Development and internal validation of an interpretable machine-learning model for predicting sepsis within 48 hours after admission in emergency-admitted patients with pulmonary infection

To develop and internally validate a machine-learning model using routinely available early clinical variables to predict clinically documented sepsis within 48 h after admission among patients with pulmonary infection in a single-center development-only setting. Adult patients with pulmonary infection admitted through the Department of Emergency Internal Medicine of Ningde Municipal Hospital between December 2021 and December 2025 were retrospectively included. The primary outcome was clinically documented sepsis within 48 h after admission, adjudicated retrospectively using a clinical-record-based Sepsis-3 conceptual framework. Patients were randomly divided into training and validation sets at a 7:3 ratio. Predictors were selected using LASSO regression and bidirectional stepwise logistic regression based on the Akaike information criterion. Six models were developed using the same predictor set: logistic regression, random forest, support vector machine, neural network, gradient boosting machine, and XGBoost. Model performance was assessed using discrimination, calibration, decision curve analysis, bootstrap confidence intervals, DeLong testing, bootstrap 0.632+ optimism correction, threshold sensitivity analyses, class-imbalance sensitivity analyses, and supplementary comparisons with NEWS/NEWS2, MEWS, and SIRS. Model-agnostic SHAP was used to interpret the selected SVM model. A total of 425 patients were included, of whom 106 (24.94%) developed clinically documented sepsis within 48 h. The final model included Glasgow Coma Scale score, respiratory rate, systolic blood pressure, procalcitonin, C-reactive protein, and lactate. In the validation set, SVM achieved an AUC of 0.869 (95% CI, 0.805–0.933), with sensitivity of 0.812, specificity of 0.760, PPV of 0.531, NPV of 0.924, and F1 score of 0.642. LR showed similar sensitivity and discrimination, with an AUC of 0.872 (95% CI, 0.808–0.937), sensitivity of 0.812, specificity of 0.771, PPV of 0.542, NPV of 0.925, and F1 score of 0.650. XGBoost achieved the numerically highest AUC of 0.879 (95% CI, 0.819–0.939), but had lower sensitivity of 0.625. Pairwise DeLong testing showed no significant AUC difference between SVM and LR or XGBoost. Model-agnostic SHAP analysis of the SVM model identified lactate, procalcitonin, Glasgow Coma Scale score, respiratory rate, C-reactive protein, and systolic blood pressure as contributors to model output. In this single-center development-only study, SVM based on six routinely available early variables showed a sensitivity-favorable profile for identifying patients with pulmonary infection at risk of clinically documented sepsis within 48 h after admission. LR provided a transparent nomogram-based statistical reference, whereas XGBoost showed higher specificity but lower sensitivity. These models may support early risk stratification and focused monitoring in conjunction with standardized early warning scores and clinical judgment, but they cannot be used as stand-alone rule-out tools or deployed clinically without external validation and prospective evaluation.