Answers are grounded in this paper and cite their sources.
OBJECTIVE: To develop and validate a machine learning (ML) framework for the rapid prediction of non-susceptibility to eight antipseudomonal antibiotics directly from routine MALDI-TOF mass spectrometry (MS) spectral data, addressing the need for accelerated antimicrobial susceptibility testing for Pseudomonas aeruginosa. METHODS: A total of 262 non-duplicate P. aeruginosa clinical isolates were collected from two geographically distinct Chinese hospitals, yielding 2792 spectra encoding 96 m/z features (2042-18 076 Da). Eight ML algorithms were systematically evaluated and ranked via DeLong tests. SHapley Additive exPlanations (SHAP) was employed for guided feature ablation to reduce models to 15-29 spectral features. External validation was performed using an independent dataset, and MIC correlation analysis was conducted to associate spectral peaks with quantitative resistance levels. An interactive web application was developed for deployment. RESULTS: Final models achieved AUCs of 0.928-0.974 on internal validation and 0.841-0.936 on independent external validation. LightGBM was selected for five antibiotics, XGBoost for two, and Random Forest for one. SHAP analysis identified conserved markers, notably m/z 6502 (selected across all eight models), and drug-specific signatures. Feature ablation was successful without significant loss in discriminative power. MIC correlation analysis confirmed significant associations between selected peaks and resistance levels. CONCLUSION: This work demonstrates that routine MALDI-TOF MS data can be successfully repurposed for interpretable, rapid antibiotic resistance prediction in P. aeruginosa without additional instrumentation. The framework provides a scalable solution for real-time susceptibility prediction in clinical settings.
No takes yet. Share an insight, caveat, or question.
Ren et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: