Predictive Score for Carbapenem-Resistant Gram-Negative Bacilli Sepsis: Single-Center Prospective Cohort Study

A clinical–epidemiological score to predict CR-GNB sepsis to guide empirical antimicrobial therapy (EAT), using local data, persists as an unmet need. On the basis of a case–case–control design in a prospective cohort study, the predictive factors for CR-GNB sepsis were previously determined as prior infection, use of mechanical ventilation and carbapenem, and length of hospital stay. In this study, each factor was scored according to the logistic regression coefficients, and the ROC curve analysis determined its accuracy in predicting CR-GNB sepsis in the entire cohort. Among the total of 629 admissions followed by 7797 patient-days, 329 single or recurrent episodes of SIRS/sepsis were enrolled, from August 2015 to March 2017. At least one species of CR-GNB was identified as the etiology in 108 (33%) episodes, and 221 were classified as the control group. The cutoff point of ≥3 (maximum of 4) had the best sensitivity/specificity, while ≤1 showed excellent sensitivity to exclude CR-GNB sepsis. The area under the curve was 0.80 (95% CI: 0.76–0.85) and the number needed to treat was 2.0. The score may improve CR-GNB coverage and spare polymyxins with 22% (95% CI: 17–28%) adequacy rate change. The score has a good ability to predict CR-GNB sepsis and to guide EAT in the future.


Introduction
Early recognition and proper empirical antimicrobial treatment (EAT) of sepsis to cover carbapenem-resistant Gram-negative bacilli (CR-GNB) are fundamental in changing the current scenario of high mortality rates [1,2]. These are especially important because be avoided (true negative), divided by the total number of episodes. Both sides of the curve and a significance level of 5% were considered in all tests.

Patients and Episodes of Sepsis
Of a total of 629 ICU admissions followed by 7797 patient-days, we identified 429 single or recurrent episodes of SIRS/sepsis in 313 patients after having completely monitored the entire cohort. We enrolled 329 episodes of hospital-acquired sepsis identified in 219 patients. Approximately 33% (108/329) of the episodes had at least one species of CR-GNB as the cause of sepsis, which were classified as Group 1 (108 episodes in 75 patients). Group 2 comprised 221 sepsis episodes in 182 patients, caused by CS-GNB (26%, 57/221), non-GNB agents (23%, 50/221), or an undetermined etiology (56%, 124/221) (Figure 1). Recurrent episodes of hospital-acquired sepsis were identified in 30% (65/219) of patients during ICU stay; 18% (39/219) were allocated in both study groups. We identified a median of one (range of 1-10) episode in the study population and a ratio of 1.4 episodes per patient in CR-GNB group versus 1.2:1 among the control group (p = 0.15). Tables 1 and 2 present the profiles of patients and episodes in each group of the study, respectively. The total SOFA score exhibited median values of seven (mean 7.1; SD 3.5; range 1-17 among Group 1 and mean 6.6; SD: 3.11; range 0-18 in Group 2) and eight points (mean The case-case-control methodology, designed to define risk factors for CR-GNB sepsis, was described previously [5]. The variables and definitions, the information sources, the microbiological methods, the forms used, and the checklist of STROBE recommenda- tions [14] of the cohort study have also already been published [5]. The sepsis-3 criteria [15] were used in an adapted manner and applied retrospectively in a random sampling of the cases included in the cohort [5], with regard to increases in two or more points of the Sequential Sepsis-Related Organ Failure Assessment (SOFA), while also considering the qSOFA when the patient no longer required mechanical ventilation or sedation. Recurrent sepsis was defined as a new episode of sepsis developing after the resolution of clinical and laboratory parameters of sepsis or the recrudescence of sepsis with evidence of new etiology by cultures during the ICU stay [5].
The following variables were also investigated at ICU admission: demographic data, origin and reason for hospitalization and clinical severity scores, such as the Charlson comorbidity index, Simplified Acute Physiology Score (SAPS 3), and SOFA. The following variables were investigated prior to the blood culture assessment upon entry to the study in each episode of sepsis: comorbidities, length of previous hospitalization in the hospital and ICU, invasive devices (central vascular catheter, mechanical ventilation, and indwelling bladder catheter), hemodialysis, enteral and parenteral nutrition, the SOFA score, and C-reactive protein levels. In addition, we investigated the etiology and type of infection that motivated sepsis, using the strict criteria developed by Klein Klouwenberg et al., (2013) [16], adapted to include Centers for Disease Control and Prevention definitions of healthcare-associated infections [17], the presence of concurrent infection, and the outcomes for discharge or death in the ICU and hospital.
To avoid potential bias, all data collected were standardized and monitored throughout the study. Post hoc analysis was performed to review all clinical, radiological, and microbiological data, including the evidence of sepsis-2 [13] and -3, and the etiology and source of infection. We also classified the plausibility of infectious source as being definitive, probable, possible, or undetermined [16].
In this cohort study, antimicrobial therapy was investigated whenever used up to the 3 days prior to the initial blood culture collection, as was empirical antimicrobial therapy comprising the period between blood culture collection and the final microbiological results that determined or not the etiology of sepsis. All the antimicrobial drugs maintained or initiated after initial blood culture collection that had no other reason to be used were considered to be for the treatment of the investigated sepsis episode. Adequate empirical treatment was defined as the use of any antimicrobial regimen with at least one drug susceptible to cover the etiology of sepsis.

Database, General Statistical Analyses, and Risk Score Development
We collected and managed study data using REDCap electronic data capture tools hosted at Instituto Oswaldo Cruz, Fundação Oswaldo Cruz (IOC/FIOCRUZ) [18]. All statistical analyses were performed using SPSS®statistics v22.0 software. Categorical variables were compared using the exact chi-square or Fisher tests; continuous variables were compared using the Mann-Whitney-Wilcoxon test [19].
Predictive factors for CR-GNB sepsis have been defined previously: (1) Table S2). ROC curve analysis determined the accuracy of the score in predicting sepsis by CR-GNB in the entire cohort by assessing the area under the curve (AUC) [20]. Therefore, the risk factors independently associated with CR-GNB sepsis, identified on the basis of this cohort study and using the case-case-control design [5], were subsequently scored in order to develop a score system. The score was applied in all episodes included in the cohort, regardless of whether the etiology was identified, in order to better estimate its ability to predict CR-GNB sepsis throughout the population studied as real life.
The case-case-control methodology is recommended for the investigation of antimicrobials as factors associated with bacterial resistance profiles [5,21]. Thus, points were attributed to each variable determined, in our previous publication [5], as a predictor of CR-GNB sepsis by multivariate analysis, using the logistic regression coefficients (Supplementary Table S2), as recommended [22]. The discriminatory capacity of the clinical risk score was calculated by applying it in all sepsis episodes in this study, regardless of whether it was recurrent or single, because each episode was treated as an individualized event; all predictive factors were investigated during the hospitalization period prior to the initial blood culture. In addition to sensitivity, specificity, negative predictive value (NPV), positive predictive value (PPV), and accuracy, we report the Youden index, calculated according to standard methods [10,20]. These indicators were used at various cutoff points to determine the best cutoff to discriminate high and low risk of CR-GNB as the etiology of sepsis. The proportion of episodes scoring 0, 1, 2, 3, and 4 points in each study group was also calculated, as well as in the set of sepsis caused by CS-GNB (carbapenemsusceptible Gram-negative bacilli), non-GNB sepsis, and undefined etiology, in order to better understand the behavior of the score in all episodes occurring in the cohort.
We analyzed the number needed to treat (NNT) to avoid empiric therapy divergence in one patient [23]. The uses of empiric carbapenems, empiric polymyxins, and adequate EAT were also analyzed in both study groups and according to the scoring system, so as to calculate the proportion of patients in which the score would be useful to increase or avoid CR-GNB coverage.
The adequation rate of the score was calculated considering the number of episodes in which it would be necessary to receive empiric polymyxin therapy to cover CR-GNB (true positive), plus the number of episodes in which empiric polymyxin therapy could be avoided (true negative), divided by the total number of episodes. Both sides of the curve and a significance level of 5% were considered in all tests.

Patients and Episodes of Sepsis
Of a total of 629 ICU admissions followed by 7797 patient-days, we identified 429 single or recurrent episodes of SIRS/sepsis in 313 patients after having completely monitored the entire cohort. We enrolled 329 episodes of hospital-acquired sepsis identified in 219 patients. Approximately 33% (108/329) of the episodes had at least one species of CR-GNB as the cause of sepsis, which were classified as Group 1 (108 episodes in 75 patients). Group 2 comprised 221 sepsis episodes in 182 patients, caused by CS-GNB (26%, 57/221), non-GNB agents (23%, 50/221), or an undetermined etiology (56%, 124/221) ( Figure 1). Recurrent episodes of hospital-acquired sepsis were identified in 30% (65/219) of patients during ICU stay; 18% (39/219) were allocated in both study groups. We identified a median of one (range of 1-10) episode in the study population and a ratio of 1.4 episodes per patient in CR-GNB group versus 1.2:1 among the control group (p = 0.15). Tables 1 and 2 present the profiles of patients and episodes in each group of the study, respectively.
The total SOFA score exhibited median values of seven (mean 7.1; SD 3.5; range 1-17 among Group 1 and mean 6.6; SD: 3.11; range 0-18 in Group 2) and eight points (mean 7.7; SD 4.2; range 0-20 among group 1 and mean 7.4; SD 4.3; range 0-18 in Group 2) in both groups, at ICU admission and at initial blood culture (p = 0.89), respectively, showing no difference in the degree of dysfunction or previous organic failure from the univariate analysis (Table 2). Nosocomial diarrhea (p < 0.001), infection/colonization by any CR-GNB prior to sepsis episode (p < 0.001), and previous infection (mostly hospital-acquired bacterial infection or sepsis) (p < 0.001) were more frequent comorbidities in CR-GNB sepsis than non-CR-GNB sepsis. The use of any mentioned invasive devices was significantly higher among episodes of CR-GNB sepsis than in the comparison group, such as the use of enteral nutrition (p < 0.001) ( Table 2). All the CR-GNB sepsis episodes were treated with antimicrobials prior to the collection of blood culture, as compared with 82% in those with another or indeterminate etiology (p < 0.001), with significant differences exhibited Antibiotics 2023, 12, 21 6 of 16 in carbapenem (88% versus 42%, p < 0.001) and polymyxin (57% versus 23%, p < 0.001) usage. The median lengths of hospital and ICU stays prior to the incidence of sepsis were significantly higher in the episodes of CR-GNB sepsis than in episodes of sepsis due to another or undetermined etiology (p < 0.001). The 30 day mortality rate (66%, 95% CI: 47-89% versus 44%, 95% CI: 34-55%, p = 0.004) and hospital mortality rate (70%, 95% CI: 51-94% versus 55%, 95% CI: 44-68%, p = 0.038) were significantly higher in patients with CR-GNB sepsis than in comparators, considering the last or unique episode of sepsis identified in each patient. Table 1. Clinical and demographic characteristics of patients with episodes of CR-GNB sepsis and sepsis due to another or undetermined etiology (non-CR-GNB sepsis).  a Comorbidities or conditions diagnosed prior to the initial blood culture collection during the hospital stay; b nosocomial diarrhea (three or more daily stool episodes for two or more days); c >10 mg prednisone for more than 50 days, >40 mg corticosteroid for more than 7 days, or immunomodulatory agent (examples: monoclonal agents, methotrexate); d granulocytes < 500 cells/mm 3 ; e prior to the initial blood culture collection; f within 24 h of the initial blood culture collection (at least two of the following criteria: body temperature > 38 • C or <36 • C, heart rate > 90 beats/minute, breath rate > 20 breaths/min or PaCO 2 < 32 mmHg, white blood cell count > 12,000 cells/mm 3 or <4000 cells/mm 3 , or ≥10% immature neutrophils); g within 24 h of the initial blood culture collection; h in relation to the SOFA at ICU admission; i Pearson's chi-square test, Fisher's exact test, or Mann-Whitney-Wilcoxon U test was used, as required and considering p < 0.05 as statistically significant (in bold). Abbreviations: AIDS, acquired immunodeficiency syndrome; CR-GNB, carbapenem-resistant Gram-negative bacilli; PCR, C-reactive protein; SIRS, Systemic Inflammatory Response Syndrome; SOFA, Sequential Organ Failure Assessment. j Pleural empyema (n = 2), vascular fistula (n = 1), and otitis (n = 1). * in bold statistically significant test.

Etiology and Infectious Focus of Sepsis
The distribution of infectious foci and its plausibility (possible, probable, or definitive) by groups of study are shown in Supplementary Figures S1 and S2. Among the sites of infection, ventilator-associated pneumonia (VAP) was more prevalent in the CR-GNB sepsis group than in the comparison group (47% and 10%, respectively, p < 0.001), whereas non-ventilator-associated hospital-acquired pneumonia was more common in Group 2 than in the CR-GNB group (23% and 13%, respectively, p = 0.04) ( Table 2).
Among episodes of sepsis classified as another etiology or non-CR-GNB (Supplementary Figure S3 Streptococcus and Enterococcus species were identified in 8% (9/119 in eight episodes) and 6% (7/119 in seven episodes) of agents, respectively. Candida albicans, non-albicans Candida spp., and yeasts were identified in 8% (9/119) of isolates in nine episodes, whereas Mycobacteria tuberculosis (diagnosed during hospitalization period or health-care associated) and other agents (Moraxella sp., Haemophylus sp., and anaerobic Gram-positive bacilli) were responsible for three episodes each.

Performance of the Clinical-Epidemiological Score to Guide EAT
The sensitivity, specificity, numbers of false negatives and false positives, negative predictive value (NPV), positive predictive value (PPV), and accuracy of clinical score using different cutoff points are presented in Table 3. Sepsis episode distributions in both groups according to the score obtained are shown in Figure 2. The AUC value of the ROC curve was 0.80 (95% CI 0.76-0.85, p < 0.001) (Figure 3). and Enterococcus species were identified in 8% (9/119 in eight episodes) and 6% (7/119 seven episodes) of agents, respectively. Candida albicans, non-albicans Candida spp., a yeasts were identified in 8% (9/119) of isolates in nine episodes, whereas Mycobacteria berculosis (diagnosed during hospitalization period) and other agents (Moraxella sp., H mophylus sp., and anaerobic Gram-positive bacilli) were responsible for three episod each.

Performance of the Clinical-Epidemiological Score to Guide EAT
The sensitivity, specificity, numbers of false negatives and false positives, negat predictive value (NPV), positive predictive value (PPV), and accuracy of clinical score ing different cutoff points are presented in Table 3. Sepsis episode distributions in b groups according to the score obtained are shown in Figure 2. The AUC value of the RO curve was 0.80 (95% CI 0.76-0.85, p < 0.001) (Figure 3).

Figure 2.
Distribution of groups of sepsis episodes according to the score obtained from the ca case-control design study already published [5]. Score for CR-GNB sepsis: infection prior to sepsis episode, one point; previous use of mechanical ventilation, one point; previous use of c bapenem, one point; length of hospital stay greater than or equal to 19 days, one point. Each sc factor present contributed to one point for the final score in each sepsis episode in the entire coh Figure 2. Distribution of groups of sepsis episodes according to the score obtained from the case-casecontrol design study already published [5]. Score for CR-GNB sepsis: infection prior to the sepsis episode, one point; previous use of mechanical ventilation, one point; previous use of carbapenem, one point; length of hospital stay greater than or equal to 19 days, one point. Each score factor present contributed to one point for the final score in each sepsis episode in the entire cohort.    We observed that when a cutoff point ≥ 3 was used, we found a high sensitivity (92%, 95% CI, 85-96%), with a significant reduction in false negatives, when compared with a score of four, and the best accuracy (74%, 95% CI 69-78%) in discriminating episodes with a higher risk for CR-GNB sepsis (Table 3). In contrast, a score ≤ 1 demonstrated high sensitivity (99%; 95% CI: 95-100%) to exclude those at risk of CR-GNB sepsis. The performance of a clinical-epidemiological score ≥ 3 to predict CR-GNB sepsis, in accordance with the etiology of sepsis, score distribution, and prior use of polymyxins, is shown in Supplementary Table S3. The NNT using this cutoff for an additional episode to benefit CR-GNB coverage was 2.0 (95% CI: 1.5-2.1, p < 0.0001).

Discussion
The prevalence of CR-GNB sepsis is increasing globally, and the rational use of antibiotics is one of the pillars to contain the agents and extend the shelf-life of antibiotics [24][25][26]. Physicians around the world, particularly those in ICUs, need to use effective antimicrobial treatments as quickly as possible to prevent disability and death in patients without indiscriminately prescribing broad-spectrum antibiotics [27]. When to escalate sepsis therapy to cover these microorganisms, as well as how to preserve last-resort antibiotics, particularly newly developed drugs, are frequent impasses. The solution to this problem seems simple, but it is not to prescribe the antibiotic with just the right spectrum. Thus, a clinical score to help decide whether or not to expect antimicrobial resistance at the bedside when deciding EAT, especially with easy-to-use components, is of great importance.
From an antimicrobial stewardship standpoint, it is recommended that EAT be adjusted to the local data [11], according to the local epidemiology of patients and pathogens. The developed score is a simple tool, because the predictor variables can be easily obtained during initial assessments, at a time when the clinical-epidemiological characteristics remain the only determinants of a patient s risk. The score aims to guide EAT until a timely antimicrobial de-escalation based on microbiological identification, susceptibility testing, and clinical response is implemented. These are essential strategies to improve the prognoses of septic patients, preserving the effectiveness of existing antimicrobials and preventing the emergence of resistance. In our study, the score represented a valuable option to effectively improve polymyxin coverage in 21% (14-30%) of CR-GNB sepsis cases and may be useful to withdraw unnecessary coverage in 26% (95% CI: 21-31%) of the studied population.
To the best of the authors' knowledge, this is the first study to develop a suitable predictive score to guide EAT for CR-GNB sepsis. Although bacteremia has been widely used as a proxy for sepsis, in our study, bacteremia did not represent all types of infections and was identified in a smaller proportion of cases, mainly catheter related bloodstream infections [5]. Bacteremia has been identified in fewer than 30% of septic cases in ICUs [5,28,29], and blood culture is relatively insensitive for the recovery of agents in VAP [30,31], the predominant infectious source of ICU sepsis [2,29]. In addition, a diversity of GNB species could be the cause of sepsis in monomicrobial and polymicrobial infections as demonstrated, making the score for specific species not useful for targeting therapy. Considering incident cases, all single and recurrent episodes in the cohort, and the variety of infections commonly identified in hospital-acquired sepsis, the estimates are likely to be more representative of what happens in real life. In addition, including episodes with unknown etiology improves our understanding of the complex epidemiological picture, aiming to understand the balance between damages and benefits of novel tool use.
Successful longitudinal follow-up of this cohort was essential to achieve the study proposal. Prior infection, mechanical ventilation, carbapenem usage, and hospitalization longer than or equal to 19 days were associated with a higher risk of sepsis due to CR-GNB [5], in agreement with the literature concerning the risk factors for the acquisition of or infection by CR-GNB [32][33][34][35].
Using data from a prospective cohort, we developed a predictive score that allowed us to accurately identify patients at high or low risk of CR-GNB sepsis. The score performance was considered to be satisfactorily estimated when applied to the entire cohort. The best combination of sensitivity (92%, 95% CI: 85-96%) and specificity (65%. 95% CI: 58-71%) was reached when we used three or more points as cutoffs. The best specificity was achieved using four as the cutoff (76%, 95% CI: 70-81%). A score of less than or equal to one showed optimal sensitivity (99%; 95% CI: 95-100%) for the exclusion of CR-GNB sepsis cases. Those exhibiting scores of two had at least 8% (8/96) false negative results, which is high considering the severity of clinical cases. Notably, higher scores (three and four) had 25% and 29% occurrences of inappropriate EAT, respectively, and these patients would benefit with a broader-spectrum EAT if the score were to be applied prospectively. Patients with lower scores (zero and one) used polymyxins unnecessarily, while exerting pressure on antimicrobial resistance and incrementing cost. Sepsis by non-GNB (including Gram-positive bacteria) also had 35% inadequate EAT, although the score would not solve all the issues, because it is designed for CR-GNB.
Additionally, the inhibition of CR-GNB growth in cultures due to polymyxin use could explain false positive or discordant results, because the drug was administered in 40% of these cases within the 3 days preceding the collection of cultures. Therefore, some of the episodes in the control group would have been in the case group if the agent was identified by microbiological tests. Thus, considering the high incidence of CR-GNB sepsis and high consumption of carbapenem and polymyxins, we expect that the score has the potential to benefit a larger number of patients who are misclassified as controls. Consequently, an NNT of two to avoid mismatch in a single case is probably an overestimation.
Taking everything in account, our findings indicate that episodes with scores of three or four should be treated with the best drugs available to cover CR-GNB sepsis, knowing that we may need to treat two patients to adequately cover one. For those exhibiting scores of zero and one, the drugs can be preserved. However, episodes presenting a score of two should be handled on a case-by-case basis. Considering that 77% (144/186) of the episodes with scores of three or four had an etiology later determined by conventional cultures, and a higher proportion would be achieved if molecular tests were incorporated into the diagnosis of sepsis [36], broad-spectrum EAT initially guided by the new score with subsequent adjustment of therapy, according to the microbiological results, contributes to its incorporation into clinical practice. Compared with standard therapy (clinical practice), we estimated an increase in the adequacy rate, considering the use of empiric polymyxins as a proxy for adequate CR-GNB cover, in 22% (95% CI: 17-28%) of study episodes using the CR-GNB sepsis score. A more in-depth analysis, however, of our results is still under evaluation to indicate the best strategy for the applicability of the score in future studies.
In fact, in this scenario of a high prevalence of CR-GNB sepsis, as we demonstrated, high intakes of both carbapenems and polymyxins are expected, because these drugs remain the most widely used drugs in double or triple combinations for the treatment of CR-GNB sepsis, especially in hospitals with restricted access to novel therapeutic options. EAT with different combinations of broad-spectrum antibiotics, generally using the old drugs polymyxins, tigecycline, amikacin, and/or meropenem, is the standard treatment regimen for hospital-acquired septic patients in public Brazilian ICUs, despite the reports of increased resistance [37]. The emergence of resistance to newer agents, such as ceftazidimeavibactam, is critical and further intensifies this problem [38]; therefore, using a bedside tool to preserve the administration of expensive wide-spectrum antibiotics in critically ill patients with a low score (≤1), limiting the use to those with high scores (≥3), indicates the importance of the developed tool and its potential.
The main limitations include the single-center nature of the study, which warrants caution for any generalization of our findings. Internal validation is required before implementation, and external, multicentric validations in hospitals with a similar size, profile, and prevalence rates of CR-GNB sepsis are important before extrapolating these results beyond our cohort. The adaptation of management strategies to the local epidemiological data is a general recommendation for hospital infection, which indicates locally cyclic evaluations. The performance of microbiological methods and the use of antimicrobials, which may have inhibited microbial growth in cultures, could also have influenced our estimates. Nonfermenting CR-GNB were more frequent than Enterobacterales CR-GNB in this cohort, whereas carbapenemase production was previously estimated in 76% of CR-GNB isolates [5]. These factors should be taken into consideration depending on the sepsis microbiology profiles across institutions, as well as the types of microbiological monitoring, once VAP is likely less represented by bacteremia [30,31].

Conclusions
Developing a predictive score for empirical sepsis therapy based on local microbiological and clinical data, and estimating NNT CR-GNB sepsis to avoid inadequate EAT and overtreatment in a patient are feasible approaches. Such a score was proposed in this study, and its overall adequacy rate in not only indicating the correct treatment of CR-GNB, but also in avoiding the overtreatment of such pathogens, sparing antibiotics such as polymyxins, is promising. However, defining strategies for the EAT of sepsis is very ambitious, considering limitations in the drug arsenal, costs, drug-drug interactions, and adverse events, as well as the clinical severity of cases. This also comprises moral principles, because decisions of when and how to use EAT affects current patients, as well as future generations. The emergence of resistance to novel drugs is a serious issue; therefore, more clinical studies are needed to determine the ideal approach for EAT, especially in the high-incidence setting of CR-GNB sepsis (33%), as well as for patients with severe sepsis and septic shock, such as the majority of our cases.
Supplementary Materials: The following are available online at https://www.mdpi.com/article/ 10.3390/antibiotics12010021/s1: Figure S1. Distribution of infectious focus in episodes of sepsis caused by CR-GNB (n = 137 infectious focus) and by another or undetermined etiology (n = 212 infectious focus); Figure S2. Plausibility of infectious focus in CR-GNB and non-CR-GNB sepsis, according to the adapted criteria of Klein Klouwenberg et al., (2013); Figure S3. Etiological agents in episodes of CR-GNB sepsis (in orange and blue: n = 162 isolates in 108 episodes) and non-CR-GNB sepsis (in gray: n = 120 isolates in 97 episodes); Figure S4. Polymyxin and carbapenem usage before (3 days preceding) and after (empirical therapy) initial blood culture in CR-GNB sepsis (n = 108 episodes) and non-CR-GNB sepsis (n = 217 episodes); Figure S5. Polymyxin and carbapenem usage before (3 days preceding) and after (empirical therapy) initial blood culture, according to the etiology of sepsis; Figure S6. Polymyxin and carbapenem usage before (3 days preceding) and after (empirical therapy) initial blood culture, according to predictive score; Figure S7. Adequate empirical antimicrobial therapy according to the etiology of sepsis and the predictive score for CR-GNB sepsis; Table S1. Sepsis-2 and adapted sepsis-3 criteria; Table S2. Score for assessing the risk of CR-GNB sepsis, according to the logistic regression coefficients; Table S3. Performance of the clinical-epidemiological score ≥ 3 to predict CR-GNB sepsis, according with the etiology of sepsis, score distribution, and prior use of polymyxins; Table S4. Polymyxins, carbapenems, and adequate empirical antimicrobial therapy according to the performance of the score ≥ 3 in predicting CR-GNB sepsis; Table S5. Pre-test and post-test adequate empirical polymyxin therapy, according to the score ≥ 3 in predicting CR-GNB sepsis. References [5,13,15,16,39,40] are cited in the Supplementary Materials.