Validation of risk prediction for outcomes of severe community-acquired pneumonia among under-five children in Amhara region, Northwest Ethiopia

Background Globally there are over 1,400 cases of pneumonia per 100,000 children every year, where children in South Asia and Sub-Saharan Africa are disproportionately affected. Some of the cases develop poor treatment outcome (treatment failure or antibiotic change or staying longer in the hospital or death), while others develop good outcome during interventions. Although clinical decision-making is a key aspect of the interventions, there are limited tools such as risk scores to assist the clinical judgment in low-income settings. This study aimed to validate a prediction model and develop risk scores for poor outcomes of severe community-acquired pneumonia (SCAP). Methods A cohort study was conducted among 539 under-five children hospitalized with SCAP. Data analysis was done using R version 4.0.5 software. A multivariable analysis was done. We developed a simplified risk score to facilitate clinical utility. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC) and calibration plot. Bootstrapping was used to validate all accuracy measures. A decision curve analysis was used to evaluate the clinical and public health utility of our model. Results The incidence of poor outcomes of pneumonia was 151(28%) (95%CI: 24.2%-31.8%). Vaccination status, fever, pallor, unable to breastfeed, impaired consciousness, CBC abnormal, entered ICU, and vomiting remained in the reduced model. The AUC of the original model was 0.927, 95% (CI (0.90, 0.96), whereas the risk score model produced prediction accuracy of an AUC of 0.89 (95%CI: 0.853–0.922. Our decision curve analysis for the model provides a higher net benefit across ranges of threshold probabilities. Conclusions Our model has excellent discrimination and calibration performance. Similarly, the risk score model has excellent discrimination and calibration ability with an insignificant loss of accuracy from the original. The models can have the potential to improve care and treatment outcomes in the clinical settings.