Statistical and Machine Learning Approaches for Virtual Population Generation in In-Silico Cardiovascular Trials

This study presents a comprehensive methodology for generating virtual human cardiovascular datasets for in-silico clinical trials, particularly for the evaluation of novel stents within the InSilc project. Our approach integrates a statistical modeling technique with a machine learning (ML)-based generative model to improve the accuracy and realism of virtual populations. The statistical model is based on a multivariate normal distribution and incorporates methods to address missing data and non-positive definite covariance matrices. The ML-based approach leverages Conditional Tabular Generative Adversarial Networks (CTGAN) to synthesize patient populations while maintaining the statistical integrity of real-world datasets. Initially, we simulated 20% of the test data for validation purposes. Following successful validation, we expanded the simulation to generate a virtual population of 10,000 patients. A comparative analysis revealed that the statistical model demonstrated higher accuracy in anatomical parameter prediction, whereas the ML approach excelled in capturing complex inter-variable relationships. The combination of these techniques enhances the ability to simulate diverse patient populations, thereby improving the robustness of in-silico clinical trials.Clinical Relevance - This methodology advances in-silico clinical trials by reducing reliance on traditional resource-intensive methods, improving trial efficiency, cost-effectiveness, and patient safety.