Background: Empirical antibiotic therapy (EAT) in febrile neutropenia (FN) remains challenging due to multidrug-resistant (MDR) Gram-negative bacteria, often leading to inappropriate empirical antibiotic therapy (IEAT). Objective: To demonstrate that risk stratification based on machine learning (ML) and prior colonisation with MDR bacteria may support the tailoring of EAT in patients with haematological malignancies. Design: Retrospective proof-of-concept cohort study. Methods: All consecutive FN episodes in patients with haematological malignancies were retrospectively included from January 2020 to March 2023 at a tertiary-level university hospital. We compared real-world, clinician-driven empirical antibiotic use with a simulated approach guided by an ML-based risk stratification model combined with prior colonisation data. The main outcomes were antibiotic selection and rates of IEAT. Results: A total of 553 FN episodes in 398 haematological patients were analysed. Bloodstream infection (BSI) occurred in 141/553 episodes (25.5%). Anti-pseudomonal (PsA) beta-lactams were prescribed in 515/553 episodes (93.1%), with carbapenems in 406/553 (73.4%). The clinician-driven approach resulted in 16/70 (22.9%) GNB-BSI episodes receiving IEAT. The ML plus colonisation-guided approach would have reduced the use of meropenem by 29.7% (−2.08 days; 95% CI, −2.42 to −1.73; p < 0.001) and anti-PsA beta-lactams by 6.7% (−0.47 days; 95% CI, −0.76 to −0.19; p = 0.001), and would also have led to a reduction in the rate of IEAT from 16/70 (22.9%) to 6/70 (8.6%) ( p = 0.035). Conclusion: ML-based risk stratification combined with colonisation status would allow for personalised antibiotic therapy in FN, potentially reducing IEAT and improving antimicrobial use. These results support integrating these tools into clinical practice.
Gallardo-Pizarro et al. (Thu,) studied this question.