From Klebsiella and Candida to artificial intelligence: a perspective from infectious diseases doctors and researchers

Daniele Roberto Giacobbe^1*Cristina Marelli²Marco Muccio³Sabrina Guastavino¹Ylenia Murgia¹Sara Mora³Alessio Signori¹Nicola Rosso³Antonio Vena¹Mauro Giacomini¹Cristina Campi¹Michele Piana¹Matteo Bassetti³

• ¹University of Genoa, Genoa, Italy
• ²Gustave Roussy, Villejuif, France
• ³IRCCS Ospedale Policlinico San Martino, Genoa, Italy

The advent of artificial intelligence (AI) and machine learning (ML) is progressively influencing clinical reasoning in infectious diseases, particularly in the management of septic shock where timely empirical antimicrobial therapy is crucial. In this perspective, we discuss how AI and ML approaches intersect with established clinical decision-making processes through two examples from our research and practice: prediction of bloodstream infection by carbapenem-resistant Klebsiella pneumoniae and prediction of candidemia. Traditionally, risk estimation has relied on interpretable models such as logistic regression, offering clinicians transparent insights into the contribution of specific risk factors. In contrast, some ML models leverage complex relationships within large datasets. Despite expectations, in several cases these complex models have not consistently outperformed classical approaches yet, a phenomenon we refer to as the "accuracy paradox," possibly stemming from limitations in data specificity and granularity. Furthermore, the opacity of many ML models still challenges their integration into clinical practice, raising ethical and practical concerns around explainability and trust. While explainable AI offers partial solutions, ML may also capture hidden patterns undetectable through classical reasoning that could be unexplainable to clinicians per definition. Achieving a reasonable and shared balance will require continued collaboration between clinicians, data scientists, and ethicists. As the field evolves, future research should prioritize the development of models that not only perform well but can also integrate meaningfully into the complex cognitive processes underpinning bedside clinical reasoning.