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

ObjectiveTo develop and validate a robust machine learning-based prediction model for assessing the risk of thrombotic events in critically ill cancer patients during their ICU stay.MethodsThis retrospective observational study utilized data from 1892 cancer patients in the MIMIC-IV database for model development and internal validation. A stringent data preprocessing pipeline was applied, including multiple imputation for missing data, exclusion of outliers, and the use of the Synthetic Minority Over-sampling Technique (SMOTE) to address class imbalance. Feature importance was evaluated using SHAP, leading to the selection of six key predictors. Nine machine learning models were constructed and compared. Model performance was assessed using the Area Under the Curve (AUC), F1-score, recall, Matthews correlation coefficient (MCC), accuracy, and specificity. The optimal model was selected, calibrated, and interpreted using SHAP. Its clinical utility was further evaluated via calibration curves and decision curve analysis (DCA). Finally, external validation was performed on an independent dataset of 200 patients from our institution.ResultsThe CatBoost model demonstrated superior performance. In internal validation, the calibrated model achieved an AUC of 0.855 (95% CI: 0.797-0.913), with a sensitivity of 0.971 and a specificity of 0.753 at an optimal threshold of 0.245. In external validation, the model maintained strong performance with an AUC of 0.83 (95% CI: 0.742-0.918), sensitivity of 0.968, and specificity of 0.698. SHAP analysis identified "history of thrombosis" as the most influential predictor. Decision curve analysis confirmed the model's clinical utility across a wide risk threshold range (0.25-0.75). The final model was deployed as an online platform to facilitate real-time, individualized risk assessment.ConclusionThe developed CatBoost model exhibits excellent discriminatory power, good calibration, and favorable clinical interpretability for predicting thrombosis risk in critically ill cancer patients. It serves as a promising and reliable clinical decision support tool to guide personalized thromboprophylaxis and improve patient outcomes.