CatBoost Machine Learning Model for Thrombosis Risk Prediction in Critically Ill Cancer Patients: A MIMIC-IV Database...

Area of Science:

• Computational biology and bioinformatics
• Machine learning in healthcare
• Critical care medicine

Background:

• Critically ill cancer patients face a high risk of thrombotic events.
• Accurate risk assessment is crucial for effective thromboprophylaxis.
• Existing prediction models may lack robustness for this specific population.

Purpose of the Study:

• To develop and validate a machine learning model for predicting thrombosis risk in critically ill cancer patients.
• To identify key predictors of thrombotic events in this cohort.
• To assess the clinical utility and interpretability of the developed model.

Main Methods:

• Retrospective analysis of 1892 cancer patients from the MIMIC-IV database for model development.
• Data preprocessing included imputation, outlier exclusion, and SMOTE for class imbalance.
• Nine machine learning models were trained and evaluated; CatBoost was selected based on AUC, F1-score, MCC, and specificity.

Main Results:

• The CatBoost model achieved an AUC of 0.855 (internal) and 0.83 (external validation).
• High sensitivity (0.971 internal, 0.968 external) and good specificity (0.753 internal, 0.698 external) were observed.
• SHAP analysis identified 'history of thrombosis' as the most significant predictor; decision curve analysis confirmed clinical utility.

Conclusions:

• The developed CatBoost model demonstrates strong discriminatory power and calibration for predicting thrombosis risk.
• The model is clinically interpretable and serves as a valuable decision support tool.
• This tool can guide personalized thromboprophylaxis strategies, potentially improving outcomes for critically ill cancer patients.