Personalized antibiograms focus verification on high-risk empiric orders

Predictive Model Development

Authors:Michihiko Goto; Anindita Bandyopadhyay; Qianyi Shi, et al.

Journal:Clinical Infectious Diseases

Personalized antibiograms from a single multi-task XGBoost model forecast per-patient resistance across eight drug classes, cutting missed carbapenem resistance in E. coli from 39.9% to 21.7% on prospective VA data. Stewardship and verification pharmacists can use probability thresholds to triage empiric orders, prioritize high-risk cases, and govern alert volume—augmented, not automated care.

January 17, 2026

Quick Take

A single multi-task XGBoost model produced per-patient "personalized antibiograms" predicting resistance probabilities across eight antimicrobial classes; on the held-out prospective test set, per-class AUROCs ranged 0.743–0.847 and overall AUROCs were 0.779 (E. coli) and 0.810 (Klebsiella spp.).

Compared with conventional hospital antibiograms in the VHA test set, the multi-task model reduced missed carbapenem resistance (false-negative rate for E. coli fell from 39.9% to 21.7% at Youden-index thresholds set on the training data) and can be used to triage empiric choices and focus pharmacist review on higher-risk orders; gains versus single-task XGBoost were generally modest and task-dependent.

Why it Matters

Conventional hospital antibiograms report population-level averages and do not account for patient-specific factors (prior antibiotics, comorbidities, prior cultures), limiting individualized empiric therapy selection.

Carbapenem resistance is uncommon but high-consequence; in low-endemicity settings, aggregate antibiograms and single-phenotype models can have high false-negative rates and poor calibration, impeding early detection during order verification.

A per-patient panel of resistance probabilities across multiple drug classes can be embedded in clinical decision support to prioritize stewardship attention and establish risk-based operating thresholds.

Using patient-level probabilities enables threshold-driven workflows (e.g., high-sensitivity screening to avoid missed carbapenem resistance) and concentrates limited pharmacist verification resources on the most at-risk empiric orders.

What They Did

Retrospective multicenter Veterans Health Administration study of Escherichia coli and Klebsiella spp. isolates: development on January 1, 2017–September 30, 2023 and simulated prospective testing on October 1, 2023–September 30, 2024 (development/test isolate counts: 536,252/75,138 E. coli and 246,898/38,015 Klebsiella).

Assembled structured electronic health record inputs from the VA Corporate Data Warehouse: local susceptibility results for eight antimicrobial classes, specimen type, demographics, 86 Hierarchical Condition Categories (HCCs), 224 procedure groups, prior microbiology, time-windowed pharmacy dispense history, and facility antibiograms.

Implemented a unified multi-task XGBoost by reshaping the dataset into culture–antimicrobial rows with an antimicrobial indicator so one model predicts resistance probabilities for all eight classes simultaneously; used patient-level 5-fold cross-validation and excluded calendar year to limit leakage.

Compared the multi-task model to single-target XGBoost and antibiogram-based logistic regression and evaluated performance on the held-out prospective year using AUROC, AUPRC, Brier scores, and pre-specified operating cutoffs (balanced, sensitivity ≥95%, specificity ≥95%) — cutoffs were determined on the training data and applied unchanged to the test set — producing per-isolate probability panels (personalized antibiograms).

What They Found

Multi-task XGBoost yielded robust probabilistic performance on the held-out prospective data: overall AUROCs were 0.779 for E. coli and 0.810 for Klebsiella spp.; per-class AUROCs ranged 0.743–0.847 (interpreted as good-to-excellent).

Relative to conventional hospital antibiograms, the model reduced missed carbapenem resistance in the VHA test set: false-negative rates declined from 39.9% to 21.7% for E. coli and from 48.2% to 29.1% for Klebsiella when using Youden-index thresholds set on the training data.

Pre-specified operating points illustrate trade-offs clinicians can choose: balanced thresholds typically gave sensitivities ~55–70%, while sensitivity-prioritized cutoffs (set to ≥95% on training data) achieved ≥95% sensitivity at the cost of much lower specificities (~10–30%), enabling high-sensitivity screening but increasing false positives.

For pharmacy practice, the personalized antibiogram supplies per-isolate resistance probabilities across eight drug classes that can populate stewardship workqueues and support threshold-driven triage so pharmacists concentrate verification on high-risk empiric orders.

Takeaways

Use the personalized antibiogram as a prioritization signal, not a replacement for clinical judgment — best used to triage empiric choices and surface higher-risk orders for pharmacist or clinician review.

Select operating thresholds deliberately: increasing sensitivity catches more carbapenem resistance but raises false positives and alert volume; set per-drug cutoffs aligned to local tolerance for false alarms and available review capacity.

Validate local data fit before deployment: the model depends heavily on prior susceptibilities and recent antibiotic exposures — confirm those inputs are available, assess calibration on local isolates, and retrain locally as needed.

Operational planning must include input-reliability monitoring, explainability (e.g., SHAP or similar) to support clinician review and regulatory positioning, and governance for threshold management; treat outputs like a calibrated 'weather forecast by drug class' that guides attention rather than automatically dictating therapy.

Strengths and Limitations

Strengths:

Large, multicenter VHA cohort with extensive patient-level features, simulated prospective held-out-year testing, and patient-level 5-fold cross-validation; reporting followed TRIPOD+AI guidance supporting reproducibility.; The multi-task XGBoost formulation enabled joint learning across correlated resistance phenotypes, which improved calibration for rare outcomes in this dataset and supported interpretable feature-importance analyses useful for deployment planning.

Limitations:

Retrospective, VHA-only sample (older, male-skewed population, integrated longitudinal EHR) limits external generalizability; local retraining and fairness audits are required before use in other populations or care settings.; Real-world risks include missing or delayed real-time inputs (data latency), inconsistent availability of genotypic data, excluded species/classes, and operating cutoffs tuned on training data that may not match local clinical priorities — these can degrade live performance and necessitate workflow/data engineering and prospective clinical validation.

Bottom Line

This study demonstrates feasibility for personalized resistance forecasting within a large integrated system; the approach supports calibration-checked pilots with threshold governance, input-reliability monitoring, explainability, and pharmacist-in-the-loop workflows, but requires local validation before clinical deployment.