Figure 2
From: ClinicalGAN: powering patient monitoring in clinical trials with patient digital twins
ClinicalGAN consists of six network components. The auto-encoder (1) (AE) transforms the data from feature space to compressed latent space; (2) supervisor network (S) to assist the generator with a supervised objective to further guide the learning process ; (3) termination network (\(T_r)\) models the drop-off likelihood of patients over their simulated journeys; (4) conditional generator (G) auto-regressively generates personalized patient journeys; (5) conditional discriminator (D) aids in unsupervised learning objective of the generator via adversarial feedback and finally; (6) auxiliary classifier (AuxC) assists the generator in learning the relation between input conditionals and their sequences. AE, \(T_r\) and S undergo pre-training on their respective tasks independently before being jointly trained with G, D and AuxC.
