q-bio.NCApr 3, 2026
SaveBridging scalp and intracranial EEG in BCI via pretrained neural representations and geometric constraint embedding
Organizations: Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China. · University of Chinese Academy of Sciences, Beijing, China.
Abstract
Electroencephalography (EEG) has become one of the key modalities underpinning brain-computer interfaces (BCIs) due to its high temporal resolution, rapid responsiveness, non-invasiveness, low cost, and portability. However, EEG signals are substantially inferior to intracranial EEG (iEEG) in signal-to-noise ratio and local spatial resolution, whereas iEEG suffers from extremely limited clinical accessibility owing to its invasive nature, hindering widespread application. To address this challenge, this study proposes a unified data-and prior knowledge-driven framework for EEG-iEEG representational enhancement. Guided by the principle that "geometric structure dictates function", the framework maps static cortical anatomy onto dynamic constraints governing neural signal propagation and integrates general-purpose neural representations extracted by a pre-trained large EEG model to explicitly model signal transmission through the brain. Enhanced EEG signals are then synthesized via a multidimensional representation diffusion process. Numerous experimental results demonstrate that the generated enhanced EEG signals effectively recover the neural activity patterns lost during propagation through the brain. This finding indicates that the performance ceiling of BCIs is constrained not only by acquisition hardware but also by the depth to which the generative model resolves the mechanisms of neural signal propagation. Collectively, the proposed framework provides a viable pathway toward acquiring high-fidelity neural signals at low cost.
Figures & tables
Figure 1: Schematic of the proposed framework. a, Data- and knowledge-based unified EEG representation enhancement framework. The model is trained with scalp EEG as input and iEEG as the target; during inference, enhanced EEG signals are generated from scalp EEG alone. b, Detailed architecture of the enhancement model, integrating a data-driven module, a neurodynamical–geometric constraint module, and a generative AI–based representation module. Individualized T1-weighted MRI can be incorporated alongside EEG to further improve performance. c, Multidimensional visualization of enhanced EEG signals, including spectral power distribution, spatial maps of band-specific activity, and temporal waveform morphology, collectively demonstrating recovery of neurophysiologically relevant features. d, Applications of the proposed framework, encompassing improved BCI decoding accuracy, enhanced BCI fairness , and support for brain functional mechanism research.
Figure 2: Multidimensional validation of consistency between synthesized iEEG and real iEEG. a, Across six complementary metrics \CJK@punctchar \CJK@uniPunct 0”80”94spectral (PSD Sim, Freq Sim), temporal (PCC, Cos Sim), and time–frequency coupling (Phase Consistency, CorrMat Sim) \CJK@punctchar \CJK@uniPunct 0”80”94synthesized iEEG generated from EEG (blue) significantly outperforms noise-driven controls (yellow), demonstrating effective recovery of high-dimensional neural dynamics lost to skull attenuation and volume conduction. b, t-SNE embeddings in the time–frequency domain show tight clustering of synthesized iEEG with real iEEG (red), while noise-driven signals (yellow) are distinctly separated, confirming acquisition of a physiologically plausible joint time–frequency representation. c, Heatmaps of full-channel correlation matrices reveal that synthesized iEEG recapitulates the local clustered and long-range synchronized connectivity patterns observed in real iEEG within five key functional networks, whereas noise-driven reconstructions lack such neurophysiologically specific topological structure.
Figure 3: Performance gains and neurophysiological validity of enhanced EEG across diverse BCI tasks. a, Across 12 EEG datasets spanning motor imagery emotion and cognitive decoding enhanced EEG boosts classification F1 scores by 2.8% to 8.6% (mean 5.2%) indicating recovery of high-frequency neural dynamics lost to tissue filtering. b, On SEED topographic maps show enhanced EEG yields focal high-intensity activation in dorsolateral prefrontal and bilateral temporal regions consistent with emotion circuitry whereas raw EEG shows diffuse weak activity. c, SHAP analysis reveals top channels in enhanced EEG such as Fp1 and Fz cluster in prefrontal limbic areas and their importance ranking correlates more strongly with emotion-related localization rising from 0.41 to 0.62.
Figure 4: Impact of T1-MRI–derived geometric constraints on iEEG reconstruction fidelity. a, Eight geometric modal features extracted from T1 MRI show coarse large-scale patterns at low N (2–5) and progressively finer gyral-level detail as N increases to 20–100. b, Consistency between synthesized and real iEEG peaks at N = 20 (Freq Sim 0.50 PSD Sim 0.51 PCC 0.38 Cos Sim 0.57) then declines with further increases in N revealing an optimal precision threshold for geometric constraint modeling. c, Correlation matrix heatmaps illustrate feature redundancy at N 10 strong diagonal dominance with moderate off-diagonal structure at N = 20–40 indicating balanced independence and physiological coherence and increasing noise and loss of structure at N 50 consistent with performance trends in b.
Figure 5: Ablation study of framework components. a, The full model (PERE + NC + GC) achieves highest fidelity across all metrics; removing PERE causes the largest drop while GC and NC provide complementary gains in spectral and phase consistency. b, Only the full model yields focal activation in emotion-relevant regions on SEED matching known neurophysiological patterns; ablated variants show diffuse activity. c, Downstream accuracy on MULTI-CLARID and SEED rises with component integration reaching 0.48 and 0.60 which are 84% and 39% above baseline confirming synergistic enhancement from data-driven representation neurodynamics and geometry.
Figure 6: A three-stage paradigm for EEG enhancement. Stage 1 builds a neural representation foundation model from multimodal data and neuroscientific priors to capture universal propagation patterns. Stage 2 adapts this model to specific BCI scenarios such as emotion or motor imagery through fine-tuning. Stage 3 personalizes the output using subject-specific calibration for high-fidelity signals; with limited data zero-shot learning provides partial adaptation.
Extended Data Fig. 1: Spectral specificity of EEG enhancement across neural oscillation bands. a, Reconstruction fidelity increases with frequency; gamma achieves highest performance with PCC 0.47 and PSD Sim 0.49 while delta and theta are lowest with PCC 0.17 and PSD Sim 0.18. b, Beta and gamma enhanced signals replicate focal iEEG-like patterns whereas delta and theta remain diffuse lacking structural specificity. c, t-SNE shows distinct band-specific clusters with adjacent bands spatially proximate; full-band embeds near beta and gamma and sigma from 13 to 15 hertz lies at the alpha–beta boundary consistent with its transitional role.
Extended Data Fig. 2: Topographic maps of EEG signal enhancement across model variants, illustrating the spatial specificity of neural activation patterns.Top rows show results on the SEED dataset; bottom rows show results on SEED-V. Columns correspond to three ablation settings: without the Pretrained EEG Representation Encoder (PERE), without the Neural Dynamics Constraint (NC), without the Geometric Constraint (GC), and the full model. Warm colors from red to orange indicate regions of high activation, while cool colors from blue to cyan denote low or baseline-level activity. As model components are incrementally integrated from ablated variants to the full framework, activation patterns become increasingly focal and anatomically coherent, converging on emotion-relevant cortical regions such as the prefrontal cortex, anterior cingulate, and temporal poles, while non-specific activations in sensorimotor and occipital areas are markedly suppressed. These findings demonstrate that the proposed framework enhances both the neurophysiological fidelity and functional specificity of reconstructed EEG signals.
Extended Data Fig. 3: Reconstruction performance of iEEG signals on a T1 MRI–free EEG–iEEG dataset (OpenNeuro ds004752), demonstrating the generalizability and robustness of the proposed framework. In the absence of subject-specific T1 MRI, the framework employs a standardized T1 MRI template as the anatomical prior to condition the iEEG reconstruction. Results are presented for 15 participants from dataset ds004752 (OpenNeuro, https://openneuro.org/datasets/ds004752/versions/1.0.1). Quantitative evaluation is conducted using four complementary metrics: Pearson correlation coefficient (PCC), cosine similarity (Cos Sim), power spectral density similarity (PSD Sim), and frequency-domain similarity (Freq Sim). The consistent high performance across these metrics confirms that the framework maintains efficacy and functional fidelity even when subject-specific structural imaging is unavailable, thereby validating its applicability in real-world clinical and research settings where high-resolution MRI may be inaccessible.
Extended Data Fig. 4: Comparative analysis of frequency distributions for original and enhanced EEG signals across representative subjects. The spectrograms demonstrate that the enhanced EEG signals exhibit a marked increase in the proportion of high-frequency bands (yellow-green hues) and a corresponding decrease in low-frequency bands (blue-purple hues) compared to the original signals. This spectral shift indicates a significant improvement in the proportion of high-frequency components, suggesting that the enhancement framework successfully and effectively recovers high-frequency information typically lost during the propagation of neural signals through brain tissue. Quantitative evaluation of the enhancement effects across different subjects reveals a widespread distribution of efficacy, with the proportion of increase in the band following the order: Sub-15 (7%) Sub-07 (6%) Sub-04 (5%) Sub-12 (3%). This hierarchical trend highlights the robustness of the method in recovering high-frequency neural dynamics across varying individual baselines.
Extended Data Fig. 5: Scatter points represent geometric modal feature samples for mode groups (mode-2 to mode-100), color-coded with a purple blue green yellow gradient reflecting the progressive diffusion stage of geometric feature decomposition. The dotted black line represents the principal regression trajectory of the Laplace–Beltrami Operator, serving as a reference axis for quantifying geometric structure deviation, with R values measuring the goodness-of-fit between each mode group and this regression line. Results show: mode-2 to mode-10 (purple-blue) exhibit highly compact clustering along the diagonal band with strong geometric constraint but limited discriminability; mode-60 to mode-100 (green-yellow) show excessive diffusion beyond the band structure, obscuring geometric-topological relationships; mode-20 to mode-50 (blue-green) maintain band-aligned structure with sufficient inter-sample discriminability while remaining tightly constrained around the regression line, preserving intrinsic cortical manifold geometry. These findings indicate that mode-20 to mode-50 achieve optimal balance between geometric constraint and feature expressiveness, representing the most suitable geometric feature decomposition modes for EEG enhancement tasks.
Extended Data Fig. 6: Scatter points represent geometric modal feature samples for mode groups (mode-2 to mode-100), color-coded with a purple blue green yellow gradient reflecting the progressive diffusion stage of geometric feature decomposition. The dotted black line represents the principal regression trajectory of the Laplace–Beltrami Operator, serving as a reference axis for quantifying geometric structure deviation. Results show: mode-2 to mode-10 (purple-blue) exhibit highly compact clustering along the diagonal band with strong geometric constraint but limited discriminability; mode-60 to mode-100 (green-yellow) show excessive diffusion beyond the band structure, obscuring geometric-topological relationships; mode-20 to mode-50 (blue-green) maintain band-aligned structure with sufficient inter-sample discriminability while remaining tightly constrained around the regression line, preserving intrinsic cortical manifold geometry. These findings indicate that mode-20 to mode-50 achieve optimal balance between geometric constraint and feature expressiveness, representing the most suitable geometric feature decomposition modes for EEG enhancement tasks.
More in this topic
Brain-Computer Interfaces
Brain-computer interfaces translate neural activity into commands or feedback for communication, movement, rehabilitation, and sensory restoration.
See all 128 papers