Speaker
Description
High-fidelity synthetic data is a critical frontier for
disease modelling, yet few generalized methods exist to quantify
whether generated virtual patients maintain physiological resemblance.
Building on our previous work in machine learning-enabled immune
profiling [1], we introduce a robust framework to assess the quality
of synthetically generated immunobiological datasets. We deploy
supervised and unsupervised generative methods, such as conditional
variational autoencoders (cVAEs) and Gaussian Mixture Models, to
generate virtual patients across multiple medical and immunobiological
datasets and develop an algorithm to assess synthetic data fidelity.
To determine ‘physiological resemblance’ between synthetic and real
data we implement a series of tests, including a
discriminator-as-evaluator layer in the algorithm. In this approach,
random forest (RF) classifiers are trained to distinguish real from
synthetic data, with classifier performance serving as a proxy for
generative success. When the classifier performs near chance, the
synthetic dataset is effectively indistinguishable from the real data.
Beyond straightforward validation, our framework provides diagnostic
feedback by identifying specific physiological deviations, such as
shifts in statistical properties or multidimensional
interdependencies, thereby allowing for the iterative refinement
towards high-fidelity virtual patients and ensuring transparency in
synthetic data applications.
[1] C.S. Korosec et al., Patterns, vol. 7, no. 3, Mar. 2026.
