Machine Learning Model to Guide Empirical Antimicrobial Therapy in Febrile Neutropenic Patients With Hematologic Malignancies

Optimal antimicrobial selection for patients with febrile neutropenia (FN) may differ depending on the underlying mechanisms. We aimed to develop a model for predicting the severity of bacteremia in patients with FN and hematologic malignancies (HMs) to help clinicians select appropriate antimicrobials using a machine-learning approach. In this single-center retrospective study, we analyzed the characteristics and microbial epidemiology of patients with FN and HMs who developed bacteremia. We applied a machine learning approach (least absolute shrinkage selection operator) to select the variables and then created a risk score. Using the risk score, a model was constructed that enabled us to estimate the probability of developing severe complications when a narrow- [cefepime (CEM)] or broad-spectrum [either piperacillin-tazobactam or meropenem (PT+MEM)] antimicrobial agent was administered. In total, 228 patients were enrolled. Of these, a microbiological cohort (n=126) and an analysis cohort (n=88) were established. In the microbiological cohort, coagulase-negative staphylococci (20.6%) were the most common pathogens, and antimicrobial resistance mechanisms were identified in 53 isolates (42.1%). In the analysis cohort, CEM and PT+MEM were administered to 53 (60.2%) and 35 (39.8%) patients, respectively. The overall incidence of severe complications was 26.1%. The performance of the machine learning model was measured by the area under the receiver operating characteristic curve (AUC) (AUC=0.813; 95% confidence interval=0.691-0.894), which showed good discrimination. This pilot study introduces a novel method for constructing predictive models tailored to specific patient groups, potentially supporting antimicrobial stewardship.