The aim of this paper was to create a tool to predict the mortality of hospitalized patients presenting Ab in microbiological samples. The clinical advantage of this predictive tool is to identify the patients at risk that require early an intensive treatment in hospital settings.

In our study sample, mortality rate was 34.2%. This finding is consistent with meta-analyses in literature, in which carbapenem-resistant Ab mortality was estimated as high as 33–44% [5, 10], and a large proportion of the estimated total number of disability-adjusted life-years (DALYs) in Europe Economic Area (EEA) was attributed to colistin-resistant Acinetobacter spp in infected survivors [11].

Polymyxin-class colistin and tigecycline remain the most effective antibiotics to treat CRAb both in ICU and non-ICU wards of Sicilian Hospitals [12]. However, these antibiotic classes are often characterized by several side effects responsible of organ damages. Colistin-induced nephrotoxicity represents a clear example [13]. Recently, Cefiderocol, a siderophore cephalosporin has been used for carbapenem-resistant Gram-negative bacteria. It showed a minor nephrotoxicity compared with colistin, but increased liver enzymes have been observed [14]. To note, the CREDIBLE-CR trial reported higher mortality rate in CRAb patients treated with cefiderocol vs. other therapies (mostly polymyxin based regimens) [15]. So, the identification of hospitalized patients with Ab infection or colonization is a relevant issue that requires a dedicated tool [16]. To our knowledge, no mortality predictive model for Ab have been published yet. Only one predictive score for Ab hospital acquired pneumonia has been developed so far [17].

In the total cohort of 140 subjects, deceased and surviving patients presented similar comorbidities, evaluated both as single diseases and as Charlson Comorbidity Index, as parameter of cumulative burden of comorbidities [18]. Our results suggest that the predictors of mortality are related to clinical characteristics of the patients (Ab infection, use of NIV, previous antibiotic therapy, acidosis, hemodynamic parameters) more than to the degree of inflammation. Probably, concomitant infections by less aggressive bacteria or other inflammatory diseases might explain the increase of inflammation indices without a contextual increase in mortality. Concomitant infections were also investigated as a possible factor influencing mortality, however no influence in mortality was found.

Regarding the variables that entered the model, the presence of penicillin/amoxicillin antibiotic therapy within 90-days before hospitalization suggests that this class of antibiotics might select beta lactamases and carbapenemases producing Ab, responsible for higher mortality rates [18, 19].

“Status” was predictive in our model as well, showing that infected patients have higher mortality rates than colonized ones. However, as mentioned before, high mortality rates were also described in colonized individuals [6], justifying their inclusion in our work.

The variable “acidosis” entered the model, and it has been proven to predict mortality independently from hyperlactatemia [20]. Our data did not discriminate respiratory from mixed or metabolic acidosis. GCS and BP entered the model too, being considered also by different mortality and sepsis score, as the de SOFA and qSOFA scores [21].

Ab biofilm-forming property explained why NIV (versus in-mask O2 supplementation) entered the model as the principal source of infection during hospitalization [22,23,24].

We compared our custom model with three other models: the “status” alone (infected/colonized), APACHE II and SAPS II. APACHE II and SAPS II have been chosen due to their known association with poor prognosis in Ab infection [18, 25].

Our model correctly classified more deceased/surviving patients (32 out of 40) in comparison with the other models (APACHE II: 30 out of 40, SAPS II: 28 out of 40, “Status” model: 24 out of 40) in the validation cohort (see Supplementary Table 5). Misclassified patients were analyzed: our model incorrectly predicted 8 patients, of whom 4 survived and 4 died. Four out of these 8 patients were equally misclassified by the other models. We could not identify a common clinical pattern that could mislead the model prediction in these patients. The remaining 4 patients (failed by our model) were correctly identified by the other models, and they presented no respiratory tract infections as common characteristic. The absence of lung involvement in these patients may have impaired our model that relies on the variable “NIV” to predict mortality. They also had a high leucocyte count, which was investigated by the other models such as SAPS II and APACHE II and not by our custom model. Nevertheless, inflammatory measures impaired our model performance. In spite of these few cases, our model proved to be globally more performing than competitors.

Our sample was very heterogeneous due to the choice of lazily stringent inclusion/exclusion criteria. Still, our sample represents a ‘real world’ experience, depicting the heterogeneity of the patients admitted to ICU and non-ICU hospital wards. We believe that this choice makes the model very exploitable in different hospital settings.

In our ward, patient surveillance was based on rectal swab at admission. It is true that other colonized sites were missed in asymptomatic patient according to this procedure.

Another limit is represented by the size of the cohorts. Larger studies with a higher numerosity of the sample might have refined the result of the present study, producing more accurate parameters estimations able to improve the model performance.

Moreover, the model is applied at T1, that is the moment in which Ab is isolated, before administering antibiotic therapy. The effect of therapy is then not evaluated by the model.