Machine learning-assisted prognostic model for mortality in ICU patients with culture-confirmed Klebsiella pneumoniae infection: a preliminary single-center retrospective cohort study.

BACKGROUND

Klebsiella pneumoniae (KP) is a major pathogen in intensive care units and is associated with substantial mortality. However, prognostic tools specifically developed for intensive care unit (ICU) patients with KP infection remain limited. This study aimed to identify stable prognostic factors and develop a predictive model for mortality in this population.

METHODS

A total of 587 adult ICU patients with culture-confirmed KP infection (first positive culture) at a tertiary hospital between August 2020 and February 2025 were retrospectively included. Candidate predictors were screened using least absolute shrinkage and selection operator (LASSO) and bootstrap-based stability selection to identify stable prognostic predictors. Using the final stable predictors, 11 models (including logistic regression and tree-based approaches) with 10-fold cross-validation were developed and compared. Model performance was evaluated using discrimination, calibration, and clinical utility. Generalized additive models (GAMs) were used to explore potential non-linear predictor-outcome associations.

RESULTS

Mortality occurred in 122/587 patients (20.8%). Five stable prognostic factors were identified: Acute Physiology and Chronic Health Evaluation II (APACHE II) score, lactate, viral co-infection, acute kidney injury (AKI), and the alveolar-arterial oxygen gradient (A-aDO2). Among the evaluated algorithms, logistic regression achieved the most favorable balance of discrimination, calibration, and interpretability, with a cross-validated area under the receiver operating characteristic curve (AUC) of 0.809. Non-linear or threshold-like patterns were observed for APACHE II (approximately 16.56), lactate (approximately 2.87 mmol/L), and the A-aDO2 (approximately 203.27 mmHg). Based on logistic regression, a nomogram and a web-based calculator were constructed for individualized risk estimation.

CONCLUSION

A parsimonious model based on five stable variables showed good performance for predicting mortality among ICU patients with KP infection. However, these findings are exploratory and require external validation before any clinical application.