MEDiT: A mask-enhanced diffusion transformer model for fetal heart rate signal generation

Fetal Heart Rate (FHR) signals are widely used in fetal monitoring and maternal-fetal health assessment. However, their acquisition is constrained by ethical limitations, motion artifacts, and signal loss, resulting in limited data availability for intelligent clinical analysis. To address these, this paper proposes a deep generative Data Augmentation (DA) framework for FHR synthesis. Specifically, we propose a Mask-Enhanced Diffusion Transformer model (MEDiT) to model various missing data scenarios and generate high-fidelity FHR signals in data-scarce clinical settings. First, a mask-enhanced reinforcement mechanism is designed to randomly mask signal segments at varying positions and lengths, forcing the model to reconstruct targeted regions from unmasked regions containing clinically critical information and thereby learn intrinsic physiological patterns under partial observability. Next, a Diffusion Transformer (DiT)-based generative architecture with a dual-branch learning strategy is developed to adaptively adjust the optimization objective under different masking ratios, improving both global temporal coherence and local waveform fidelity. Owing to self-attention-based global interaction and the threshold-guided branch design, MEDiT is able to capture long-range temporal dependencies while maintaining stable. Comprehensive experiments on a publicly available CTG dataset, including parameter sensitivity analysis, robustness evaluation, comparisons with state-of-the-art methods, downstream classification tasks, and clinician assessment, demonstrate the effectiveness of MEDiT. The results show that diffusion-based DA approaches perform better in both generation quality and training stability, and MEDiT further improves classification accuracy by 14.0% over the best baseline method (DDPM).