Personalized Antibiogram: A Novel Multitask Machine Learning Framework for Simultaneous Prediction of Antimicrobial...

Conventional hospital antibiograms summarize aggregated resistance rates, limiting their utility for individualized antimicrobial selection. Existing statistical and machine learning models predict each phenotype separately, ignoring correlations among resistance profiles. We developed novel multitask extreme gradient boosting (XGBoost) models utilizing structured data in electronic health records (EHRs) to predict resistance to 8 antimicrobial classes simultaneously and evaluated their performance within the Veterans Health Administration (VHA).

We conducted a retrospective multicenter study of Escherichia coli and Klebsiella spp. isolates collected at 127 hospitals and >1400 clinics from January 2017 to September 2024. Data from January 2017 to September 2023 were used for model development, while data from October 2023 to September 2024 were used for simulated prospective testing. Model performances were compared to hospital antibiograms and single-target XGBoost models.

The training cohort included 536 252 E. coli and 246 898 Klebsiella spp. isolates; the test cohort included 75 138 and 38 015 isolates, respectively. On the test data, the multitask model achieved overall areas under the receiver operating characteristic curve (AUROCs) of 0.779 (E. coli) and 0.810 (Klebsiella spp.), with good to excellent per-class performance (AUROC range, 0.743-0.847). A multitask approach improved calibration and decreased false-negative rates for carbapenem resistance while predicting individualized resistance probabilities for all target antimicrobials simultaneously ("personalized antibiograms").

A multitask XGBoost framework can accurately predict individualized resistance profiles for common Gram-negative pathogens, outperforming conventional antibiograms and single-target models. Personalized antibiograms may enhance the selection of empiric therapy, including the detection of carbapenem resistance in low-endemicity settings.