Generative Models for Clinical Data Augmentation

Generative models can assist in augmenting clinical datasets to address the data scarcity that limits the application of complex predictive models. The opportunity remains to evaluate how generative models can be applied to the augmentation of clinical datasets while satisfying a number of requirements. First, the generative model should be suitable for modeling high-dimensional tabular data with missingness, as all clinical data have the tabular representation of the schema of medical record. Second, the model should support proven techniques for dealing with imbalanced class distributions, since the clinical task usually appears as predictive modeling with imbalanced class distribution. Finally, prior to the use of augmented data for predictive modeling, it needs to demonstrate that augmentation of training samples using generative models can improve model fairness-aware fairness with respect to sensitive attributes.