Bacteremia in Patients With Fever and Acute Lower Extremity Pain in a Non-Lyme Endemic Region

Recent evidence reports that a high proportion of pediatric patients presenting to the emergency department (ED) for fever and acute lower extremity pain have positive blood cultures (BCs).¹ However, these results have not yet been validated at a different, non-Lyme endemic medical center. METHODS We conducted a cross-sectional study from July 1, 2018, to June 30, 2022, at a large Midwest US pediatric ED with ~50 000 patient encounters per year. Lyme disease is rarely seen at this center.² We included patients aged 1 to 18 years presenting with fever within 24 hours of presentation and acute, unilateral lower extremity pain. We excluded patients with acute traumatic injury (presenting within 24 hours of the injury), multifocal pain, immunocompromised status, indwelling central line, preexisting orthopedic comorbidity, antibiotic use within the previous 24 hours, preexisting poor bone density or joint health, and those with an identified alternative reason for limping (eg, neurologic deficits and abdominal pain) as determined by the ED physician’s initial examination. The Washington University institutional review board approved this study. We used a natural language processing-assisted manual chart review that has been previously described^1,3,4 to identify our patient cohort. Regular expressions identifying fever and lower-extremity pain (eg, “limp,” “refused to bear weight,” “nonweightbearing”) assisted with narrowing manual review of the large cohort of patients. We trained a support-vector machine model on the manual review of charts. We reviewed additional charts identified by the support-vector machine model to achieve 95% sensitivity to identify the patients meeting inclusion criteria. We manually abstracted laboratory results, physical exam findings, and consultation notes to determine whether BC results prompted a change in patient management. We used Document Review Tools as a graphical user interface for natural language processing review and machine learning. We used SPSS (IBM, Armonk, New York) for statistical analysis. Descriptive statistics included proportions with confidence intervals (CIs) and medians with interquartile ranges.