AI Improves Clinical Microbiology Generalization Across Centers

This research introduces DALMA, a novel probabilistic representation learning framework designed to overcome a major hurdle in deploying machine learning models for clinical microbiology: the variability between different institutions' data acquisition systems. Existing methods often struggle with "domain shift," where models learn technical artifacts rather than true biological information, limiting their transferability. DALMA addresses this by simultaneously modeling both the acquisition-specific variations and the inherent biological supervision available in microbiology datasets. By combining domain-specific reconstruction with biologically guided representation learning, DALMA generates latent representations that are highly transferable. This allows for "zero-shot" deployment, meaning the model can be used effectively on data from previously unseen clinical centers without requiring any site-specific adjustments during inference. Evaluated on a multi-center benchmark across seven datasets from three countries, DALMA consistently achieved state-of-the-art microbial identification and also effectively transferred to antimicrobial resistance prediction. Furthermore, its ability to estimate novelty in the latent space enables reliable selective prediction under new domain shifts.