HyperNSDE: Generating Personalized Clinical Data with Neural Stochastic Differential Equations

The Data Dilemma in Medical AI

Machine learning in healthcare has long faced a fundamental paradox: there is never enough high-quality data, yet what data exists is locked behind strict privacy constraints. Real patient clinical records are difficult to obtain, and when obtained, they bring significant compliance and ethical risks. Synthetic patient data has emerged as a promising solution — using generative models to create realistic but non-identifiable clinical records that can ease data scarcity while sidestepping privacy concerns.

Yet achieving true realism is far from straightforward. A complete patient-level clinical record typically contains three tightly coupled components: heterogeneous static covariates (such as age, sex, and baseline medical history), irregularly sampled longitudinal trajectories (such as lab values changing over time), and the observation timestamps themselves, which carry their own information. In reality, these three components influence one another, yet prior work rarely models them together within a unified framework.

The Core Idea Behind HyperNSDE

HyperNSDE, introduced in a new paper on arXiv, is a continuous-time generative model designed precisely to address this gap. Its key innovation lies in using a hypernetwork to connect a latent Neural SDE with static patient representations.

Conventional approaches typically allow static features to determine only the "initial condition" of a trajectory, after which the dynamics decouple from the baseline features. HyperNSDE instead lets baseline features continuously shape the evolution of the entire trajectory through the hypernetwork — not just its starting point. More elegantly, this mechanism eliminates the need for a separate trajectory encoder, simplifying the overall model architecture. Meanwhile, stochastic latent dynamics capture the genuine variability present in generation paths, ensuring that synthetic trajectories are neither overly smooth nor homogeneous.

Modeling Observation Times as Well

An often-overlooked detail is that when a patient is observed is itself informative — more severe illness typically leads to more frequent testing. HyperNSDE addresses this by jointly modeling observation times through a latent-state-dependent intensity process, keeping "when measurements are taken" consistent with "what is measured." This makes the temporal structure of generated data far more faithful to real clinical settings.

This phenomenon is known in statistics as informative missingness or an informative observation process: whether and when an observation occurs is correlated with the value of the variable being observed. In settings like the ICU, clinicians tend to increase testing frequency as a patient's condition deteriorates, so "dense observations" is itself a signal of poor health status. Ignoring this mechanism — modeling only the observed values — implicitly assumes that observation times are independent of patient state, systematically underestimating the correlation between illness severity and observation frequency. Modeling observation times with an intensity process is the standard approach from point process theory: the intensity function describes the instantaneous probability of the next observation given the current state, and when this depends on the latent health state, the joint distribution of observation times and trajectories can be correctly captured.

Technical Design for Stable Training

Training a generative model on irregular stochastic paths is inherently unstable. The paper addresses this with a deterministic–stochastic path decomposition strategy, splitting trajectories into a predictable component and a stochastic perturbation, optimized with a non-adversarial signature-kernel objective. Compared to adversarial training, this approach avoids the instability issues common in GAN-style models and is better suited to continuous-time stochastic trajectories.

A Hypernetwork is a "network that generates the parameters of another network": rather than processing data directly, it takes some conditional input (here, static patient features) and dynamically produces the weights or parameters of another network (here, the Neural SDE). This design allows the primary network's behavior to vary continuously with the conditional input, without requiring a separate set of parameters for each patient. By contrast, traditional conditional generative models typically concatenate the conditioning information to the input vector without affecting the network's structure. The hypernetwork's advantage lies in its greater expressiveness — it allows a patient's baseline features to influence the entire logic of the dynamic process in a deeper, more structured way, rather than simply shifting the initial state.

Neural SDEs replace the drift and diffusion terms of classical stochastic differential equations with neural networks, enabling continuous-time modeling of stochastic dynamical processes. They retain the rigorous mathematical treatment of randomness from classical SDEs while gaining the flexible fitting capacity of neural networks, making them particularly well-suited to representing the irregular, stochastic evolution of patient vital signs over time.

The Signature Kernel combines path signature theory with kernel methods. A path signature maps a time series or continuous path to a collection of iterated integrals, compactly capturing the geometric and sequential properties of the path. It is naturally robust to irregular sampling and variable path lengths. The signature kernel defines a similarity measure between two paths and can serve as a training objective for generative models: by minimizing the distance between real and synthetic trajectories in signature kernel space, the generator is driven to produce paths with similar statistical properties. Compared to GAN-based adversarial training, a signature-kernel objective is non-adversarial — there is no need to simultaneously optimize a generator and discriminator, eliminating common issues like mode collapse and training oscillation. This is especially valuable for continuous-time stochastic trajectories, where GAN instability is further amplified by the temporal dimension.

Experimental Results and Key Findings

The paper validates HyperNSDE on both simulated and real clinical datasets. Results show a clear improvement in observation-time fidelity, with overall performance that is competitive across the board.

One particularly noteworthy analysis is the so-called "matched-grid" experiment. The authors find that predictive and correlation metrics are sensitive to the regularity of the observation grid and the smoothness of trajectories. In other words, without controlling for these factors, comparing predictive performance across models can be misleading — the more regular the data and the smoother the trajectories, the better certain metrics will look, but this does not necessarily reflect the model's true generative capability. This finding raises an important caution for how synthetic medical data is evaluated.

Implications for Medical Data Synthesis

The value of HyperNSDE lies not simply in proposing yet another generative model, but in attempting to unify three components — static features, longitudinal trajectories, and observation times — that have long been treated in isolation, within a single continuous-time framework. For medical machine learning researchers, this means that future synthetic data can be more temporally faithful, supporting more reliable downstream tasks such as prognostic prediction and clinical decision support.

At the same time, the paper's observation that evaluation metrics are influenced by data regularity serves as a reminder to the broader field that reporting synthetic data quality requires more careful experimental design. Methodological reflections of this kind often carry more long-term value than any single performance number.

It is worth noting that, as an arXiv preprint, HyperNSDE has not yet undergone peer review, and its generalization to larger-scale real-world data remains to be further validated.