Early Antimicrobial Resistance Prediction Using Urinary Proteomics and Machine Learning

Presenter: Emelie Lemann

This study evaluated the use of urinary proteomics combined with machine learning (ML) to predict antimicrobial resistance (AMR) in E. coli urinary tract infections using clinical samples. Proteomic data from urine samples were analysed alongside resistance phenotypes, and random forest ML models were trained to predict resistance to 10 antibiotics. Single-label models (one antibiotic at a time) and an exploratory multi-label model (multiple antibiotics simultaneously) were developed. Single-label models showed high predictive accuracy, with the best performance for cefadroxil (99.5%) and cefotaxime (99.3%), as well as meropenem (98.9%) and mecillinam (98.4%). Other antibiotics were predicted with accuracies ranging from 77.2% to 98.2%.

The multi-label model achieved 59.3% average accuracy, with a macro-F1 score of 68.6% and a weighted F1 score of 91.4%, indicating strong performance for common resistance patterns.

Overall, urinary proteomic signatures enabled accurate prediction of AMR, supporting their potential use in early, sample-specific resistance assessment.

A Next-Generation Platform for Culture-Free AMR Gene Identification in Human Infections

Presenter: Tanshi Mehrotra

This multicentric genomics study characterized the antimicrobial resistance (AMR) landscape in India and explored its relevance for developing rapid diagnostic tools for sepsis, urinary tract infections (UTIs), and respiratory infections. A total of 2,720 genomes were analysed, including 203 clinical isolates from 13 Gram-negative bacterial species collected between 2016 and 2022. Phenotypic susceptibility data for 17 antibiotics supported genomic identification of AMR alleles. Phylogeographic analysis, multilocus sequence typing (MLST), and pan-genome reconstruction demonstrated substantial genetic diversity, lineage expansion, and evolution of resistant pathogens. Extracellular pathogens showed extensive drug resistance, with over 70% of isolates exhibiting extensively drug-resistant profiles, largely driven by horizontally acquired AMR genes. A wide range of AMR determinants was identified, associated with 38 plasmid types, indicating high transmission potential. Regionally enriched high-risk international clones were also observed.

The study further evaluated key AMR gene sets in Klebsiella pneumoniae, Acinetobacter baumannii, Escherichia coli, and Pseudomonas aeruginosa across antibiotic contexts.

In conclusion, the validated AMR gene panels provide a foundation for developing region-specific rapid diagnostics targeting major resistance determinants.

Evaluating a Machine Learning-Based Decision Support Tool for Early Antibiotic Discontinuation in Lower Respiratory Tract Infections (LRTI): A Randomised Controlled Trial

Presenter: Sarah Tang

This study evaluated the impact of a machine learning-based clinical decision support tool (AI²D) on antibiotic prescribing in adults with suspected lower respiratory tract infections (LRTIs) at Singapore General Hospital (January–September 2025). Among 1,066 eligible patients, AI²D identified bacterial LRTI as unlikely in 457 (42.9%). Patients were managed either with AI²D-supported recommendations (INT, n=246) or standard of care (SOC, n=211). Physicians accepted 64.2% (158/246) of AI²D recommendations.

The INT group had a significantly higher rate of early antibiotic discontinuation (≤3 days) compared to SOC (41.1% vs 16.1%, p<0.001). Antibiotic duration was shorter in the INT group (6 vs 7 days, p<0.001), and 30-day hospital readmissions were lower (15.0% vs 24.2%, p<0.05). There were no statistically significant differences in 30-day mortality or length of hospital stay between groups.

Overall, AI²D use improved antibiotic stewardship by reducing unnecessary antibiotic exposure without compromising patient safety.

Cost-Effectiveness of a Machine Learning (ML) Approach for Reducing Unnecessary Antibiotic Use in Patients with Suspected Lower Respiratory Tract Infections (LRTI): An Economic Analysis of a Randomised Clinical Trial

Presenter: Yiying Cai

This economic analysis evaluated the cost-effectiveness of a machine learning (ML)–guided antimicrobial stewardship approach for suspected lower respiratory tract infections (LRTIs), using data from a randomized trial in Singapore. A decision tree model compared the ML approach—where identified patients undergo antimicrobial stewardship review—versus usual care without review. Outcomes were assessed from the hospital perspective, including costs (2025 SGD), life-years gained, and antibiotic use. In the base-case analysis, the ML approach reduced costs by $477 per patient (95% UI: -518 to -437) and improved outcomes by 0.17 life-years per patient (95% UI: 0.17 to 0.18). Antibiotic use decreased by 1.34 days per patient. At the hospital level, this translated to an estimated reduction of 3,348 antibiotic days annually.

Scenario analyses showed that improved adherence to ML recommendations further enhanced patient outcomes and reduced antibiotic use.

The ML-guided approach was both cost-saving and cost-effective compared to usual care.

ESCMID 2026, 17-21 April, Munich, Germany.