Abstract
BACKGROUND: Respiratory pathogens are a common cause of disease. Currently there is not a practical tool to predict the putative etiology of each case with an inexpensive, fast point-of-care assay. Here, we describe a decision support tool that enables the prediction of both bacterial and viral respiratory pathogen infections in a single patient, using a Machine Learning model. METHODS: The data were obtained from the Hadassah-Hebrew University Medical Center during a period of 10 years beginning from 2007 and contained more than 40,000 patients from a 1,000,000-population community for whom specimens were tested by either PCR or culture. The pathogens included were, H. influenzae; M. catarrhalis; S. pneumoniae; M. pneumoniae; Adenovirus; Human metapneumovirus; Influenza H1N1, A, B; parainfluenza 1,2 and 3; and RSV. We then created a Machine-Learning algorithm to simulate the spread of infection in the entire Jerusalem area. We defined transmission areas based on geographical distances of patients’ home-addresses. Then we prospectively tested the tool accuracy over a 4-month period, in addition to real-time improvement of the model. RESULTS: Initial model was created based on gender, age, home addresses and the diagnostics test results. We then reconstructed a putative spread pattern for each of the pathogens that can be correlated to potential “transmission routes.” The initial prediction tool had an AUC for most pathogens around 0.85. It ranged from 0.75 to 0.8 for the bacterial and 0.82 to 0.89 for the viral pathogens. In almost all pathogens the NPV was 0.98–0.99. We then tested the decision support tool prospectively over four consecutive months (January to April 2019—1,700 patients with respiratory complaints from whom samples were sent to the lab). While the AUC in the prospective cohort was 0.81 on average, the NPV remained high on 0.98. CONCLUSION: The implementation of the decision support tool on respiratory pathogen diagnostics enables better prediction of patients not infected with either viral or bacterial pathogens. The use of such a tool can save more than 50% of diagnostic tests expenses as well as real-time mapping of disease spread. Improvement of the Machine Learning protocol may further promote the optimization of positive predictive values. DISCLOSURES: All authors: No reported disclosures.