Refinement and Validation of a Clinical-Based Approach to Evaluate Young Febrile Infants

For febrile infants, predictive models to detect bacterial infections are available, but clinical adoption remains limited by implementation barriers. There is a need for predictive models using widely available predictors. Thus, we previously derived 2 novel predictive models (machine learning and regression) by using demographic and clinical factors, plus urine studies. The objective of this study is to refine and externally validate the predictive models.

This is a cross-sectional study of infants initially evaluated at one pediatric emergency department from January 2011 to December 2018. Inclusion criteria were age 0 to 90 days, temperature ≥38°C, documented gestational age, and insurance type. To reduce potential biases, we derived models again by using derivation data without insurance status and tested the ability of the refined models to detect bacterial infections (ie, urinary tract infection, bacteremia, and meningitis) in the separate validation sample, calculating areas-under-the-receiver operating characteristic curve, sensitivities, and specificities.

Of 1419 febrile infants (median age 53 days, interquartile range = 32-69), 99 (7%) had a bacterial infection. Areas-under-the-receiver operating characteristic curve of machine learning and regression models were 0.92 (95% confidence interval [CI] 0.89-0.94) and 0.90 (0.86-0.93) compared with 0.95 (0.91-0.98) and 0.96 (0.94-0.98) in the derivation study. Sensitivities and specificities of machine learning and regression models were 98.0% (94.7%-100%) and 54.2% (51.5%-56.9%) and 96.0% (91.5%-99.1%) and 50.0% (47.4%-52.7%).

Compared with the derivation study, the machine learning and regression models performed similarly. Findings suggest a clinical-based model can estimate bacterial infection risk. Future studies should prospectively test the models and investigate strategies to optimize clinical adoption.