Development and validation of a risk score for predicting positivity of blood cultures and mortality in patients with bacteremia and fungemia

Abstract

Introduction Bacteremia is a major cause of morbidity and mortality in hospitalized patients. Predictors of mortality are critical for the management and survival of hospitalized patients. The objective of this study was to determine the factors related to blood culture positivity and the risk factors for mortality in patients whose blood cultures were collected. Methods A prospective 2-cohort study (derivation with 784 patients and validation with 380 patients) based on the Pitt bacteremia score for all patients undergoing blood culture collection. The score was obtained from multivariate analysis. The Kaplan-Meier survival curve of the cohort derivation and the cohort validation groups was calculated, and the difference was assessed using a log-rank test. Mortality-related factors were older age, extended hospitalization, > 10% of immature cells in the leukogram, lower mean blood pressure, elevated heart rate, elevated WBC count, and elevated respiratory rate. These continuous variables were dichotomized according to their significance level, and a cut-off limit was created. ResultsThe area under the ROC curve (AUC) was 0.789. The score was validated in a group of 380 patients who were prospectively evaluated. Conclusion Prolonged hospitalization, body temperature, and elevated heart rate were related to positive blood cultures. The Pitt score can be used to assess the risk of death; however it can be individualized according to the epidemiology of each hospital.

Sequences of Events from the Electronic Medical Record and the Onset of Infection

“…[3] Such analyses have been increasingly implemented to predict sepsis, [4] cardiac arrest, [4] acute kidney injury, [5] clinical decompensation, [6] and inhospital mortality. [6][7][8] Although recent studies have proposed methods to predict hospital-acquired infection from EHR data, these implemented algorithms, including random forests, [9] multivariate survival analysis, [10] logistic regression, [11] and gradient boosting models, [9,11] have remained retrospective and static in approach. Our method presented in this article represents a novel contribution to analyzing timeseries critical care data.…”