Open access publication: nature.com/articles/s41...
Web application: uol.de/carba-detector
Thanks to all co-authors.

#AntimicrobialResistance #MachineLearning #ClinicalMicrobiology #InfectionControl

The model integrates into existing laboratory workflows without modification of standard procedures. Predictions are instantaneous, theoretically enabling rapid infection control decisions and more efficient use of confirmatory testing resources.

Performance: 96% sensitivity, 87% specificity for carbapenemase detection. This significantly outperforms existing screening algorithms while using only routine laboratory data.

CarbaDetector uses random forest classification on up to 8 standard antibiotic zone diameters from disk diffusion testing. Training dataset: 385 isolates with whole genome sequencing validation. External validation: 800+ isolates from multiple sites.

Machine learning excels at identifying complex, multidimensional patterns in large datasets. Carbapenemase production creates subtle resistance signatures across multiple antibiotics that rule-based algorithms cannot capture effectively.

Carbapenemase-producing Enterobacterales are WHO priority pathogens. Rapid detection is critical for infection control, but current screening algorithms have poor specificity (EUCAST: 8.2%), leading to unnecessary confirmatory testing.