OBJECTIVE  To develop an interpretable machine learning model to predict the risk of poor prognosis in patients with Klebsiella pneumoniae bloodstream infection （BSI）.

METHODS  Clinical data from 393 patients with K. pneumoniae BSI admitted to the Affiliated Jiangning Hospital of Nanjing Medical University between Jan. 1, 2019, and Dec. 31, 2024, were retrospectively analyzed. Seven machine learning methods, including logistic regression, decision tree, random forest, extreme gradient boosting, light gradient boosting machine, support vector machine and artificial neural network, were employed to construct prediction models for poor prognosis in these patients. Model predictive performance was evaluated through metrics such as F1 score and area under the receiver operating characteristic curve （AUC-ROC）. SHAP values were adopted to assess the contribution of each feature in the best-performing model.

RESULTS  Among the seven machine learning models evaluated （logistic regression, decision tree, random forest, extreme gradient boosting, light gradient boosting machine, support vector machine and artificial neural network）, the random forest model demonstrated the best predictive performance. The accuracy rates for the models in the test set were 0.853, 0.844, 0.933, 0.875, 0.875, 0.718 and 0.853, respectively. The precision rates were 0.957, 0.940, 0.974, 0.957, 0.957, 0.897 and 0.957, respectively. The F1 scores were 0.920, 0.885, 0.949, 0.918, 0.918, 0.823 and 0.921, respectively. The AUC values were 0.985, 0.959, 0.987, 0.982, 0.985, 0.983 and 0.985, respectively. SHAP analysis identified key influencing factors, including septic shock, prior antimicrobial use and ICU/EICU admission history.

CONCLUSIONS  The random forest model exhibited optimal performance in predicting the prognosis of K. pneumoniae BSI. SHAP analysis highlighted critical risk factors, thereby providing valuable foundation for auxiliary diagnosis.