	[15]
Yoo et al. (2021)	Train test split: 70:30; parameters tuning with grid search	Auto-ML (TPOT), DNN, gradient boosting, RF	Sensitivity, specificity, PPV, NPV, F1 score, AUROC	RF: sensitivity† = 0.901 Specificity† = 0.722	R, Python	[16]

^†Evaluated at the optimal threshold, based on specific conditions.

AUROC: Area under the receiver operating characteristics curve; Auto-ML: Automated machine learning; DNN: Deep neural network; ET: ExtraTrees; HLT: Hosmer–Lemeshow test; RF: Random forest; SMOTE: Synthetic minority over-sampling technique; SVM: Support vector machine; TPOT: Tree-based pipeline optimization tool.

The future of ML techniques for diagnosis of candidemia

There are currently a few studies evaluating the approach through ML techniques for the early diagnosis of candidemia in clinical practice (as a prediction task based on clinical and routine laboratory features) [^15-18]. In our opinion, there are four important considerations that should be made regarding the future development of these techniques.

The first consideration consists in the overlap of two points: almost all current studies do not have very large sample sizes, rather they are similar to those of some previous prediction models of candidemia based on logistic regression [^28-33]; logistic regression is also frequently presented as a ML model (although we did not include studies employing logistic regression in the present review), and most teaching courses on ML for students or other interested professional figures start with logistic regression [^15-18]. This underscores the lack of a clear conceptual border between classical statistical techniques and ML models, since both are able to ‘learn’ from data. Then, learned knowledge is used for making a prediction for a new patient (for example, this is what is usually done when considering the Candida Score by Leon and colleagues in clinical practice, which was based on a logistic regression model) [30]. The crucial difference between classical approaches and modern ML techniques may rely on the highly interconnected concepts of ML and ‘big data’ [34,35]. Some modern ML techniques are able to learn complex associations/correlations across features (of which some may be difficult to, or even impossible, to be spotted or understood by humans) and then to predict an outcome/result with high accuracy, usually unachievable when using classical techniques (although this is not absolute rule). Provided there are no other biases (or they are limited) in the data, very large training samples are needed for these complex models to learn to predict without overfitting to learning data, which partly explains why the concepts of ML and ‘big data’ are highly interconnected [34,36].

The second consideration regards the fact that, in clinical practice, the overall diagnostic accuracy, considering the best compromise across performance measures (for example, sensitivity, specificity, PPV, NPV), may be not the most desired result from model training (see also previous section). Indeed, at the bedside of patients with signs and symptoms of sepsis, clinicians may prefer an early prediction tool favoring sensitivity over specificity, considering that the least desirable situation (in other words, with the worse prognostic impact for the patient) is not to treat a true episode of candidemia (in other words, delaying up to blood cultures results the initiation of an antifungal agent). In our opinion, the crucial question to be answered in the near future is whether ML models, provided extremely high sensitivity is guaranteed in order not to perilously miss the early diagnosis of true cases, would be able to dramatically improve specificity, limiting as much as possible (and more than what is currently obtained with classical predictions tools) the number of false positive cases, thereby avoiding useless antifungal treatments in line with antifungal stewardship principles [^37-40].

The third consideration regards the explainability of ML model results. Indeed, while models like logistic regression are interpretable (the ‘weight’ on the prediction conferred by clinical and laboratory features is evident from the model equation and can easily allow to calculate the related odds ratios), some complex ML models (for example, neural networks) are commonly defined as ‘black boxes’, since the calculations leading to the final prediction are hidden and more obscure to both data scientists and clinicians [^41-43]. From this perspective, we support the use of interpretable models whenever their performance equals that of black box models. However, when this is not the case and a black model is used, there are crucial issues regarding the degree of explainability achievable for the employed model. Explainability is the ability to understand how the model produced its output (which inherently cannot be full explainability, as this would be interpretability), which is an intense matter of research and should be considered essential for healthcare in general and for candidemia. This because the lack of understanding (lack of explainability) of the reasons leading to a possible mistake of the model (for example, classifying as ‘no candidemia’ a patient who conversely has candidemia) may hamper the recognition of biases leading to the mistake, consequently increasing the risk of wrong clinical decision unfavorably impacting patients' outcomes [44,45].

Finally, the importance of data quality should not be overlooked. An accurate, transparent, and large dataset is as least as important as the choice of the best suited ML model for training. Both increased speed of data collection (automation of extraction could be crucial for healthcare personnel overwhelmed by clinical tasks that have limited or no time to collect research data) and increased data quality can be achieved by well-designed and calibrated automatic extraction tools, provided format and semantics of extraction are properly standardized [46].

Conclusion

The use of ML techniques has started to be explored also for the early diagnosis of candidemia. As in other fields of medicine, the use of ML tools hold promise to provide highly accurate and real-time support to clinicians for relevant therapeutic decisions at the bedside of patients with suspected candidemia, although due caution is still necessary in terms of extrapolation and generalization of this potential and the application of ML tools for the early diagnosis of candidemia in real life practice, considering the limited number of currently available studies.

Future perspective

As described in depth in the main body of the article, further research is needed in terms of sample size, data quality, recognition of biases and interpretation of model outputs by clinicians to better understand if and how these techniques could be safely adopted in the daily clinical practice.

Author contributions

Conceptualization: DR Giacobbe; writing—original draft preparation, DR Giacobbe, C Marelli, S Mora, A Cappello, S Guastavino, A Vena, A Signori; writing—review and editing, DR Giacobbe, C Marelli, S Mora, A Cappello, S Guastavino, A Vena, A Signori, N Rosso, C Campi, M Giacomini, M Bassetti; supervision: N Rosso, C Campi, M Giacomini, M Bassetti.

Financial disclosure

The authors have no financial involvement with any organization or entity with a financial interest in or financial conflict with the subject matter or materials discussed in the manuscript. This includes employment, consultancies, honoraria, stock ownership or options, expert testimony, grants or patents received or pending, or royalties.

Competing interests disclosure

Outside the submitted work, M Bassetti has received funding for scientific advisory boards, travel and speaker honoraria from Angelini, Astellas, BioMérieux, Cidara, Gilead, Menarini, MSD, Pfizer, Shionogi, Tetraphase, Nabriva. Outside the submitted work, DR Giacobbe reports investigator-initiated grants from Pfizer, Shionogi, BioMérieux, and Gilead Italia and speaker/advisor fees from Menarini, Pfizer and Tillotts Pharma. The authors have no other competing interests or relevant affiliations with any organization or entity with the subject matter or materials discussed in the manuscript apart from those disclosed.

Writing disclosure

No writing assistance was utilized in the production of this manuscript.