Predicting Gram-negative bloodstream infection in elderly patients after isolation of GNB from non-blood specimens: a machine learning-based tool.

BACKGROUND

This study aimed to develop a machine learning-based model for the early identification of elderly inpatients at high risk of Gram-negative bloodstream infection immediately following the first detection of Gram-negative bacteria in non-blood specimens (e.g., urine, sputum, wound secretions), during the 48-72 h window before blood culture results become available.

METHODS

A retrospective cohort study was conducted across three centers of Sichuan Provincial People's Hospital (Qingyang, Caotang, and Chengdong Campuses). It enrolled 9,646 elderly inpatients with Gram-negative bacteria positivity in any specimen during their hospitalization between January 2017 and December 2022. Predictor variables were screened using LASSO regression and Boruta algorithm, and their clinical significance was validated via the Delphi expert consultation method. Six machine learning models-random forest, XGBoost, logistic regression, artificial neural network, k-nearest neighbors, and decision tree-were developed. The discriminatory performance of the models was evaluated using a comprehensive set of metrics, including the area under the receiver operating characteristic curve along with other key indicators such as accuracy and recall.

RESULTS

The XGBoost model demonstrated the optimal performance among the developed models, achieving a test set AUC of 0.816 (95% CI: 0.766-0.861), accuracy of 0.733, recall of 0.760, while logistic regression offered a simpler alternative with acceptable accuracy. Seven predictive variables were identified: maximum procalcitonin level, max neutrophil percentage, maximum C-reactive protein level, minimum white blood cell count, venous catheter, age, and length of hospital stay. SHAP variable importance analysis identified maximum procalcitonin level, length of hospital stays, and max_neutrophil_percentage as the top three most important predictors. External validation in an independent MIMIC-IV database confirmed acceptable generalizability.

CONCLUSION

The predictive model developed for Gram-negative bloodstream infections in elderly patients demonstrates promising predictive ability. It overcomes the temporal limitations of traditional blood culture, enabling the identification of high-risk elderly patients at the earliest stage of infection. This approach provides clinicians with a valuable intervention window and represents a novel pathway toward precision infection control.