Purpose <p>Sepsis is a complex and life-threatening condition that arises from an abnormal and dysregulated host response to an infection. Developing a practical and interpretable machine learning model to predict mortality in patients with sepsis is essential. To address the limitations of existing sepsis mortality prediction models (insufficient validation and limited interpretability), we developed a streamlined extreme gradient boosting (XGBoost) model that predicts mortality risk using only 7 routine clinical features.</p> Methods <p>Medical Information Mart for Intensive Care-IV (MIMIC-IV) data were used to train and internally validate our model. External validation was performed using the eICU-Collaborative Research Database (eICU-CRD). The area under the receiver operating characteristic curve (AUC) was used as the key evaluation index. Differences in model performances were assessed using DeLong's test. The optimal model was simplified using a recursive feature elimination algorithm, and model interpretability was evaluated using SHAP.</p> Results <p>Our 7-feature model demonstrated strong performance across training (AUC 0.920), internal validation (AUC 0.900) and external validation (AUC 0.839) cohorts. The selected features—platelet count (at discharge), leukocyte levels (post-antibiotic therapy and at discharge), BUN (post-antibiotic therapy), anion gap (post-antibiotic therapy and at discharge), and ICU length of stay—are objective (eliminating subjective/recall bias) and readily available in clinical practice. SHAP analysis showed that the prognostic contributions of these variables differed before and after antibiotic&#xa0;treatment.</p> Conclusion <p>The XGBoost model, built on 7 routine and objective clinical features, quantifies inter-feature relationships to provide mortality risk and treatment thresholds, thereby&#xa0;enabling clinicians to formulate evidence-based therapeutic strategies.</p>