A Comparison of Synthetic Data Generation Techniques for Control Group Survival Data in Oncology Clinical Trials: Simulation Study (Preprint)

Abstract

UNSTRUCTURED Introduction: Synthetic patient data (SPD) generation for survival analysis in oncology trials holds significant potential for accelerating clinical development. Various machine learning methods, including CART, random forest (RF), Bayesian network (BN), and CTGAN, have been employed for this purpose, but their performance in reflecting actual patient survival data remains under investigation. Method: Utilizing multiple clinical trial datasets, survival SPD was generated and evaluated using mean survival time (MST), hazard ratio distance (HRD), and visual analysis of Kaplan‒Meier (KM) plots. Each method's ability to mimic the statistical profile of real patient data was compared. Results: CART consistently demonstrated promising results across various evaluation metrics, outperforming other methods such as RF, BN, and CTGAN. However, while RF is known for its high generalization performance, CART exhibited closer resemblance to actual data, emphasizing the importance of similarity in SPD generation. Conclusion: It seems that the reason that CART indicated better similarity than RF is that the ensemble learning of RF prevents overfitting, and CART overfits SPD. In SPD generation, the statistical properties close to the actual data should be the focus, not a well-generalized prediction model. Both the BN and CTGAN methods cannot accurately reflect the statistical profile of the actual data, primarily due to the small datasets. As a method for generating SPD for survival data from small datasets, such as clinical trial data, CART is considered the most effective method. Additionally, it is necessary to improve CART-based generation methods by incorporating feature engineering and other methods in future work.