Synthetic medical time‑series can alleviate data scarcity and provide reliable training data for clinical prediction models. Existing approaches mainly aim at matching the overall distribution and temporal dynamics of real data, which does not guarantee strong downstream performance on imbalanced medical datasets. Clinically informative patterns often appear at heterogeneous temporal scales, and minority‑class characteristics can be masked by dominant majority patterns.
To tackle these issues, we introduce MedFlow, a class‑aware multi‑scale flow‑matching framework. MedFlow employs a vector‑quantized multi‑scale tokenizer that encodes medical sequences at complementary temporal resolutions, capturing coarse clinical trends while preserving fine‑grained dynamics.
On top of this, we propose Token Marginal Guidance, which injects class‑conditional token statistics directly into the flow‑matching process, steering generation toward class‑specific regions of the learned token space. This mechanism strengthens minority‑class patterns while preserving both the global and tail distributions of the real data.
Experiments on four public datasets covering electronic health records, EEG, and ECG demonstrate that MedFlow consistently outperforms recent diffusion‑based baselines on downstream prediction tasks. It improves AUPRC by $5.8\%$, reduces Context‑FID by $88.6\%$, and achieves a $3.8\times$ increase in sampling throughput.
Review: MedFlow’s combination of multi‑scale token representation and class‑aware flow matching effectively addresses scale heterogeneity and class imbalance in medical time‑series synthesis, offering a powerful and efficient solution for generating data with strong downstream utility.