gms | German Medical Science

Longitudinal datasets are prevalent in clinical and biomedical areas. Because of privacy concerns, high-quality synthetic data generation from these datasets is of great importance. Artificially building a cross-sectional structure from longitudinal data causes loss of important information about the changing pattern between two time-points [1]. Therefore, researchers need specific synthetic data generation methods for longitudinal data. We consider generative deep learning approaches for this task. A major challenge in clinical longitudinal datasets is the varying number of visits and the irregular time intervals between them, due to different needs of patients. Different deep generative approaches will correspondingly have different characteristics with a varying number of time points, sample size, number of variables, and general complexity of the data structure. In this work, we therefore compare different approaches with simulated and real data. As a starting point, we investigate a standard variational autoencoder (VAE). In this approach, we consider measurements from different visits and the time interval between visits simultaneously as input variables. As a more complex approach, we investigate a combination of ordinary differential equations with VAEs for dynamic modeling [2], which can infer individual-specific trajectories. As a third approach, we consider a combination of a transformer approach [3] with point processes. Specifically, we employ point processes for event modeling to handle the irregular time points in the longitudinal data (e.g. [4], [5]). We find that more complex and thus flexible methods do not necessarily perform better when it comes to simple structures. However, the performance of the rather naive simultaneous VAE considerably degrades as the data structure becomes more complex. Furthermore, when the number of visits changes from patient to patient, the standard VAE needs more parameters and a more complex architecture, which also is seen to decrease performance. In contrast, we demonstrate how the more complex approaches become feasible with increasing sample size. We thus provide more general guidance for deep generative modeling of longitudinal data for clinical settings in different scenarios.

The authors declare that a positive ethics committee vote has been obtained.