Rapid antibiotic sensitivity prediction in Pseudomonas aeruginosa using UV–vis-NIR spectroscopy and gray-box one-vs-all models

Abstract

Pseudomonas aeruginosa is a widespread pathogen known to cause infections in various hosts, particularly threatening immunocompromised patients. Although determining antibiotic sensitivity is crucial for appropriate patient care, existing diagnostic methods remain time-consuming, which can delay targeted therapy. In this study, we propose a novel, interpretable, and cost-effective framework that combines ultraviolet-visible-near-infrared (UV–Vis-NIR) spectroscopy with subgroup discovery and a one-vs-all multilayer perceptron (MLP) model to predict antibiotic sensitivity without the need for traditional culture methods. Unlike prior approaches that depend on expensive instruments or black-box algorithms, our method leverages spectral pattern interpretability to identify key wavelength features associated with distinct resistance categories. Testing on clinical isolates of P. aeruginosa, the model achieved optimal prediction accuracy within 10 min of culture time, significantly reducing the typical 48–72 h turnaround time of conventional culture-based susceptibility testing. This work demonstrates a promising direction for rapid, low-cost, and clinically actionable antimicrobial susceptibility testing that balances performance with explainability.