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

Area of Science:

• Computational epidemiology
• Pediatric infectious diseases
• Bayesian network modeling

Background:

• Childhood pneumonia is a major global health concern, leading to significant hospitalizations and deaths.
• Accurate differentiation between bacterial and non-bacterial pneumonia is crucial for appropriate antibiotic prescription.
• Causal Bayesian networks (BNs) offer a robust framework for modeling complex probabilistic relationships in diagnostics.

Purpose of the Study:

• To construct and validate a causal BN for predicting causative pathogens in childhood pneumonia.
• To provide explainable and quantitative predictions to aid in clinical decision-making regarding antibiotic use.
• To develop a tool that integrates expert knowledge and data for improved pneumonia diagnosis.

Main Methods:

• Iterative construction and validation of a causal BN using domain expert knowledge and clinical data.
• Expert knowledge elicitation through workshops, surveys, and one-on-one meetings with 6-8 specialists.
• Model performance evaluation using quantitative metrics and qualitative expert validation, including sensitivity analyses.

Main Results:

• The developed BN accurately predicts clinically-confirmed bacterial pneumonia with an area under the ROC curve of 0.8.
• Achieved 88% sensitivity and 66% specificity in predicting bacterial pneumonia under specific input scenarios.
• Demonstrated the model's utility in various clinical scenarios, highlighting the impact of input data and trade-off preferences.

Conclusions:

• This is the first causal model designed to identify causative pathogens for pediatric pneumonia.
• The BN framework provides actionable insights for antibiotic decision-making in clinical practice.
• The model and methodology are adaptable for broader respiratory infections across different settings.