To develop and externally validate an explainable machine-learning model for early prediction of sepsis in ICU patients with hemorrhagic stroke.
Approach:
Study Design: Conducted a multicenter retrospective cohort study using three large publicly available critical care databases.
Data Sources: Utilized the MIMIC-IV database for model derivation and internal validation, MIMIC-III as a supplementary validation dataset, and the eICU Collaborative Research Database for true multicenter external validation.
Model Development: Integrated multiple feature-selection strategies and compared various machine-learning algorithms to identify the optimal predictive model.
Interpretability: Applied Shapley additive explanations (SHAP) to enhance model interpretability and created an interactive web-based risk calculator.
Key Findings:
Machine-learning techniques improved predictive performance for sepsis compared to conventional scoring systems.
The model demonstrated robust interpretability through the use of SHAP.
The web-based risk calculator facilitates real-time bedside risk stratification.
Interpretation:
The integrative prediction framework aims to provide accurate and clinically interpretable risk estimation for early identification of high-risk patients in neurocritical care settings.
Limitations:
Models developed using single-center or homogeneous datasets may limit generalizability.
Focus primarily on predictive performance without robust interpretability frameworks may restrict clinical adoption.
Few studies have translated prediction models into accessible clinical tools.
Conclusion:
The study presents a novel machine-learning approach for predicting sepsis in hemorrhagic stroke patients, with a focus on interpretability and clinical application.
