Send to

Choose Destination
Clin Microbiol Infect. 2019 Jan;25(1):108.e1-108.e7. doi: 10.1016/j.cmi.2018.03.029. Epub 2018 Apr 25.

Decision-support models for empiric antibiotic selection in Gram-negative bloodstream infections.

Author information

Division of Infectious Diseases, University of Toronto, Canada. Electronic address:
Division of Infectious Diseases, University of Toronto, Canada.
Division of Infectious Diseases, NorthShore University Health Systems, Chicago, IL, USA.
Critical Care and Population Health, Providence St. Joseph Health, Seattle, Washington, USA.
Toronto General Hospital Research Institute, University of Toronto, Canada; Dalla Lana School of Public Health, University of Toronto, Canada.
Department of Pharmacy, Sunnybrook Health Sciences Centre, Toronto, Canada.
Division of Infectious Diseases, University of Toronto, Canada; Sunnybrook Research Institute, Sunnybrook Health Sciences Centre, University of Toronto, Canada.



Early empiric antibiotic therapy in patients can improve clinical outcomes in Gram-negative bacteraemia. However, the widespread prevalence of antibiotic-resistant pathogens compromises our ability to provide adequate therapy while minimizing use of broad antibiotics. We sought to determine whether readily available electronic medical record data could be used to develop predictive models for decision support in Gram-negative bacteraemia.


We performed a multi-centre cohort study, in Canada and the USA, of hospitalized patients with Gram-negative bloodstream infection from April 2010 to March 2015. We analysed multivariable models for prediction of antibiotic susceptibility at two empiric windows: Gram-stain-guided and pathogen-guided treatment. Decision-support models for empiric antibiotic selection were developed based on three clinical decision thresholds of acceptable adequate coverage (80%, 90% and 95%).


A total of 1832 patients with Gram-negative bacteraemia were evaluated. Multivariable models showed good discrimination across countries and at both Gram-stain-guided (12 models, areas under the curve (AUCs) 0.68-0.89, optimism-corrected AUCs 0.63-0.85) and pathogen-guided (12 models, AUCs 0.75-0.98, optimism-corrected AUCs 0.64-0.95) windows. Compared to antibiogram-guided therapy, decision-support models of antibiotic selection incorporating individual patient characteristics and prior culture results have the potential to increase use of narrower-spectrum antibiotics (in up to 78% of patients) while reducing inadequate therapy.


Multivariable models using readily available epidemiologic factors can be used to predict antimicrobial susceptibility in infecting pathogens with reasonable discriminatory ability. Implementation of sequential predictive models for real-time individualized empiric antibiotic decision-making has the potential to both optimize adequate coverage for patients while minimizing overuse of broad-spectrum antibiotics, and therefore requires further prospective evaluation.


Readily available epidemiologic risk factors can be used to predict susceptibility of Gram-negative organisms among patients with bacteraemia, using automated decision-making models.


Antibiotic resistance; Antimicrobial resistance; Antimicrobial-resistant organisms (ARO); Clinical decision-making; Decision support; Predictive models


Supplemental Content

Full text links

Icon for Elsevier Science
Loading ...
Support Center