Incremental learning-based risk prediction for carbapenems resistance in Klebsiella pneumoniae

OBJECTIVE  To achieve early prediction for carbapenems resistance of Klebsiella pneumoniae based on the IXGBoost model under an incremental learning framework.

METHODS  Medical records of 35 060 patients infected with K. pneumoniae at the Central Hospital of Dalian University of Technology from Jan. 2015 to Dec. 2024 were collected. After data cleaning and preprocessing, significant characteristic variables related to the risk of drug resistance were extracted. The dataset was randomly split into training and test sets at an 8:2 ratio. The initial batch was based on data from 2015 to 2020, followed by incremental training with weekly new data from 2021 to 2024. Models, including support vector machine, decision tree, random forest, logistic regression, XGBoost and LightGBM, were selected for comparison. Model performance was evaluated on the test set through K-fold cross-validation.

RESULTS  The IXGBoost model not only enabled real-time learning and prediction for data, but also outperformed other models in performance indicators, including precision （0.9823）, recall rate （0.9813）, F1-score （0.9818） and AUC （0.8421）.

CONCLUSIONS  This study confirms the effectiveness of incremental learning in predicting the carbapenems resistance of K. pneumoniae, identifies key risk factors influencing resistance development, and is of great significance for clinical prevention and understanding the underlying mechanisms of drug resistance.