Development and validation of venous thrombosis risk prediction model and scale for lung cancer patients in intensive care unit based on interpretable machine learning.

BACKGROUND: Venous thromboembolism (VTE) is a prevalent and grave complication in patients with lung cancer, with an incidence of more than 10% and a substantial increase in mortality risk. Therefore, it is imperative to construct a high-precision, interpretable, predictive model and develop a practical risk assessment scale for the early identification of high-risk lung cancer patients. METHODS: Based on data from 3,694 intensive care unit (ICU) patients with lung cancer in the Medical Information Mart for Intensive Care IV (MIMIC-IV), eICU, and Northwestern ICU (NWICU) databases, independent predictors were jointly screened using the Boruta algorithm, variance inflation factor (VIF) analysis, and least absolute shrinkage and selection operator (LASSO) regression. Predictive models were constructed using logistic regression, Adaptive Boosting (AdaBoost), Categorical Boosting (CatBoost), K-nearest neighbors (KNN), Light Gradient Boosting Machine (LightGBM), Naïve Bayes (NBS), random forest, support vector machine (SVM), and eXtreme Gradient Boosting (XGBoost). Model interpretability was elucidated using SHapley Additive exPlanations (SHAP) values. A practical clinical scale was developed based on the optimal model, and the cutoff value was determined using the receiver operating characteristic (ROC) curve and the Youden index. RESULTS: The CatBoost model demonstrated the best performance [training set, area under the curve (AUC) =0.7648; internal validation set, AUC =0.6437; eICU external validation, AUC =0.6822; NWICU external validation, AUC =0.7153]. The clinical scoring system comprised nine indicators [antiplatelet, vasoactive, atrial fibrillation and flutter (AF_AFL), heart failure (HF), respiratory failure (RF), creatinine, platelet count (PLT), partial thromboplastin time (PTT), white blood cell count (WBC)]. Cutoff values of 46 (sensitivity 43.1%) and 66.5 (specificity 96.9%) were used for screening high-risk patients and ruling out low-risk populations, respectively. The validation set AUC was 0.604, with external validation AUCs of 0.6822 for the eICU cohort and 0.7153 for the NWICU cohort. SHAP analysis revealed that PTT and RF contributed significantly to the prediction. CONCLUSIONS: The CatBoost model demonstrated the best performance (training set, AUC =0.7648; internal validation set, AUC =0.6437; eICU external validation, AUC =0.6822; NWICU external validation, AUC =0.7153). The clinical scoring system comprised nine indicators [red blood cell distribution width (RDW), activated partial thromboplastin time (APTT), mean corpuscular hemoglobin (MCH), length of hospital stay, and duration of mechanical ventilation]. Cutoff values of 46 (sensitivity 43.1%) and 66.5 (specificity 96.9%) were used for screening high-risk patients and ruling out low‑risk populations, respectively. The validation set AUC was 0.604, with external validation AUCs of 0.6822 for the eICU cohort and 0.7153 for the NWICU cohort. SHAP analysis revealed that PTT and RF contributed significantly to the prediction.