JAC Advance Access originally published online on April 4, 2008
Journal of Antimicrobial Chemotherapy 2008 62(1):184-188; doi:10.1093/jac/dkn141
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Original research |
Predicting pathogens causing ventilator-associated pneumonia using a Bayesian network model
1 Department of Internal Medicine and Infectious Diseases, University Medical Center Utrecht, Utrecht, The Netherlands 2 Institute for Computing and Information Sciences, Radboud University, Nijmegen, The Netherlands 3 Department of Medical Microbiology, University Medical Center Utrecht, Utrecht, The Netherlands 4 Julius Center for Health Sciences and Primary Care, University Medical Center Utrecht, Utrecht, The Netherlands
Received 30 January 2008; returned 2 March 2008; revised 7 March 2008; accepted 10 March 2008
* Correspondence address. Department of Medical Microbiology, University Medical Center Utrecht, Utrecht, The Netherlands. Tel: +31-88-755-7676; Fax: +31-30-252-3741; E-mail: mbonten{at}umcutrecht.nl
Background: We previously validated a Bayesian network (BN) model for diagnosing ventilator-associated pneumonia (VAP). Here, we report on the performance of the model to predict microbial causes of VAP and to select antibiotics.
Methods: Pathogens were grouped into seven categories based upon the antibiotic susceptibility and epidemiological characteristics. Colonization of the upper respiratory tract was modelled in the BN and depended—in additional steps—on (i) duration of admission and ventilation, (ii) previous culture results and (iii) previous antibiotic use. A database with 153 VAP episodes and their microbial causes was used as reference standard. Appropriateness of antibiotic prescription, with fixed choices for pathogens predicted, was determined.
Results: One hundred and seven VAP episodes were monobacterial and 46 were caused by two pathogens. Using duration of admission and ventilation only, areas under the receiver operating curve (AUC) ranged from 0.511 to 0.772 for different pathogen groups, and model predictions significantly improved when adding information on culture results, but not when adding information on antibiotic use. The best performing model (with all information) had AUC values ranging from 0.859 for Acinetobacter spp. to 0.929 for Streptococcus pneumoniae. With this model, 91 (85%) and 29 (63%) of all pathogen groups were correctly predicted for monobacterial and polymicrobial VAP, respectively. With fixed antibiotic choices linked to pathogen groups, 92% of all episodes would have been treated appropriately.
Conclusions: The BN models' performance to predict pathogens causing VAP improved markedly with information on colonization, resulting in excellent pathogen prediction and antibiotic selection. Prospective external validation is needed.
Keywords: decision support , ICUs , appropriate antibiotic treatment