Risk prediction model for distinguishing Gram-positive from Gram-negative bacteremia based on age and cytokine levels: A retrospective study

BACKGROUND: Severe infection often results in bacteremia, which significantly increases mortality rate. Different therapeutic strategies are employed depending on whether the blood-borne infection is Gram-negative (G(-)) or Gram-positive (G(+)). However, there is no risk prediction model for assessing whether bacteremia patients are infected with G(-) or G(+) pathogens. AIM: To establish a clinical prediction model to distinguish G(-) from G(+) infection. METHODS: A total of 130 patients with positive blood culture admitted to a single intensive care unit were recruited, and Th1 and Th2 cytokine concentrations, routine blood test results, procalcitonin and C-reactive protein concentrations, liver and kidney function test results and coagulation function were compared between G(+) and G(-) groups. Least absolute shrinkage and selection operator (LASSO) regression analysis was employed to optimize the selection of predictive variables by running cyclic coordinate descent and K-fold cross-validation (K = 10). The predictive variables selected by LASSO regression analysis were then included in multivariate logistic regression analysis to establish a prediction model. A nomogram was also constructed based on the prediction model. Calibration chart, receiver operating characteristic curve and decision curve analysis were adopted for validating the prediction model. RESULTS: Age, plasma interleukin 6 (IL-6) concentration and plasma aspartate aminotransferase concentration were identified from 57 measured variables as potential factors distinguishing G(+) from G(-) infection by LASSO regression analysis. Inclusion of these three variables in a multivariate logistic regression model identified age and IL-6 as significant predictors. In receiver operating characteristic curve analysis, age and IL-6 yielded an area under the curve of 0.761 and distinguished G(+) from G(-) infection with specificity of 0.756 and sensitivity of 0.692. Serum IL-6 and IL-10 levels were upregulated by more than 10-fold from baseline in the G(-) bacteremia group but by less than ten-fold in the G(+) bacteremia group. The calibration curve of the model and Hosmer-Lemeshow test indicated good model fit (P > 0.05). When the decision curve analysis curve indicated a risk threshold probability between 0% and 68%, a nomogram could be applied in clinical settings. CONCLUSION: A simple prediction model distinguishing G(-) from G(+) bacteremia can be constructed based on reciprocal association with age and IL-6 level.