Predicting Persistent Fever in Patients with Cancer Receiving Antibiotics: A Machine Learning Approach

Persistent fever after starting intravenous antibiotics is common in hospitalized patients with hematologic and solid cancers. This often triggers additional imaging and broader antimicrobial treatment. Earlier identification of patients unlikely to remain febrile could help to prevent unnecessary antimicrobial escalation and diagnostic procedures. We developed and externally validated machine learning models to predict persistent fever 48–72 hours after initiation of antibiotic therapy using electronic health records from two large academic hospitals. Adult patients with cancer who had fever within 24–48 hours of treatment and at least one temperature measurement between 48–72 hours were included. Models incorporated demographics, comorbidities, laboratory results, vital signs, and temperature patterns. Across internal cross-validation and external testing, the models showed consistent discrimination for predicting persistent fever 48–72 hours after antibiotic initiation; the best-performing model achieved an area under the receiver operating characteristic curve of 0.82 (95% CI, 0.79–0.84) in internal validation and 0.78 (95% CI, 0.77–0.79) in external validation. Temperature-related variables were the most impactful predictors. Decision-curve analyses demonstrated positive net benefit across clinically relevant thresholds. In retrospective assessment, approximately one quarter of antibiotic escalations and chest CT scans were performed in patients who were correctly predicted not to remain febrile, suggesting opportunities to reduce unnecessary interventions. These findings show that short-term fever persistence in patients with cancer can be predicted with clinically meaningful accuracy using routine data. Prospective studies are needed to assess safety, clinician uptake, and real-world impact of integrating such tools into decision-making pathways.