The nasal microbiota in the development of a multi-parametric prediction model to differentiate bacterial versus viral infections in lower respiratory tract infections

Background: The ability to rapidly and accurately determine 'bacterial infection' from 'viral infection' in cases of suspected lower respiratory tract infections would help clinicians better target antimicrobial therapy. Further, technological developments make the rapid determination of an individual's microbiota now feasible, such that patient-specific microbiota profiles may rapidly be generated and integrated with personalized clinical data. However, evidence is required to show the clinical value of such combined personalized data in the (rapid) diagnosis of infections. Therefore, as part of a European Union-funded study (TAILORED-Treatment, Grant ID 602860), we determined if the addition of nasal cavity microbiota profiles increased the predictive value of a machine learning classifier developed to distinguish between bacterial and viral infection in patients with lower respiratory tract infections. Results: Three machine learning prediction models were evaluated on 293 patients with diagnostic class labels of 'viral' or 'bacterial' infection, as determined by an international expert panel. 29 clinical variables and 7 genus-level nasal cavity microbiota variables were selected to be input into our Random Forests prediction models. The best predictor when used alone was C-reactive protein (CRP), reaching an AUC of 0.75. When adding the 7 most common microbiota variables to the CRP model, the average AUC showed a marginal improvement of 6%. The best model yielded an average AUC of 0.90 and correctly predicted 85% of the 'bacterial' patients and 82% of the 'viral' patients in the 5-fold cross-validation test sets. This model included 13 clinical variables and 3 genus-level nasal cavity microbiota variables (i.e. Staphylococcus, Moraxella, and Streptococcus). However, simply adding absolute neutrophil count (ANC), consolidation on X-ray, and age group to CRP results (a total of 4 variables) significantly improved the prediction model towards an average AUC of 0.85.Conclusion: The development of rapid microbiota sequencing techniques and predictive modelling means that microbiota profiling could become a useful aid in facilitating rapid, personalized and accurate clinical decision making. In this respect, our results indicate that nasal cavity microbiota profiling may not add significant value to diagnostic algorithms that aim to differentiate between bacterial and viral infections in both children and adult patients with LRTI.