Machine Learning Algorithms to Predict Venous Thromboembolism in Patients With Sepsis in the Intensive Care Unit:...

Area of Science:

• Intensive care medicine
• Machine learning applications in healthcare
• Thrombosis research

Background:

• Venous thromboembolism (VTE) is a significant complication for intensive care unit (ICU) patients with sepsis.
• Existing risk assessment tools may not adequately address sepsis-specific factors or complex patient data interactions.

Purpose of the Study:

• To develop and validate an interpretable machine learning (ML) model for early VTE prediction in ICU patients with sepsis.
• To improve upon conventional risk stratification methods with a sepsis-specific approach.

Main Methods:

• Utilized the Medical Information Mart for Intensive Care IV database for model development and internal validation, with external validation from Changshu Hospital.
• Developed nine ML models, including light gradient boosting machine, and employed Shapley Additive Explanations (SHAP) for interpretability.
• Evaluated model performance using area under the curve (AUC), calibration, and decision curve analysis, with subgroup analysis by sepsis severity.

Main Results:

• The light gradient boosting machine model achieved high performance (AUC 0.956 internally, 0.786 externally).
• The model demonstrated robust generalization and improved discrimination in severe sepsis cases (AUC 0.816).
• Key predictors included central venous catheterization, serum chloride/bicarbonate, arterial catheterization, and prolonged partial thromboplastin time, with both linear and nonlinear relationships identified.

Conclusions:

• A high-performing, interpretable ML model for VTE prediction in ICU sepsis patients was successfully developed.
• The model shows strong performance across different cohorts and patient severities, particularly in severe sepsis.
• This tool offers potential for personalized VTE prophylaxis and early diagnosis in critical care settings.