Hospitalized cancer patients who develop fever present a clinical dilemma: aggressive antibiotic treatment is often necessary to prevent life-threatening infections in immunocompromised individuals, yet overtreatment carries its own risks. Now, researchers have developed machine learning models that predict which patients will remain febrile 48-72 hours after starting antibiotics, potentially guiding more personalized treatment decisions while maintaining safety.

The models, developed at University Medicine Essen in Germany and validated at Brigham and Women's Hospital in the United States, achieved an area under the receiver operating characteristic curve (AUROC) of 0.82 in internal validation and 0.78 in external validation—substantially outperforming a baseline approach that scored just 0.66.

Identifying Candidates for De-escalation

The study analyzed 987 hospital stays at the German site and 7,409 at the U.S. hospital, spanning January 2020 through March 2025 and June 2015 through September 2024, respectively. Baseline persistent fever rates were 54.0% at University Medicine Essen and 46.5% at Brigham and Women's Hospital.

The researchers found that approximately one quarter of antibiotic escalations and chest CT scans occurred in patients the model correctly predicted would not remain febrile. This suggests substantial opportunity for reducing unnecessary interventions in a population where conservative management has traditionally dominated clinical practice.

Temperature-related variables from electronic health records emerged as the dominant predictors, alongside demographics, comorbidities, laboratory results, and vital signs. The best-performing model used TabPFN, though researchers also tested regularized logistic regression, Random Forest, XGBoost, and CatBoost approaches.

Cross-Continental Validation Demonstrates Robustness

The external validation proved particularly noteworthy given significant differences between the two patient populations. The German cohort was dominated by hematologic malignancies (82.4%), while the U.S. cohort included 56.8% solid tumors. Temperature monitoring frequency also varied substantially, with Brigham and Women's Hospital averaging 12.2 measurements in the first 24 hours post-antibiotics compared to 6.1 at University Medicine Essen.

Despite these differences, the model maintained clinically meaningful accuracy across both settings, suggesting potential for broader implementation in diverse oncology environments.

Adjunctive Tool, Not Standalone Decision-Maker

The researchers emphasized critical limitations: these predictions should serve only as adjunctive decision support rather than standalone diagnostic tools. Given the potentially life-threatening nature of infections in immunocompromised cancer patients, conservative management remains essential, and clinical judgment must remain paramount.

Prospective studies will be needed to evaluate real-world safety and clinical impact before any deployment in routine practice. The reliance on routine clinical data from electronic health records—rather than specialized inputs—makes this approach technically feasible for integration into existing hospital workflows, though regulatory approval would require demonstration of safety in prospective trials.

For healthcare providers, the tool could support more individualized antibiotic stewardship during the standard 48-72 hour reassessment window, providing an additional data point to inform decisions about treatment escalation, de-escalation, or continuation. However, implementation must carefully balance the potential benefits of reduced antibiotic overuse against the imperative to prevent missed infections in this vulnerable population.

Sources

  1. Predicting persistent fever in patients with cancer receiving antibiotics: a machine learning approach — Nature