Validation of Prediction Models for Pneumonia Among Children in the Emergency Department

Abstract

Background: Several prediction models have been reported to identify patients with radiographic pneumonia, but none have been validated or broadly implemented into practice. We evaluated 5 prediction models for radiographic pneumonia in children.

Methods: We evaluated 5 previously published prediction models for radiographic pneumonia (Neuman, Oostenbrink, Lynch, Mahabee-Gittens, and Lipsett) using data from a single-center prospective study of patients 3 months to 18 years with signs of lower respiratory tract infection. Our outcome was radiographic pneumonia. We compared each model's area under the receiver operating characteristic curve (AUROC) and evaluated their diagnostic accuracy at statistically-derived cutpoints.

Results: Radiographic pneumonia was identified in 253 (22.2%) of 1142 patients. When using model coefficients derived from the study dataset, AUROC ranged from 0.58 (95% confidence interval, 0.52-0.64) to 0.79 (95% confidence interval, 0.75-0.82). When using coefficients derived from original study models, 2 studies demonstrated an AUROC >0.70 (Neuman and Lipsett); this increased to 3 after deriving regression coefficients from the study cohort (Neuman, Lipsett, and Oostenbrink). Two models required historical and clinical data (Neuman and Lipsett), and the third additionally required C-reactive protein (Oostenbrink). At a statistically derived cutpoint of predicted risk from each model, sensitivity ranged from 51.2% to 70.4%, specificity 49.9% to 87.5%, positive predictive value 16.1% to 54.4%, and negative predictive value 83.9% to 90.7%.

Conclusions: Prediction models for radiographic pneumonia had varying performance. The 3 models with higher performance may facilitate clinical management by predicting the risk of radiographic pneumonia among children with lower respiratory tract infection.