Smart diagnostic algorithms for automated detection of childhood pneumonia in resource-constrained settings

Pneumonia is the leading cause of death in children under five, with 1.1 million deaths annually more than the combined burden of HIV/AIDS, malaria, and tuberculosis for this age group; the majority of these deaths occur in resource-constrained settings. Accurate diagnosis of pneumonia relies on expensive human expertise and requires the evaluation of multiple clinical characteristics, measured using advanced diagnostic tools. The shortage of clinical experts and appropriate diagnostic tools in many low and middle income countries impedes timely and accurate diagnosis. We demonstrate that the diagnostic process can be automated using machine learning techniques, processing several clinical measurements that could be obtained with affordable and easy-to-operate point-of-care tools. We evaluated our findings on a dataset of 1093 children, comprising 777 diagnosed with pneumonia and 316 healthy controls, on the basis of 47 clinical characteristics. Seven feature selection techniques were used to identify robust, parsimonious subsets of clinical characteristics, which could be measured reliably and affordably. Standard machine learning techniques, such as support vector machines and random forests, were used to develop a predictive algorithm based on the four jointly most predictive characteristics (temperature, respiratory rate, heart rate and oxygen saturation); this approach led to 96.6% sensitivity, 96.4% specificity, and an Area Under the Curve (AUC) of 97.8%. The proposed approach can be easily embedded in a mobile phone application, allowing for point-of-care assessment and identification of children in need of clinical attention by basically trained healthcare workers in resource-constrained settings.