Abstract
Early recognition of bloodstream infection (BSI) in infants can be difficult, as symptoms may be non-specific, and culture can take up to 48 h. As a result, many infants receive unneeded antibiotic treatment while awaiting the culture results. In this study, we aimed to develop a model that can reliably identify infants who do not have positive blood cultures (and, by extension, BSI) based on the full blood count (FBC) and C-reactive protein (CRP) values. Several models (i.e. multivariable logistic regression, linear discriminant analysis, K nearest neighbors, support vector machine, random forest model and decision tree) were trained using FBC and CRP values of 2693 infants aged 7 to 60 days with suspected BSI between 2005 and 2022 in a tertiary paediatric hospital in Dublin, Ireland. All models tested showed similar sensitivities (range 47% - 62%) and specificities (range 85%-95%). A trained decision tree and random forest model were applied to the full dataset and to a dataset containing infants with suspected BSI in 2023 and showed good segregation of a low-risk and high-risk group. Negative predictive values for these two models were high for the full dataset (> 99%) and for the 2023 dataset (> 97%), while positive predictive values were low in both dataset (4%-20%). Conclusion: We identified several models that can predict positive blood cultures in infants with suspected BSI aged 7 to 60 days. Application of these models could prevent administration of antimicrobial treatment and burdensome diagnostics in infants who do not need them. What is Known: • Bloodstream infection (BSI) in infants cause non-specific symptoms and may be difficult to diagnose. • Results of blood cultures can take up to 48 hours. What is New: • Machine learning models can contribute to clinical decision making on BSI in infants while blood culture results are not yet known.