A machine learning-driven multitask framework for antibiotic resistance prediction in Escherichia coli, Klebsiella pneumoniae, and Enterobacter cloacae bloodstream infections among hematological malignancy patients.
Inappropriate empirical antimicrobial therapy (IEAT) significantly increases mortality in patients with resistant Gram-negative bacteremia. We developed a multitask machine learning framework to predict carbapenem resistance (CR), β-lactam/β-lactamase inhibitor combinations resistance (BL/BLI-R), and third-/fourth-generation cephalosporins resistance (3GC/4GC-R) in HM patients with monomicrobial BSIs caused specifically by Escherichia coli, Klebsiella pneumoniae, or Enterobacter cloacae.
Using retrospective data from 1,353 HM patients with three specific Enterobacterales BSIs (2017-2023), we trained support vector machines, eXtreme Gradient Boosting, and logistic regression models through 5-fold cross-validation with hyperparameter tuning and conducting internal validation via bootstrap. Model thresholds were optimized via Pareto front analysis to minimized IEAT and carbapenem use while maximizing sensitivity.
The models achieved AUCs of 0.81 (95% CI, 0.78-0.85) for CR model, 0.81 (95% CI, 0.77-0.84) for BL/BLI-R model, and 0.72 (95% CI, 0.68-0.76) for 3GC/4GC-R model before optimization. Threshold optimization improved sensitivity for CR prediction increased from 0.10 to 0.77, for BL/BLI-R prediction from 0.19 to 0.94, and for 3GC/4GC-R prediction from 0.54 to 0.98. SHAP analysis identified key predictors including prolonged neutropenia, prior carbapenem exposure, and tumor consolidation stage. Integrating pathogen species enhanced 3GC/4GC-R prediction (AUC 0.72 vs. 0.76). Clinically, by employing the threshold optimization strategy, the model successfully reduced empirical carbapenem use from 71.7% to 32.1% without increasing the rate of IEAT compared to clinical practice (4.0% vs. 6.11%).
This multitask framework supports resistance prediction in HM patients with monomicrobial BSIs due to three specific Enterobacterales species, mainly used to assist clinical decision-making after pathogen identification and strongly suspected infection with one of the three pathogens. Prospective validation is required before clinical application.
Using retrospective data from 1,353 HM patients with three specific Enterobacterales BSIs (2017-2023), we trained support vector machines, eXtreme Gradient Boosting, and logistic regression models through 5-fold cross-validation with hyperparameter tuning and conducting internal validation via bootstrap. Model thresholds were optimized via Pareto front analysis to minimized IEAT and carbapenem use while maximizing sensitivity.
The models achieved AUCs of 0.81 (95% CI, 0.78-0.85) for CR model, 0.81 (95% CI, 0.77-0.84) for BL/BLI-R model, and 0.72 (95% CI, 0.68-0.76) for 3GC/4GC-R model before optimization. Threshold optimization improved sensitivity for CR prediction increased from 0.10 to 0.77, for BL/BLI-R prediction from 0.19 to 0.94, and for 3GC/4GC-R prediction from 0.54 to 0.98. SHAP analysis identified key predictors including prolonged neutropenia, prior carbapenem exposure, and tumor consolidation stage. Integrating pathogen species enhanced 3GC/4GC-R prediction (AUC 0.72 vs. 0.76). Clinically, by employing the threshold optimization strategy, the model successfully reduced empirical carbapenem use from 71.7% to 32.1% without increasing the rate of IEAT compared to clinical practice (4.0% vs. 6.11%).
This multitask framework supports resistance prediction in HM patients with monomicrobial BSIs due to three specific Enterobacterales species, mainly used to assist clinical decision-making after pathogen identification and strongly suspected infection with one of the three pathogens. Prospective validation is required before clinical application.
Authors
Cui Cui, Feng Feng, Shao Shao, Zhang Zhang, Li Li, Lin Lin, Wang Wang, Liu Liu, Pan Pan, Zhen Zhen, Fan Fan, Zhang Zhang, Mi Mi, Xiao Xiao, Jiang Jiang, Han Han, Wang Wang, Chen Chen, Feng Feng
View on Pubmed