BACKGROUND: We analyzed data from the Department of Veterans Affairs trial of steroid therapy for systemic sepsis to identify predictors of bacteremia and gram-negative bacteremia. METHODS: Of the 2568 patients screened for entry in the trial, 465 met the following criteria: presence of four of seven clinical signs of sepsis; blood cultures at the time of screening; and complete data on nine clinical parameters. The multivariate logistic regression model was used to identify predictors of bacteremia and gram-negative bacteremia. Predicted probabilities of having these types of infections were calculated using the identified predictors. Patients were then classified into groups with and without bacteremia (and gram-negative bacteremia) based on the predicted probability. Misclassification error rates were calculated for each method of categorization by comparing the true with the predicted grouping of patients. RESULTS: Three factors were independently predictive of bacteremia and gram-negative bacteremia: elevated temperature, low systolic blood pressure, and low platelet count. Using these three factors, classification methods were identified that predicted blood infection better than chance, but misclassification was also high. For predicting bacteremia, the maximum predicted positive rate was 83%, with a specificity of nearly 100% and a sensitivity of only 5%. For predicting gram-negative bacteremia, the maximum predicted positive accuracy was 100%, with a specificity also of 100% and a sensitivity of almost 0%. CONCLUSIONS: Using simple clinical parameters, we could not predict either bacteremia or gram-negative bacteremia with sufficient accuracy to be clinically meaningful; however, our approach represents a step in the direction of forecasting the bacterial organism responsible for sepsis in advance of culture results.
RCT Entities:
BACKGROUND: We analyzed data from the Department of Veterans Affairs trial of steroid therapy for systemic sepsis to identify predictors of bacteremia and gram-negative bacteremia. METHODS: Of the 2568 patients screened for entry in the trial, 465 met the following criteria: presence of four of seven clinical signs of sepsis; blood cultures at the time of screening; and complete data on nine clinical parameters. The multivariate logistic regression model was used to identify predictors of bacteremia and gram-negative bacteremia. Predicted probabilities of having these types of infections were calculated using the identified predictors. Patients were then classified into groups with and without bacteremia (and gram-negative bacteremia) based on the predicted probability. Misclassification error rates were calculated for each method of categorization by comparing the true with the predicted grouping of patients. RESULTS: Three factors were independently predictive of bacteremia and gram-negative bacteremia: elevated temperature, low systolic blood pressure, and low platelet count. Using these three factors, classification methods were identified that predicted blood infection better than chance, but misclassification was also high. For predicting bacteremia, the maximum predicted positive rate was 83%, with a specificity of nearly 100% and a sensitivity of only 5%. For predicting gram-negative bacteremia, the maximum predicted positive accuracy was 100%, with a specificity also of 100% and a sensitivity of almost 0%. CONCLUSIONS: Using simple clinical parameters, we could not predict either bacteremia or gram-negative bacteremia with sufficient accuracy to be clinically meaningful; however, our approach represents a step in the direction of forecasting the bacterial organism responsible for sepsis in advance of culture results.
Authors: A Gikas; G Samonis; A Christidou; J Papadakis; D Kofteridis; Y Tselentis; N Tsaparas Journal: Infection Date: 1998 May-Jun Impact factor: 3.553