Can the risk of a resistant pathogen be assessed before microbiology results arrive?

Before microbiology results are available, clinicians must choose empiric antimicrobial treatment without knowing the susceptibility of the pathogen causing the current infection. Prior antibiotic use, recent hospitalization, and known colonization with a resistant organism are often considered risk factors. But can published estimates of their association with resistance be used for patients in another hospital?

A systematic review by Kuzmenkov and colleagues, published in Clinical Microbiology and Antimicrobial Chemotherapy, analyzes studies of risk factors and models for predicting the detection of bacteria with clinically significant resistance phenotypes.

Read the article at the journal · Download the Russian-language PDF

What the review covered

The authors reviewed work published from January 2000 through April 2026. The analysis included 382 publications: 317 studies of individual risk factors and 65 studies of prediction models. They examined these groups separately. The studies differed in patient populations, clinical events, and definitions of resistance, so the authors did not calculate a single risk estimate for all settings.

Which risk factors were studied?

The studies of individual factors most often examined the following groups of characteristics (the number of publications is shown in parentheses):

• Prior antibiotic treatment (172) – including use of carbapenems, fluoroquinolones, and cephalosporins.
• Hospitalization and other healthcare exposure (91) – for example, recent hospitalization.
• Demographic characteristics (75) – age and sex.
• Comorbidities (72).
• Invasive devices and procedures (71) – including mechanical ventilation and surgery.
• Microbiology history (55) – previous colonization or infection with a resistant organism.

The review also includes studies of the site of the current infection (54), severity and ICU stay (48), immunosuppression (30), and geographic factors or residential setting (29). A study could examine more than one group, so the counts sum to more than 317.

How often a factor was studied does not establish its predictive value. Even for similar factors, estimates of association varied by pathogen, resistance phenotype, and clinical setting.

One factor, different estimates of association

The authors examined factors in relation to specific resistance phenotypes and clinical events, rather than as universal predictors. The table shows three combinations from the broader analysis. Even within these combinations, estimates differed substantially across studies:

ESBL means extended-spectrum beta-lactamase; MRSA means methicillin-resistant Staphylococcus aureus.

OR (odds ratio) is a statistical measure of association between a factor and an outcome. The table shows ranges of point estimates from different studies, not confidence intervals, pooled effects, or probabilities for an individual patient. Estimates may differ because of the patient population, how the factor was defined, or the care setting. More combinations appear in Table 4 of the original article and in the interactive evidence map below.

Published studies can help identify variables for local analysis. But their numerical estimates should not be used in another hospital without checking whether they apply there.

Why a published score may not work in another hospital

A prediction model uses several characteristics to estimate the probability of a specified outcome. Before applying it, clinicians need to check both how well it separates higher-risk from lower-risk patients and its calibration: do predicted probabilities match the observed frequency of that form of resistance among patients in the hospital?

Some models were incompletely reported; many had not been tested in independent patient populations (external validation), and calibration results were rarely reported. Good performance during development therefore does not show that a model will work in another hospital.

The role of local data

Choosing empiric treatment calls for knowing which pathogens occur with a given clinical syndrome, how often they are resistant to candidate agents in the hospital, and which patient characteristics are associated with a particular form of resistance. A cumulative antibiogram summarizes local susceptibility data; a prediction model estimates probability for an individual patient. Neither substitutes for the other.

Local microbiology surveillance can track pathogens and resistance and, when additional information is available, examine how these patterns relate to characteristics of the test and patient. Individual prediction also requires standardized clinical data, an appropriate model, and validation in the intended population. An antibiogram dataset alone does not provide a patient-level prediction.

The review mentions ABioGram among resources for working with microbiology data. We discuss the possibilities and limits of machine learning for AMR prediction in another article.

Interactive evidence map

The article includes an interactive evidence map (Supplement S10). Readers can search studies by risk factor, resistance phenotype, clinical event, and prediction model. The map helps navigate the published evidence, but it does not calculate risk for an individual patient and is not intended to guide treatment selection. Its interface is in Russian.

Takeaway

The review offers candidate factors for further evaluation, not a universal risk score. Their relevance and the performance of prediction models must be checked in the population where they are intended for clinical use.

Source: Kuzmenkov A.Yu. et al. Clinical Microbiology and Antimicrobial Chemotherapy. 2026;28(2):209–240. DOI: 10.36488/cmac.2026.2.209-240.