MedFlow: Class-Aware Multi-Scale Generation for Medical Time-Series Synthesis

MedFlow is a perceptual multi-scale flow matching framework for medical time-series synthesis. This study uses a vector quantization multi-scale tokenizer to represent medical sequences with complementary time resolutions, capturing both coarse-grained clinical trends and fine-grained dynamics. The study introduces the Token Marginal Guidance mechanism, which directly integrates conditional tokenization statistics into the flow matching process, guiding the generation towards learned specific category regions, thereby strengthening minority class patterns while preserving the overall and tail distributions of real data. Experiments on four publicly available datasets involving electronic health records, EEG, and ECG signals show that MedFlow outperforms recent diffusion-based baseline methods in downstream prediction tasks, with an average improvement of 5.8% in AUPRC and a reduction of 88.6% in Context-FID, and a 3.8-fold increase in sampling throughput.