DFG Project Number: FK 515800538

Machine learning in clinical research often suffers from small, imbalanced, and privacy-restricted datasets. These limitations make it difficult to train robust and generalizable models, especially when sensitive patient data cannot be shared across institutions. Synthetic data generation offers a promising solution by producing artificial yet statistically representative datasets that preserve the structure and relationships of the original data while maintaining patient privacy.

Building on our previous work on convex space learning, we have developed several state-of-the-art algorithms, including LoRAS, ProWRAS, and ConvGeN, for generating synthetic data in imbalanced tabular datasets. The newly funded DFG project extends this research through the development of the Convex-space Generative Network for Tabular Data (NextConvGen). NextConvGeN will generalize convex space learning beyond imbalanced classification, enabling the generation of synthetic data with both continuous and categorical features, even in the presence of missing values.

The project investigates how convex space learning can be tailored for specific machine learning tasks, including regression, classification, and unsupervised patient stratification. Through collaborations with clinical partners at Universitätsmedizin Rostock, NextConvGeN will be validated on real-world clinical datasets involving osteoporosis and kidney transplantation.

All methods and results will be made open-source to ensure transparency and reproducibility. By enabling realistic, privacy-preserving synthetic data generation, NextConvGeN aims to advance the use of artificial intelligence in medicine and contribute to more inclusive, data-driven clinical research.