Predicting the causative pathogen among children with pneumonia using a causal Bayesian network

• DOI: 10.1371/journal.pcbi.1010967
• Publication Date: 2023-03-13T17:39:59Z
• Authors: Yue Wu; Steven Mascaro; Mejbah Bhuiyan; Parveen Fathima; Ariel O. Mace; Mark P. Nicol; Peter C. Richmond; Lea-Ann Kirkham; Michael Dymock; David A. Foley; Charlie McLeod; Meredith L. Borland; Andrew Martin; Phoebe C. M. Will...; Julie A. Marsh; Thomas L. Snelling; Christopher C. Blyth
• Keywords: 03 medical and health sciences; 0302 clinical medicine; QH301-705.5; Surveys and Questionnaires; Australia; Humans; Bayes Theorem; Pneumonia; Biology (General); Research Article; Anti-Bacterial Agents; 3. Good health

Pneumonia remains a leading cause of hospitalization and death among young children worldwide, the diagnostic challenge differentiating bacterial from non-bacterial pneumonia is main driver antibiotic use for treating in children. Causal Bayesian networks (BNs) serve as powerful tools this problem they provide clear maps probabilistic relationships between variables produce results an explainable way by incorporating both domain expert knowledge numerical data.We used data combination iteratively, to construct, parameterise validate causal BN predict causative pathogens childhood pneumonia. Expert elicitation occurred through series group workshops, surveys one-on-one meetings involving 6-8 experts diverse areas. The model performance was evaluated based on quantitative metrics qualitative validation. Sensitivity analyses were conducted investigate how target output influenced varying key assumptions particularly high degree uncertainty around or knowledge.Designed apply cohort with X-ray confirmed who presented tertiary paediatric hospital Australia, resulting offers predictions range interest, including diagnosis pneumonia, detection respiratory nasopharynx, clinical phenotype episode. Satisfactory numeric has been achieved area under receiver operating characteristic curve 0.8 predicting clinically-confirmed sensitivity 88% specificity 66% given certain input scenarios (i.e., information that available entered into model) trade-off preferences relative weightings consequences false positive versus negative predictions). We specifically highlight desirable threshold practical very dependent upon different preferences. Three commonly encountered demonstrate potential usefulness outputs various pictures.To our knowledge, first developed help determine pathogen have shown method works it would decision making antibiotics, providing insight computational may be translated actionable decisions practice. discussed next steps external validation, adaptation implementation. Our framework methodological approach can adapted beyond context broad infections geographical healthcare settings.