Machine learning prediction of in-hospital mortality risk among hospitalized patients with secondary bloodstream infection: a retrospective cohort study

Secondary bloodstream infection (SBSI) carries substantial mortality burden. Dedicated mortality prediction models with interpretable predictions remain limited. To develop an interpretable machine learning model for predicting in-hospital mortality in SBSI patients and establish a clinically applicable web-based tool. Data from 340 SBSI patients (January 2020–December 2024) were analyzed. Candidate variables were selected using the union of multivariate Cox regression and LASSO-Cox regression. Five prognostic models were constructed including Cox regression, LASSO-Cox regression, Random Survival Forest (RSF), Gradient Boosting Machine (GBM), and XGBoost, with 10 selected variables as predictors. Model performance was evaluated using time-dependent Area Under the Receiver Operating Characteristic Curve (AUC) at 7, 14, and 28 days as the primary metric. Internal validation was conducted using bootstrap resampling with 1,000 iterations. The optimal RSF model was interpreted using SHapley Additive exPlanations (SHAP) analysis. An interactive web-based prediction tool was subsequently developed. In-hospital all-cause mortality rate was 34.41% (117/340). Among the five models, RSF demonstrated superior discriminative ability with AUC values of 0.869 (95% CI: 0.824–0.914), 0.897 (95% CI: 0.856–0.937), and 0.878 (95% CI: 0.822–0.933) at 7, 14, and 28 days, respectively, substantially outperforming traditional Cox regression and LASSO-Cox models. RSF exhibited excellent calibration with corrected C-index of 0.8809. Decision curve analysis confirmed RSF’s clinical superiority, demonstrating the greatest net benefit across a wide range of threshold probabilities. SHAP analysis identified log Lactate Dehydrogenase, Multidrug-Resistant Organism infection, and log creatinine as the top three predictive features. An interactive web-based prediction tool was successfully developed enabling real-time risk assessment at 7, 14, and 28 days. A interpretable machine learning model (RSF) for SBSI mortality prediction was successfully established. SHAP-based transparency and the interactive web-based tool facilitate clinical implementation and support personalized risk stratification for improved patient outcomes.