REBALANCING DATA FOR CANCER-ASSOCIATED THROMBOSIS: COMPARISON OF DIFFERENT RESAMPLING APPROACH

Objective: Cancer-associated thrombosis (CAT) presents a complex challenge in oncology, exacerbated by data imbalances in related datasets that often lead to suboptimal outcomes in machine learning (ML) classification. Many ML algorithms were originally designed for balanced datasets, prompting this study to evaluate the interaction between logistic regression (LR) and eXtreme Gradient Boost (XGBoost) and data resampling techniques for improving prediction on imbalances in Malaysian data on CAT (MDCAT). Methods: Random oversampling (ROS), random undersampling (RUS), and a combined oversampling and undersampling approach (BOTH) were applied to MDCAT dataset. Classification tasks were performed using LR and XGBoost in R version 4.3.1. Classifier performance was assessed using accuracy, sensitivity, specificity, and the area under the ROC curve (AUROC) to evaluate the impact of different resampling techniques. Results: Applying LR and XGBoost to the imbalanced data revealed high specificity but low sensitivity in testing samples. A substantial decline in XGBoost performance was observed, with the AUC decreasing from 0.794 in training to 0.381. Metastasis, surgery, and Indian ethnicity showed statistically significant associated with the CAT event across all resampling techniques. Among XGBoost models, oversampling (XO) exhibited excellent training performance (Accuracy 0.99; AUC 0.98) but showed a large performance drop on the test set (Accuracy 0.82; AUC 0.72). Among LR models, logistic undersampling yielded the highest training accuracy (0.83) and AUC of 0.82. Tuning amplified the differences between resampling strategies and highlighted clear classifier–resampling interactions. XGBoost benefited most, particularly when trained on mixed and oversampled datasets, while LR remained comparatively stable. Conclusion: This study demonstrated that the effectiveness of prediction models in imbalanced MDCAT dataset is strongly influenced by the interaction between classifier characteristics and resampling strategies. A tuned XGBoost model with mixed resampling outweighed the benefits of LR’s simplicity and stability, making it our recommended approach given the primary importance of AUC.