Development and validation of a novel clinical risk score to predict hypoxemia in children with pneumonia using the WHO PREPARE dataset

Background Hypoxemia predicts mortality at all levels of care, and appropriate management can reduce preventable deaths. However, pulse oximetry and oxygen therapy remain inaccessible in many primary care health facilities. We aimed to develop and validate a simple risk score comprising commonly evaluated clinical features to predict hypoxemia in 2-59-month-old children with pneumonia. Methods Data from 7 studies conducted in 5 countries from the Pneumonia Research Partnership to Assess WHO Recommendations (PREPARE) dataset were included. Readily available clinical features and demographic variables were used to develop a multivariable logistic regression model to predict hypoxemia (SpO2<90%) at presentation to care. The adjusted log coefficients were transformed to derive the PREPARE hypoxemia risk score and its diagnostic value was assessed in a held-out, temporal validation dataset. Results We included 14,509 children in the analysis; 9.8% (n=2,515) were hypoxemic at presentation. The multivariable regression model to predict hypoxemia included age, sex, respiratory distress (nasal flaring, grunting and/or head nodding), lower chest indrawing, respiratory rate, body temperature and weight-for-age z-score. The model showed fair discrimination (area under the curve 0.70, 95% CI 0.67 to 0.73) and calibration in the validation dataset. The simplified PREPARE hypoxemia risk score includes 5 variables: age, respiratory distress, lower chest indrawing, respiratory rate and weight-for-age z-score. Conclusion The PREPARE hypoxemia risk score, comprising five easily available characteristics, can be used to identify hypoxemia in children with pneumonia with a fair degree of certainty for use in health facilities without pulse oximetry. Its implementation would require careful consideration to limit inappropriate referrals on patients and the health system. Further external validation in community settings in low- and middle-income countries is required.