Knowledge of the bacterial group involved in the process of bacteremia promotes early initiation of the most appropriate empirical therapy and increases the chances of patient survival4. Although many RLBs have been reported as predictors of bacteremia, sepsis, or mortality, they are not commonly used to distinguish between BSI-GN or BSI-GP5,6,8,10,11,15,16,17. On the 68 RLB evaluated in this study, we implemented a new statistical model with 4 covariates, predicting BSI-GN with an AUC of 0.69, an accuracy of 0.64, a sensitivity of 0.61 and a specificity of 0.67, which could help clinicians choose the antimicrobial before the final BC result (Table 4, Fig. 3).
This model involves several analyzes where the predictor variables are analyzed simultaneously, so that the effect of each variable is adjusted for the effect of the others. Some of these biomarkers have already been proposed in other studies to predict BSI or the bacterial group involved in BSI, but few have shown good sensitivity and specificity as an independent test.9,11,14,17.
Regarding some patient data included in the study, the inclusion of the pediatric population is a controversial subject. While Bash et al. demonstrated that the large differences in the immunological system would make this population special, Oksuz et al. and Colak et al., when using pediatric patients in the proportion of 12.5 and 35%, respectively, reported that the inclusion of children and adults was beneficial in their studies1,7,18. We maintained the pediatric population (28.6% in this study) and agreed with Colak et al. that the inclusion of age may have helped reduce the bias of the selected group, since age in our study was considered a covariate, which was related to other biomarkers in all statistical analyzes (Table 3).
In this study, the prevalence of BSI-GP was 47.7% and BSI-GN was 52.3% (Tables 1, 2). Among the BSI-GNs, species of the Enterobacteriaceae family were the most represented, mainly by E.coli (26.5%) and Klebsiella pneumoniae (19.7%). We highlight the high mortality associated with non-glucose-fermenting species of GN bacilli such as P. aeruginosa 51.5% and A.baumannii 55%. Among the BSI-GPs, the highest frequency was S. aureus (43.7%), whose mortality was 30.4%. Current data supports the literature9,11,19,20,21demonstrating greater difficulty in processing BSI-GN, sometimes related to specific characteristics of this bacterial group, such as known lipopolysaccharide with GN-specific endotoxin21,22,23; or greater resistance to antibacterial drugs of clinical use, primarily found in A.baumannii and P. aeruginosa, which increases their death rate24.25.
Thus, if RLB data could predict the bacterial group involved in BSI, even with an estimated 70% accuracy, it would be extremely useful for initiating more targeted empirical therapy when blood culture results are not yet available to the physician. , which can take an average of 2-3 days5,6,7.
Levy, 2017 demonstrated that prior knowledge of the presence of an infectious focus can help signal the infectious agent in BSIs26. In this study, the initial infectious focus of the abdomen had an OR 2.48 (1.43–4.41) p
In our study, mortality between BSI-GP and BSI-GN showed a statistically significant difference only when the initial focus was intra-abdominal (p = 0.002). Clinical parameters added to the initial infectious focus, in combination with RLB parameters, can increase the chances of predicting the bacterial group in BSI, helping to choose the most appropriate antimicrobial treatment and thus contributing to a reduction in morbidity and disease. mortality.
Unexpectedly, ED had the highest number of BSI (37.3%), followed by USI, with only 15.4%. Wang et al. also found a high frequency of BSI in the ED due to delays in identifying causative agents of BSI and in initiating appropriate antimicrobial therapy, which alter patient prognosis.12,27,28,29,30,31.
When analyzed individually, some RLBs showed statistical differences in univariate analysis, such as creatinine (p 11. Ljunsgstrom et al. determined a predictive model of bacteremia composed of four biomarkers with high AUC (0.78), and which was also significantly higher (all p8.
We present here, for the selection of variables, a new selection model capable of discriminating bacterial groups in BSI using four sequential filters (Fig. 1). Next, we present a predictive model with 16 RLBs (CM) that showed an AUC of 0.72, which is slightly higher than the AUC of the RM (0.69) consisting of only four easily achievable RLBs in most scenarios. laboratories. The CM and RM had a sensitivity of 0.62 and 0.61, respectively, and a specificity of 0.67 for both (Table 4 and Fig. 3). The CM and the RM are two hypotheses with equal efficiencies and, according to the principle of parsimony which advocates the simplest, the RM would be more easily usable, allowing faster interpretation and a lower cost.
LR models built with automatic variables already described in the literature, such as forward selection, backward selection, stepwise selection, RFE and Boruta feature selection, do not outperform the values of the estimated metrics by the RM proposed here32.
Tang et al. determined specific combinations involving lymphocyte count, PLT, neutrophil to lymphocyte ratio (NLCR), mean platelet volume (MPV), MPV/PLT ratio (MPV/PLT), platelet to larger cell ratio (P- LCR), and C-reactive protein (CRP), and obtained good ability to distinguish various pathogens in BSI from negative BC. The highest AUC in their study was for BSI-GP, 0.715, and 0.777 for BSI-GN, with 0.797 for E.coli BSI, 0.943 for Enterobacter aerogenes BSI, 0.830 for P. aeruginosa BSI and 0.767 for S. aureus BSI14. Our work was carried out in patients with BSI confirmed by GN or GP, and did not include patients with negative BC, so the changes in RLB were greater in both groups, which reduced the difference between them. and may explain the lower AUC values obtained here. When Tang et al. compared BSI-GP with BSI-GN, the highest AUC obtained was 0.63, while our CM and RM had an AUC of 0.71 and 0.69, respectively. Our work assessed RLB at the group level, not at the species level. Although PLT was one of the selected biomarkers in our statistical model, the MPV and P-LCR parameters that assess the platelet series are not part of the RLB routine at the study hospital.
Biomarkers such as PCT, lipopolysaccharide binding protein, CD14-ST isoform and interleukin-6 measurement are described as potential biomarkers to distinguish BSI-GN, however, these data are inconsistent , as some studies have shown promising results with PCT1,12,33,34. This is contrary to Ruddel et al., who found low discriminating power of PCT to guide treatment decisions.9,35,36. Our study was based on the RLB, and therefore did not evaluate the aforementioned tests. This fact could be considered a limitation since they presented promising results; however, these markers are not part of the routine of most clinical laboratories.
The retrospective nature of the study may introduce bias into the analysis of the results. We worked with one dataset to generate the predictive model, and with another dataset to test it, so while we believe our model is validated, it needs to be applied in other healthcare settings and its applicability has yet to be tested in clinical practice. We emphasize that the models should not be applied to outcomes after blood transfusions and electrolyte replacement.
Challenging the proposed model with unused data to develop the predictive model has not yet been a common practice in the literature. The model built in our study was tested with a database that was not part of the model validation. The predicted values obtained for the CM (AUC of 0.67) and for the RM (AUC of 0.68) confirmed the discriminating capacity of the model developed for the BSI-GN and demonstrated that the two models are similar (Table 4 and Fig. 3) .
It would be ideal to provide a model with high specificity for the indication of BSI-GN or BSI-GP, so that we can indicate high and low values for creatinine, PLT count, RBC and MCH. However, as we chose to do a robust study with multiple variables, our specificity was around 70%. As described in Table 3 and Figure 1, we can see that the values referring to creatinine, PLT count, RBCs and MCH for BSI-GN and BSI-GP have significant differences in the medians. For example, creatinine values for BSI-GN are generally higher for BSI-GN than for BSI-GP (median of 1.66 for BSI-GN and 0.8 for BSI-GP) as well than for the number of PLTs, conversely for the BSG-GN the values are generally lower than those of the BSI-GP (median for the BSI-GN of 166 and for the BSI-GP of 204). However, as described in the text and seen in Fig. 1 and Table 3, the standard deviation is high and therefore we cannot establish a threshold that indicates BSI-GP or BSI-GN. These discrepancies are inherent in the unique characteristics of each patient, demonstrating that univariate analyzes are insufficient, since these differences can only be minimized or corrected with robust statistical models.
In conclusion, our proposed model using four RLBs, easy to obtain, could be used daily without extra cost (creatinine, PLT, RBC and MCH) and can be an early warning system in at least 13 h to detect the bacterial group of l ‘etiology. officer in charge of the BSIs, via a simple computer system or even a mobile phone application. We believe that the association of this MR with the patient’s clinical data could increase the chances of success in the search for the bacterial group involved in BSI, and thus help in the management of antibiotic therapy in a more precise manner. It is also important to add that the use of these models can help decision-making for empirical therapy without ever forgetting that the empirical prescription of antimicrobials must also take into account the specificities of each patient.