Routine laboratory biomarkers used to predict Gram-positive or Gram-negative bacteria involved in bloodstream infections

Abstract

This study evaluated routine laboratory biomarkers (RLB) to predict the infectious bacterial group, Gram-positive (GP) or Gram-negative (GN) associated with bloodstream infection (BSI) before the result of blood culture (BC). A total of 13,574 BC of 6787 patients (217 BSI-GP and 238 BSI-GN) and 68 different RLB from these were analyzed. The logistic regression model was built considering BSI-GP or BSI-GN as response variable and RLB as covariates. After four filters applied total of 320 patients and 16 RLB remained in the Complete-Model-CM, and 4 RLB in the Reduced-Model-RM (RLB p > 0.05 excluded). In the RM, only platelets, creatinine, mean corpuscular hemoglobin and erythrocytes were used. The reproductivity of both models were applied to a test bank of 2019. The new model presented values to predict BSI-GN of the area under the curve (AUC) of 0.72 and 0.69 for CM and RM, respectively; with sensitivity of 0.62 and 0.61 (CM and RM) and specificity of 0.67 for both. These data confirm the discriminatory capacity of the new models for BSI-GN (p = 0.64). AUC of 0.69 using only 4 RLB, associated with the patient's clinical data could be useful for better targeted antimicrobial therapy in BSI.

Figure 1

(A) Criterium for exclusion of patients; (B) exclusion criterium for routine laboratory biomarkers (RLB). F1 = first filtering process variables with different quartiles were kept in the bank at 0%, 25%, 50%, 75% and 100%. F2 = second filtering process, Pearson's correlation coefficient of 0.8 in absolute value was adopted as the cutoff. F3 = third filtering process, variables with 30% or more of lost values and/or zeros were removed. F4 = fourth filtering process, The Kolmogorov–Smirnov tests, the t-Student test and the Wilcoxon-Mann–Whitney test were used and variables with p ≤ 0.1 in at least one of the tests were maintained.

Figure 2

Comparisons of levels of single biomarkers stratified by patient categories. GN = Gram-negative (red); GP = Gram-positive (blue); PLT = platelet count; RBC = red blood cell count; MCH = mean corpuscular haemoglobin.

Figure 3

Comparison between the Complete and Reduced Models to predict Gram-negative bacteremia. AUC = area under curve. Complete Model = Monocytes%, Bands (/mm³), p50 (mmHg); Mean Corpuscular Haemoglobin (CHCM) (g/dl), Bands%, Age, Monocytes (/mm³), Hydrogen potential (pH), total CO₂ (mmol/L), Methaemoglobin (%), ionized Calcium (iCa—mg/dL), Lactate (mmol/L), platelet count (PLT—10³/mm³), Creatinine (mg/dl), Mean corpuscular haemoglobin (MCH—fl) and Red blood cell count (RBC—millions/μL). Reduced Model = PLT, creatinine, MCH and RBC. Blue: Validation (n = 320). Red: Prediction using the 2019 database (n = 69).