TOPLINE:

A novel multi-task Extreme Gradient Boosting (XGBoost) model using electronic health record (EHR) data showed good-to-excellent accuracy in predicting antimicrobial resistance profiles of Escherichia coli and Klebsiella species and appeared to outperform conventional hospital antibiograms.

METHODOLOGY:

• Researchers conducted a retrospective study in the US to develop and validate a novel XGBoost model using EHR data to predict resistance to eight antimicrobial groups — including extended-spectrum cephalosporins, carbapenems, and fluoroquinolones — in E coli and Klebsiella species.
• They used data from two patient cohorts:
  • Training cohort — Data from 287,045 patients (E coli) and 139,377 patients (Klebsiella), collected between January 2017 and September 2023, were used to develop the model.
  • Test cohort — Data from 53,913 patients (E coli) and 26,907 patients (Klebsiella), collected between October 2023 and September 2024, were used to evaluate predictive performance.
• The performance of the multitask XGBoost model in predicting personalized antimicrobial resistance profiles was compared with the performance of conventional hospital antibiograms and single-task models.
• Model probabilities were converted into “yes/no” predictions with three training set cutoffs per antimicrobial: a balanced cutoff that optimally balanced sensitivity and specificity, a high‑sensitivity cutoff (sensitivity ≥ 95%), and a high‑specificity cutoff (specificity ≥ 95%). These cutoffs were then applied unchanged to the test cohort.
• They also assessed the predictive performance of the models for each antimicrobial and overall using the area under the receiver operating characteristic curve (AUROC) (0.5-0.7 indicated poor, 0.7-0.8 good, 0.8-0.9 excellent, and values > 0.9 indicated outstanding performance).

TAKEAWAY:

• The XGBoost multi-task model demonstrated excellent overall performance for Klebsiella species (AUROC, 0.810) and good overall performance for E coli (AUROC, 0.779) in the test cohort.
• In the training cohort, the novel model achieved good-to-excellent performance for all antimicrobials, with the highest prediction accuracy for fluoroquinolones in E coli (AUROC, 0.832) and carbapenems in Klebsiella species (AUROC, 0.869).
• In the test cohort, the novel model achieved good-to-excellent performance for all antimicrobials, with the highest prediction accuracy for extended-spectrum cephalosporins in E coli (AUROC, 0.825) and for carbapenems in Klebsiella species (AUROC, 0.847).
• Both multitask and single-task models showed superior prediction performance compared with conventional hospital antibiograms. Overall improvements in prediction accuracy were comparable between single-task and multitask models, but the multitask model had lower false-negative rates for carbapenem resistance in E coli, indicating better detection of carbapenem‑resistant isolates.

IN PRACTICE:

“A multitask machine learning model, such as ours, can potentially serve as an identification tool by setting thresholds to prioritize specificities and increase the efficiency of including patients who likely benefit,” the authors of the study wrote.

SOURCE:

This study was led by Michihiko Goto, MD, MS, University of Iowa Carver College of Medicine, Iowa City, Iowa. It was published online on January 17, 2026, in Clinical Infectious Diseases.

LIMITATIONS:

The model depended on historical Veterans Health Administration (VHA) data; therefore, its accuracy hinged on data quality, and the model will require ongoing updates and external validation to maintain performance. The VHA cohort is generally older and predominantly includes men (about 20% of isolates in this study were from women), limiting generalizability. Resistance gene markers were inconsistently available — especially at smaller or rural sites — and were excluded, potentially introducing selection bias.

DISCLOSURES:

This study was supported by the Agency for Healthcare Research and Quality. One author reported receiving a grant from the same agency.

This article was created using several editorial tools, including AI, as part of the process. Human editors reviewed this content before publication.