Background: Data-driven approaches to effectively select antibiotics are crucial to improving patient outcomes and reducing antibiotic resistance. This study aimed to determine whether routinely collected clinical and microbiological data can be used to train machine learning (ML) models to predict antibiotic resistance in patients with bacterial infections. Methods: We conducted a retrospective observational study on clinical data from two separate Italian hospitals, analyzing 15,581 bacterial isolates collected from 9,966 patients between 2018 and 2024 at a multi-department hospital located in Rome and secondary clinic located in Capranica, about 70 Km away from Rome. Multiple ML models were trained, using both unbalanced and SMOTE-balanced datasets. Performance was assessed using cross-validation and independent test sets, comprising 2023-2024 isolates. Performance was measured using the Area Under the Receiver Operating Characteristic curve (AUROC), F1 score, accuracy, precision, and recall. Results: XGBoost consistently outperformed other trained models, achieving AUROCs of 0.882 and 0.878 for Gram-positive and Gram-negative datasets, respectively. Species-specific models further improved discrimination, reaching AUROC scores up to 0.946 for P. aeruginosa, 0.941 for K. pneumoniae, 0.938 for E. faecalis, 0.919 for E. coli, 0.894 for P. mirabilis, and 0.891 for S. aureus. Conclusions: These results demonstrate the utility of ML models in accurately predicting antibiotic susceptibility from routine clinical data, thus facilitating rapid initiation of targeted therapy. Such an approach can potentially reduce treatment delays by up to 48 hours compared to traditional diagnostic methods, presenting a useful tool to manage patients in critical conditions and combat antibiotic resistance in clinical practice.
Investigating Infection and Clinical Data to Predict Effective Antibiotic Therapy and Monitor Resistance Trends
Facchiano, Antonio
2026-01-01
Abstract
Background: Data-driven approaches to effectively select antibiotics are crucial to improving patient outcomes and reducing antibiotic resistance. This study aimed to determine whether routinely collected clinical and microbiological data can be used to train machine learning (ML) models to predict antibiotic resistance in patients with bacterial infections. Methods: We conducted a retrospective observational study on clinical data from two separate Italian hospitals, analyzing 15,581 bacterial isolates collected from 9,966 patients between 2018 and 2024 at a multi-department hospital located in Rome and secondary clinic located in Capranica, about 70 Km away from Rome. Multiple ML models were trained, using both unbalanced and SMOTE-balanced datasets. Performance was assessed using cross-validation and independent test sets, comprising 2023-2024 isolates. Performance was measured using the Area Under the Receiver Operating Characteristic curve (AUROC), F1 score, accuracy, precision, and recall. Results: XGBoost consistently outperformed other trained models, achieving AUROCs of 0.882 and 0.878 for Gram-positive and Gram-negative datasets, respectively. Species-specific models further improved discrimination, reaching AUROC scores up to 0.946 for P. aeruginosa, 0.941 for K. pneumoniae, 0.938 for E. faecalis, 0.919 for E. coli, 0.894 for P. mirabilis, and 0.891 for S. aureus. Conclusions: These results demonstrate the utility of ML models in accurately predicting antibiotic susceptibility from routine clinical data, thus facilitating rapid initiation of targeted therapy. Such an approach can potentially reduce treatment delays by up to 48 hours compared to traditional diagnostic methods, presenting a useful tool to manage patients in critical conditions and combat antibiotic resistance in clinical practice.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
