What is it about?
Pneumonia is a leading cause of hospital visits among young children. It can be very hard for doctors to tell which pneumonias are caused by bacteria and which are caused by viral infections. But doctors need to make decisions quickly, like whether to give antibiotics, balancing the risk of clinical deterioration if appropriate treatment is delayed against the consequences of prescribing antibiotics when they are not required. This paper shows how this clinical challenge has been addressed by a cross-disciplinary team of mathematical modelers and medical experts, through the creation of a causal Bayesian network (BN), a type of explainable artificial intelligence (AI). The BN produces reliable and explainable predictions on things that doctors need to know when they make clinical decisions; e.g., whether bacteria might have caused a pneumonia episode, and therefore whether antibiotic should be used. The authors give examples to demonstrate how the BN might be used by doctors to help them make better decisions in clinical care.
Featured Image
Photo by Neil Thomas on Unsplash
Why is it important?
• The diagnostic challenge of differentiating bacterial from viral infections, and fear of the consequences of failing to treat bacterial pneumonia, are the main drivers of antibiotic use for treating pneumonia. The published BN is the first causal model that aims to help determine the pathogens responsible for a child's pneumonia, offering the opportunity to inform better clinical decision making and patient outcomes, as well as improved quality in the use of antibiotics. • We have established a modelling framework and a methodological approach that can be adapted beyond our context to respiratory tract infections more broadly, and other geographical and healthcare settings. • This work has also offered insights into how computational predictions may be translated into decisions in practice. In particular, how trade-offs can play an important role in the way clinicians might use the predictions, e.g., the extent to which clinicians accept false positives (i.e., over-treating infections that are not bacterial with antibiotics) in order to avoid false negatives (i.e., missing or delaying antibiotic therapy for bacterial infections).
Perspectives
Recent advances in AI offer plenty of opportunities to improve healthcare, but there are concerns about using them to support decisions in clinical practice. Understandably, clinicians are reluctant to accept predictions made by ‘black box’ algorithms that cannot be easily understood. Building on several years of work with clinical experts and target end users, our paper shows how AI researchers and clinicians can build a causal BN together that explicitly depicts the hidden processes underlying the disease, and ties it to the typically available tests, diagnostics and management available for pneumonia. Ultimately, this means our causal AI offers transparent, understandable predictions that can be readily monitored and challenged and can therefore genuinely earn the trust of clinicians.
Dr Yue Wu
University of Sydney
Read the Original
This page is a summary of: Predicting the causative pathogen among children with pneumonia using a causal Bayesian network, PLoS Computational Biology, March 2023, PLOS,
DOI: 10.1371/journal.pcbi.1010967.
You can read the full text:
Contributors
The following have contributed to this page
