Bridging the state gap: Multifractal Conditional Diffusion with Latent Feature Hallucination for physiologically consistent EEG

Motor imagery electroencephalography (MI-EEG) is a core modality for studying noninvasive brain–computer interfaces (BCIs), but its practical application is limited by the scarcity of high-quality data. While diffusion probability models (DPMs) have shown potential in data synthesis, they often struggle to maintain physiological consistency under complex neurodynamics. In response to these limitations, we propose MFH-Diff, a novel Multifractal Conditional Diffusion framework integrated with Latent Feature Hallucination, designed to achieve physiologically consistent EEG synthesis. MFH-Diff captures the complex cross-scale dynamics of EEG signals through a multifractal framework, guiding the diffusion process to achieve coarse-to-fine signal reconstruction. Furthermore, to mitigate coverage bias caused by intra-subject nonstationarity, we introduce a Latent Feature Hallucination (LFH) module to fill non-overlapping state gaps in the feature space. Extensive experiments on the BCIC-IV-2a and BCIC-IV-2b datasets demonstrate that MFH-Diff significantly outperforms existing augmentation methods, significantly improving accuracy and kappa on both datasets. Qualitative analysis also reveals that the generated signals retain neurophysiological patterns in the time, frequency, and feature embedding domains without introducing false artifacts. These results prove that MFH-Diff offers a physiologically consistent and interpretable EEG data augmentation method, effectively bridging the state gap and significantly enhancing the robustness and decoding accuracy of BCI systems in complex motor imagery tasks.