ABSTRACT
BACKGROUND: In 2015, pneumonia remained the leading cause of mortality in children aged 1-59 months. METHODS: Data from 1802 human immunodeficiency virus (HIV)-negative children aged 1-59 months enrolled in the Pneumonia Etiology Research for Child Health (PERCH) study with severe or very severe pneumonia during 2011-2014 were used to build a parsimonious multivariable model predicting mortality using backwards stepwise logistic regression. The PERCH severity score, derived from model coefficients, was validated on a second, temporally discrete dataset of a further 1819 cases and compared to other available scores using the C statistic. RESULTS: Predictors of mortality, across 7 low- and middle-income countries, were age <1 year, female sex, ≥3 days of illness prior to presentation to hospital, low weight for height, unresponsiveness, deep breathing, hypoxemia, grunting, and the absence of cough. The model discriminated well between those who died and those who survived (C statistic = 0.84), but the predictive capacity of the PERCH 5-stratum score derived from the coefficients was moderate (C statistic = 0.76). The performance of the Respiratory Index of Severity in Children score was similar (C statistic = 0.76). The number of World Health Organization (WHO) danger signs demonstrated the highest discrimination (C statistic = 0.82; 1.5% died if no danger signs, 10% if 1 danger sign, and 33% if ≥2 danger signs). CONCLUSIONS: The PERCH severity score could be used to interpret geographic variations in pneumonia mortality and etiology. The number of WHO danger signs on presentation to hospital could be the most useful of the currently available tools to aid clinical management of pneumonia.