Geometry- and Relation-Aware Diffusion for EEG Super-Resolution

TL;DR

This work introduces TopoDiff, a diffusion-based framework for EEG spatial super-resolution that explicitly enforces geometry and cross-channel relations. By fusing topology-aware topographic embeddings with a dynamic graph of inter-electrode connectivity, the model generates high-density EEG signals that are spatially faithful and temporally coherent. Across SEED, SEED-IV, PhysioNet MI/MM, and TUSZ, TopoDiff yields consistent gains in reconstruction fidelity (NMSE, SNR, PCC) and downstream task performance, especially under aggressive 8x upsampling. The approach demonstrates how global scalp geometry and time-varying electrode relationships can be leveraged within a conditional diffusion model to improve wearable EEG performance and cross-subject generalization. Limitations include potential gains from subject-specific head models and more sophisticated connectivity priors, with future work aimed at uncertainty-aware and cross-montage extension.

Abstract

Recent electroencephalography (EEG) spatial super-resolution (SR) methods, while showing improved quality by either directly predicting missing signals from visible channels or adapting latent diffusion-based generative modeling to temporal data, often lack awareness of physiological spatial structure, thereby constraining spatial generation performance. To address this issue, we introduce TopoDiff, a geometry- and relation-aware diffusion model for EEG spatial super-resolution. Inspired by how human experts interpret spatial EEG patterns, TopoDiff incorporates topology-aware image embeddings derived from EEG topographic representations to provide global geometric context for spatial generation, together with a dynamic channel-relation graph that encodes inter-electrode relationships and evolves with temporal dynamics. This design yields a spatially grounded EEG spatial super-resolution framework with consistent performance improvements. Across multiple EEG datasets spanning diverse applications, including SEED/SEED-IV for emotion recognition, PhysioNet motor imagery (MI/MM), and TUSZ for seizure detection, our method achieves substantial gains in generation fidelity and leads to notable improvements in downstream EEG task performance.

Figures (9)

• Figure 1: Comparison between previous pipelines and the proposed framework. (a) Prior methods only implicitly encode spatial geometry through coordinates and positional encoding. (b) TopoDiff enables geometry- and relation-aware generation by explicitly incorporating spatial and relational modeling.
• Figure 2: Overview of the TopoDiff framework. (a) Global topological conditioning, where EEG topographic representations are used to provide global scalp geometry and spatial context. (b) Dynamic graph-based inter-channel relation modeling, in which graphs are constructed in a time-varying manner and a GNN is used to enrich channel representations with explicit relational information. (c) The diffusion Transformer architecture employed for EEG spatial super-resolution.
• Figure 3: Topographical visualization of Ground Truth, ESTformer, and TopoDiff super-resolution outputs for the TUSZ 2$\times$ setting at 3 different time steps.
• Figure 4: Convergence behavior with topological conditioning on the SEED dataset under $2\times$, $4\times$, and $8\times$ super-resolution settings.
• Figure 5: Visualization of EEG channels in full, 2 $\times$ SR, 4 $\times$ SR, and 8 $\times$ SR configurations.