Objective:

To develop and validate a machine learning model for predicting 30-day mortality in immunocompromised sepsis patients using clinical data within 6 h of ICU admission.

Approach:

• Study Design: Retrospective cohort study utilizing data from MIMIC-IV and eICU databases.
• Patient Selection: Included adult immunosuppressed patients meeting Sepsis-3 criteria.
• Data Collection: Clinical indicators within 6 h of ICU admission were extracted.
• Model Development: Constructed predictive models using seven machine learning algorithms.
• Model Assessment: Evaluated discriminative capabilities using AUC, calibration curves, and DCA.
• External Validation: Best-performing model validated using eICU dataset and SHAP analysis.

Key Findings:

• Included 2,494 immunosuppressed sepsis patients with a 30-day mortality rate of 33.4%.
• Final prediction model included 10 feature variables: Weight, APS-III score, Urine output, Prothrombin Time, Blood Urea Nitrogen, SOFA score, Red Blood Cell count, Platelet count, Age, and Mean Corpuscular Hemoglobin Concentration.
• Support Vector Machine (SVM) model showed best predictive performance with an AUC of 0.794 in the validation set.
• External validation on the eICU dataset yielded an AUC of 0.847.

Interpretation:

The SVM-based prediction model effectively predicts 30-day mortality risk in immunosuppressed sepsis patients using easily obtainable clinical indicators.

Limitations:

• Retrospective design may introduce bias.
• Data derived from specific databases may limit generalizability.

Conclusion:

The study developed a practical tool for early identification of high-risk immunosuppressed sepsis patients.