Functional data—random functions observed over a continuous domain—are common in biomedical research but difficult to analyze due to privacy constraints, sparse sampling, and their infinite-dimensional nature. Existing deep generative methods often rely on regular grids or restrictive assumptions that do not hold for real-world clinical records.
The authors introduce Smooth Flow Matching (SFM), a copula-based semiparametric framework that learns a smooth flow to generate functional data. SFM avoids Gaussianity and low-rank assumptions, works with irregular observations, and ensures the generated functions are smooth. In simulations, it showed advantages in both synthetic data quality and computational efficiency.
On the MIMIC-IV electronic health records database, SFM generated clinical trajectory data that proved useful for downstream tasks, highlighting its potential to increase the utility of EHR data without exposing sensitive patient information. The paper is accepted for publication in The Annals of Statistics; this digest is based on that single source.