Related Experiment Video
Updated: Jan 17, 2026
A Data-Driven Approach to Quantifying Immune States in Sepsis
Published on: February 7, 2025
A Retrospective Cohort Study on 28-Day Mortality in Immunosuppressed Sepsis: An Interpretability-Based Predictive
Zhiru Zhong1, Huiwei He, Zhiying Lin
1Jiangxi Provincial People's Hospital, The First Affiliated Hospital of Nanchang Medical College, Nanchang, China.
Background:
Sepsis in immunosuppressed patients is associated with significantly higher mortality rates, yet predictive models tailored to this high-risk population remain limited. This study aims to develop an interpretable machine learning model to predict 28-day mortality in immunosuppressed sepsis patients, with a focus on model transparency and clinical applicability.
Methods:
A retrospective cohort study was conducted using clinical, laboratory, and demographic data from immunosuppressed sepsis patients. Feature selection was performed using LASSO regression, followed by the development of predictive models, including XGBoost. The model's performance was evaluated using the area under the receiver operating characteristic curve (AUROC). To enhance clinical interpretability, Shapley additive explanations (SHAP) were employed to provide insights into the contribution of individual features to mortality predictions.
Results:
The final model identified key predictors of 28-day mortality, including lactate levels, red cell distribution width, platelet count, and Sequential Organ Failure Assessment score. XGBoost demonstrated superior predictive accuracy with an AUROC of 0.93 (95% confidence interval: 0.90-0.96), outperforming other models. SHAP analysis revealed that elevated lactate levels and reduced platelet counts were strong risk factors for mortality, while lower lactate and higher platelet counts were protective. The model's interpretability provided clear insights into the role of each predictor, facilitating individualized risk stratification.
Conclusion:
The XGBoost model, combined with SHAP analysis, offers an accurate and interpretable tool for predicting 28-day mortality in immunosuppressed sepsis patients. This approach enhances clinical decision-making by providing transparent insights into the factors driving mortality risk, thus supporting personalized and timely interventions aimed at improving patient outcomes.
Insights
An interpretable machine learning model accurately predicts 28-day mortality in immunosuppressed sepsis patients. Key factors include lactate, platelet count, and SOFA score, aiding clinical decisions.
Area of Science:
- Machine Learning in Medicine
- Critical Care Medicine
- Immunology
Background:
- Sepsis poses a higher mortality risk for immunosuppressed patients.
- Existing predictive models for this group are limited.
- Need for transparent, clinically applicable predictive tools.
Purpose of the Study:
- Develop an interpretable machine learning model.
- Predict 28-day mortality in immunosuppressed sepsis patients.
- Focus on model transparency and clinical utility.
Main Methods:
- Retrospective cohort study of immunosuppressed sepsis patients.
- XGBoost model development with LASSO feature selection.
- Performance evaluated by AUROC; interpretability via SHAP analysis.
Main Results:
- XGBoost model achieved high accuracy (AUROC 0.93).
- Key predictors identified: lactate, RDW, platelets, SOFA score.
- SHAP analysis clarified risk factors (e.g., elevated lactate, low platelets).
Conclusions:
- XGBoost with SHAP provides accurate, interpretable mortality prediction.
- Enhances clinical decision-making for immunosuppressed sepsis patients.
- Supports personalized interventions and improved outcomes.
