ReCoGen Synthesizes Missing Physiological Signals from Multimodal Data

In clinical monitoring, continuous physiological time series are vital, but many informative signals are invasive, expensive, or simply unavailable. Conditional generation offers a solution by synthesizing absent signals from co-recorded data and routine clinical variables. However, current generators struggle with the diverse, irregularly missing mix of time-variant signals and static covariates common in real-world practice. Researchers propose ReCoGen (Represent Conditions, then Generate), a two-stage framework designed to overcome these limitations. The first stage involves training a masked autoencoder for each modality, distilling time-variant conditions into compact, missingness-tolerant token sequences. The second stage then trains a flow-matching generator that combines these tokens with static conditions to synthesize the desired target signal. ReCoGen was validated across three physiological benchmarks, including continuous glucose monitoring and arterial blood pressure generation. It achieved superior downstream utility in all sixteen tested settings, often matching or exceeding the utility of real signals. The gains were attributed to its sophisticated conditioning path, which includes learnable cross-attention over frozen per-modality encoders and a dual token-plus-AdaLN route for static conditions. This innovation allows routinely collected signals to serve as informative surrogates for invasive or unavailable ones, paving the way for less invasive and lower-cost continuous clinical monitoring.