Improved Prediction of Invasive Bacterial Infections in Febrile Infants Using Machine Learning

Existing guidelines for febrile infants aged 8 to 60 days use clinical appearance, age, and laboratory test results to assess the risk of invasive bacterial infections (IBIs) with variable performance. This study aimed to (1) develop and validate 5 supervised learning models (SLMs) capable of predicting IBIs in febrile infants aged 8 to 60 days and (2) compare diagnostic performance between the SLMs and the 2021 American Academy of Pediatrics Clinical Practice Guideline (AAP CPG).

This single-center, retrospective cohort study included febrile infants (≥38 °C) aged 8 to 60 days who presented between January 1, 2018, and December 31, 2023. Demographics, procalcitonin levels, complete blood count indices, urinalysis, viral test results, and blood/cerebrospinal culture results were collected. IBI was defined as growth in blood/cerebrospinal cultures. Derivation (80%) and validation (20%) data sets were created using proportionate stratified random sampling. Univariable logistic regression identified potential predictors. Five SLMs were trained on the derivation set. Internal validation was performed using the validation set.

A total of 3682 infants were screened, and 2057 (56%) met inclusion criteria; 55 (3%) had an IBI. Procalcitonin, absolute neutrophil count, viral status, and positive urinalysis results were included in all models. Internal validation demonstrated that SLMs retained similar sensitivity to the AAP CPG (92%-100%) but significantly greater specificity (81%-89% vs 60%).

In febrile infants aged 8 to 60 days, our SLMs had comparable sensitivity but greater specificity than the AAP CPG. These models show promise as clinical decision support for early identification of febrile infants at risk for IBI. Future studies are needed to externally validate the models and investigate whether one SLM exhibits the greatest performance.