Machine learning–based decision support for ICU admission in febrile oncology patients: Optimizing care delivery and resource allocation.

1598 Background: In febrile oncology patients, distinguishing self-limiting illness from life-threatening sepsis is a critical challenge for care delivery, particularly in resource-constrained settings. Over-triage strains limited critical care capacity, while under-triage compromises patient safety. We aimed to develop and validate a machine learning–based decision support tool to assist clinical judgment in identifying patients requiring intensive care. Methods: In a retrospective cohort of 149 febrile oncology patients, we extracted routinely available clinical features (including MASCC score, qSOFA, hypotension status, tumor characteristics, and comorbidities). We developed an XGBoost classifier to predict ICU admission and compared its performance against logistic regression and standard clinical risk scores (MASCC, qSOFA). Model performance was rigorously evaluated using 10×5-fold cross-validation with bootstrap confidence intervals, with specific attention to its utility as a supportive screening tool. Results: Of 149 patients, 81 (54.4%) required ICU admission. The XGBoost model demonstrated superior discrimination with an AUROC of 0.934 (95% CI: 0.863-1.000), significantly outperforming logistic regression (0.917), MASCC (0.656), and qSOFA (0.838). The model showed excellent calibration (Brier score: 0.092). Crucially, in a sensitivity analysis excluding hypotension—a clinically obvious trigger for ICU transfer—the model retained high discriminative ability (AUROC 0.887), indicating its value in identifying high-risk patients before hemodynamic collapse. Conclusions: This machine learning model effectively identifies febrile oncology patients requiring ICU admission, significantly outperforming traditional scores. By functioning as a high-fidelity decision support tool, it holds the potential to improve care delivery outcomes: ensuring timely escalation for high-risk patients while preventing unnecessary resource utilization for stable patients. Ultimately, this tool is designed to augment, rather than supersede, expert clinical judgment in complex acute care settings.