Pneumonia is a major cause of illness and mortality in children under 5, especially in low-resource settings (GBD, 2019). The World Health Organisation (WHO) provides clinical guidelines for severity in order to help manage childhood pneumonia cases; however, recent research has highlighted changes in the markers of severity over time, suggesting that the efficacy of these guidelines may be reduced (Agweyu, 2017). In the Biotope study, in which the project supervisor is a co-investigator, clinicians and scientists have developed a machine learning model to predict the severity of pneumonia from clinical variables that are easily accessible in a community healthcare setting in Malawi (manuscript under revision; cohort described in Gallagher et al. 2021). Among the variables used to predict severity are the concern of clinicians and caregivers, who each provided a rating of how concerned they were about the child’s condition. These concern ratings are important because they can indicate early signs of severe pneumonia. However, clinicians and caregivers may base their concern on different clinical features. In this project we are interested in understanding the differences between the clinical variables that cause concern for caregivers and clinicians. This is relevant both for optimising the machine learning model itself and for gaining an improved general understanding of differences in how caregivers and clinicians respond to warning signs of disease. The goal of the internship is to develop visualisations that will help researchers to understand the relative importance of different predictors of clinician and caregiver concern. This will aid the model interpretability and make complex model outputs easier to understand by translating them into graphical representations. The expectation is that these visualizations will highlight similarities and differences in the factors that influence concern among clinicians and caregivers.
