Abstract
Carbapenemase-producing Enterobacterales (CPE) are considered among the highest threats to global health by WHO. Their detection is difficult and time-consuming. We developed a random-forest machine learning (ML) model, CarbaDetector, to predict carbapenemase production from inhibition zone diameters of eight antibiotics, using 385 isolates for training with whole genome sequencing as reference. Validation on two external datasets (A = 282, B = 518 isolates) shows high performance: sensitivity/specificity are 96.6%/84.4% (training), 96.3%/86.1% (A), and 91.2%/87.0% (B, five antibiotics). In contrast, the algorithms of EUCAST and the Antibiogram Committee of the French Society of Microbiology (CA-SFM) exhibit lower specificity (8.2% and 40.1%, respectively on the training dataset). In this work, we show that CarbaDetector, available as a web-app, reduces unnecessary confirmatory testing and accelerates the time to result. This approach offers high sensitivity and improved specificity compared to standard algorithms and has the potential to improve CPE detection, especially in resource-limited settings.