Predicting a failure of postoperative thromboprophylaxis in non-small cell lung cancer: A stacking machine learning approach

• DOI: 10.1371/journal.pone.0320674
• Publication Date: 2025-04-01T21:11:42Z
• Authors: Ligang Hao; Junjie Zhang; Yonghui Di; Zheng Qi; Peng Zhang

Background Non-small-cell lung cancer (NSCLC) and its surgery significantly increase the venous thromboembolism (VTE) risk. This study explored VTE risk factors established a machine-learning model to predict failure of postoperative thromboprophylaxis. Methods retrospective included patients with NSCLC who underwent between January 2018 November 2022. The were randomized 7:3 training test sets. Nine machine learning models constructed. three most predictive classifiers chosen as first layer stacking model, logistic regression was second meta-learning model. Results 362 patients, including 58 (16.0%) VTE. Based on multivariable analysis, age, platelets, D-dimers, albumin, smoking history, epidermal growth factor receptor (EGFR) exon 21 mutation used develop nine models. LGBM Classifier, RandomForest GNB for area under received operating characteristics curve (ROC-AUC), accuracy, sensitivity, specificity in training/test set 0.984/0.979, 0.949/0.954, 0.935/1.000, 0.958/0.887, respectively. In validation set, final demonstrated an ROC AUC 0.983, accuracy 0.937, sensitivity 0.978, 0.947. decision analyses revealed high benefits. Conclusion based EGFR clinical had value NSCLC.