Represent, Then Generate: Multimodal-Conditioned Time-Series Generation under Irregular Missingness
ReCoGen introduces a two-stage framework that decouples multimodal condition representation from target generation for time-series synthesis under irregular missingness Stage I uses per-modality masked autoencoders to distill time-variant conditions into compact, missingness-tolerant token sequences Stage II employs a flow-matching generator that fuses encoded tokens with static clinical covariates to synthesize target physiological signals ReCoGen achieves best-in-class downstream utility acros
Analysis
TL;DR
- ReCoGen introduces a two-stage framework that decouples multimodal condition representation from target generation for time-series synthesis under irregular missingness
- Stage I uses per-modality masked autoencoders to distill time-variant conditions into compact, missingness-tolerant token sequences
- Stage II employs a flow-matching generator that fuses encoded tokens with static clinical covariates to synthesize target physiological signals
- ReCoGen achieves best-in-class downstream utility across 16 (dataset, task, metric) settings on three physiological benchmarks (AI-READI, MIMIC-III, MIMIC-IV), outperforming six existing conditional generators
- On 13 of 16 settings, generated signal utility reaches or exceeds that of real signals, suggesting near-lossless surrogate generation
Why It Matters
This work directly addresses a critical bottleneck in clinical AI: the frequent absence of invasive or expensive physiological signals in real-world patient monitoring. By demonstrating that routinely collected data can serve as high-fidelity surrogates for unavailable signals, ReCoGen enables more accessible, lower-cost continuous monitoring pipelines and reduces patient burden from invasive procedures.
Technical Details
- Two-stage architecture: Stage I trains one masked autoencoder per conditioning modality, producing missingness-robust token representations; Stage II trains a flow-matching generator conditioned on these tokens plus static covariates
- Conditioning pathway innovations: Learnable cross-attention over frozen per-modality encoders, combined with a dual token-plus-AdaLN (adaptive layer normalization) route for static conditions
- Benchmarks: Continuous glucose monitoring on AI-READI; arterial blood pressure generation on MIMIC-III and MIMIC-IV
- Evaluation: Downstream utility measured across 16 (dataset, task, metric) combinations, comparing against six representative conditional generators
- Key finding: Generated signal utility matching or exceeding real signal utility on 13/16 settings, indicating the synthetic signals preserve clinically actionable information
Industry Insight
- The decoupled represent-then-generate paradigm offers a generalizable blueprint for handling heterogeneous, irregularly missing multimodal data beyond clinical time series, applicable to any domain with sparse sensor coverage
- The finding that synthetic signals can match real-signal utility challenges the assumption that generated data is inherently inferior, opening doors for data augmentation and privacy-preserving clinical data sharing
- Hospitals and health-tech companies should prioritize building multimodal conditioning pipelines rather than single-modality generators, as the latter degrade sharply under the irregular missingness patterns typical of real-world clinical data
Disclaimer: The above content is generated by AI and is for reference only.