HyperNSDE
A continuous-time generative model that combines hypernetworks with neural stochastic differential equations, designed to synthesize personalized clinical patient data. The model jointly captures static covariates, irregular longitudinal trajectories, and observation times, enabling the generation of realistic patient records that reflect individual-level heterogeneity in clinical settings.
Timeline (last 90 days)
HyperNSDE is a continuous-time generative model that connects latent neural stochastic differential equations with static patient representations via a hypernetwork to generate synthetic clinical data.
HyperNSDE allows baseline static features to continuously shape the evolution of the entire trajectory through the hypernetwork, not merely the initial conditions.
HyperNSDE's mechanism requires no additional trajectory encoder, simplifying the model architecture.
HyperNSDE jointly models observation times through an intensity process that depends on the latent state.
HyperNSDE adopts a deterministic–stochastic path decomposition strategy, splitting the trajectory into a predictable component and a stochastic perturbation component.
HyperNSDE uses a non-adversarial signature kernel objective function for optimization, avoiding the training instability issues of GAN-type models.
The paper validates HyperNSDE on simulated datasets and real clinical datasets, with results showing significant improvements in observation time fidelity.
Researchers found that predictive metrics and correlation metrics are affected by the regularity of the observation grid and the smoothness of trajectories, such that the more regular the data and the smoother the trajectories, the better certain metrics appear.
HyperNSDE unifies static features, longitudinal trajectories, and observation timestamps — three components that have long been handled separately — into a continuous-time framework.
HyperNSDE is an arXiv preprint and has not yet undergone peer review.
All Facts (10)
HyperNSDE is a continuous-time generative model that connects latent neural stochastic differential equations with static patient representations via a hypernetwork to generate synthetic clinical data.
50%UnverifiedHyperNSDE allows baseline static features to continuously shape the evolution of the entire trajectory through the hypernetwork, not merely the initial conditions.
50%UnverifiedHyperNSDE's mechanism requires no additional trajectory encoder, simplifying the model architecture.
50%UnverifiedHyperNSDE jointly models observation times through an intensity process that depends on the latent state.
50%UnverifiedHyperNSDE adopts a deterministic–stochastic path decomposition strategy, splitting the trajectory into a predictable component and a stochastic perturbation component.
50%UnverifiedHyperNSDE uses a non-adversarial signature kernel objective function for optimization, avoiding the training instability issues of GAN-type models.
50%UnverifiedThe paper validates HyperNSDE on simulated datasets and real clinical datasets, with results showing significant improvements in observation time fidelity.
50%UnverifiedResearchers found that predictive metrics and correlation metrics are affected by the regularity of the observation grid and the smoothness of trajectories, such that the more regular the data and the smoother the trajectories, the better certain metrics appear.
50%UnverifiedHyperNSDE unifies static features, longitudinal trajectories, and observation timestamps — three components that have long been handled separately — into a continuous-time framework.
50%UnverifiedHyperNSDE is an arXiv preprint and has not yet undergone peer review.
50%