Machine learning-assisted risk assessment for fluoroquinolone treatment in Chryseobacterium indologenes bacteremia: A comparative study of model performance and clinical calibration.

INTRODUCTION

Chryseobacterium indologenes bacteremia poses significant therapeutic challenges due to intrinsic multidrug resistance and the absence of established Clinical and Laboratory Standards Institute breakpoints for fluoroquinolones. We aimed to develop machine learning models to predict mortality and guide fluoroquinolone treatment decisions.

METHODS

In this retrospective study of 61 patients, we compared Logistic Regression, Conservative Random Forest, and two Calibrated Random Forest variants (Sigmoid/Isotonic) using Borderline Synthetic Minority Over-sampling Technique and conservative parameter settings to address small-sample limitations.

RESULTS

Mortality was 21.3%, with shock identified as the strongest predictor (p < 0.001). Conservative Random Forest achieved the highest area under the curve (0.908) and good calibration stability. Notably, complex calibration methods yielded unstable probability curves, revealing a "calibration paradox" in small-sample settings. A risk stratification system based on optimal models successfully identified low-risk patients suitable for fluoroquinolone therapy.

CONCLUSIONS

For rare infections with limited data, robust, conservative RF demonstrated a trend of stable calibration characteristics. This study provides an objective clinical decision support tool for Chryseobacterium indologenes treatment when standard susceptibility interpretation is unavailable.