Cross-region neural signal reconstruction to lift electrode placement constraints in SSVEP brain-computer interfaces Authors: Wenqiang Yan, Qi Luo, Chenghang Du, Hui Li, Shengchao Chen, Ziwen Yuan, Xinyu Hou, Haochong Wang Published: 2026 ABSTRACT Steady-state visual evoked potential (SSVEP)-based brain–computer interfaces (BCIs) rely on occipital EEG recordings, which are infeasible in many clinical scenarios (e.g., supine positioning, traumatic brain injury), restricting SSVEP-BCI access for patients with urgent communication needs. We propose dynamic sliding point-wise reconstruction with a triple-band cross-fusion network (DSTF-Net), a cross-brain-region framework enabling accurate SSVEP decoding exclusively from frontal EEG signals. DSTF-Net integrates four core innovations: ①a dynamic sliding point-wise reconstruction strategy that abandons static n-to-m mapping (directly mapping n frontal EEG samples to m occipital samples) to capture fine-grained temporal dependencies; ②a triple-band cross-fusion sub-network processing frontal EEG frequency bands through parallel convolutional streams and cyclic cross-attention; ③a stage-wise hierarchical training mechanism mitigating gradient interference by sequentially training single-band streams before fusion optimization; and ④a neurophysiologically constrained loss function enforcing validated unidirectional occipital-to-frontal SSVEP propagation. We trained DSTF-Net using paired frontal-occipital EEG from a healthy participant, then transferred it to 20 new users (including 8 brain-injured patients maintaining a supine position) via cross-subject transfer, reconstructing occipital activity solely from frontal EEG. Our framework achieves a maximum 33.47% decoding accuracy improvement over baselines. By eliminating occipital electrode requirements, our work expands SSVEP-BCI accessibility for clinically constrained populations and establishes a generalizable cross-brain-region neural mapping framework. Introduction Brain–computer Interface (BCI) is defined as a direct interface enabling bidirectional information exchange between biological neural systems (including the brain and spinal cord) and external computational devices1,2. This advanced technology has found significant applications in multiple fields, such as electroencephalography (EEG)-based speech decoding3–6, neural prosthetic control systems7,8, neurorehabilitation engineering9–11, and spinal cord stimulation therapy12,13. BCI systems utilize predefined encoding paradigms to establish the mapping relationship between neural intent and external device control14–22. Among these, the steady-state visual evoked potential (SSVEP) paradigm activates the occipital lobe through visual stimulation at specific frequencies, generating synchronous EEG activity that matches the stimulation frequency23–28. This represents a primary application paradigm for realizing external device control in BCI technology1,26,29–31. The SSVEP represents synchronous neural oscillations generated by occipital visual cortex neurons in response to fixed-frequency visual stimulation, which generally requires the placement of EEG recording electrodes over the occipital region32–38. However, real-world clinical scenarios present significant implementation challenges: neurological patients often require long-term supine positioning due to medical conditions or intensive care requirements39,40; traumatic brain injury and intracranial pathologies may result in occipital bone defects41,42; and postoperative cranial fixation devices frequently interfere with electrode placement. These clinical constraints effectively preclude occipital EEG acquisition, thereby eliminating BCI accessibility for populations with urgent communication needs. To the best of our knowledge, no existing studies have addressed this critical technical barrier despite its urgent unmet clinical need. When the occipital visual cortex is unavailable for neural signal acquisition, the frontal region—characterized by its hairless surface and sufficient contact area—represents an optimal alternative site for EEG recording electrode placement. However, the human skull acts as an inhomogeneous conductor43,44, inducing significant signal attenuation and dispersion effects on SSVEP-generated neural oscillations propagating from the occipital lobe to other brain regions45–48. Specifically, the amplitude of SSVEP-related neural information decays exponentially with increasing distance from the occipital cortex, rendering frontal EEG signals insufficient for reliable SSVEP decoding. As detailed in technical pathway ① of Fig. 1, we introduce a novel approach combining synchronous placement of EEG electrodes over the occipital and frontal region in a healthy subject with the development of advanced deep learning architectures to establish a cross-brain-region mapping between frontal and occipital neural oscillations. This framework enables non-invasive reconstruction of occipital SSVEP activity from frontal EEG recordings. Technical pathway ② involves transferring the cross-brain-region reconstruction model—trained on healthy subject datasets—to patient data via cross-subject and cross-device transfer, allowing exclusive acquisition of patients’ frontal EEG to computationally reconstruct occipital neural activity and decode SSVEP information using the reconstructed occipital EEG.Fig. 1Breaking the constraints of electrode placement in specific brain regions: a cross-brain-region neural signal reconstruction approach.Schematic illustration of a novel framework enabling cross-brain-region reconstruction of occipital lobe neural activity from frontal EEG signals when direct occipital electrode placement is infeasible for patients with brain injury. Technical Pathway ①: Simultaneous placement of EEG electrodes on the occipital and frontal regions of a healthy subject to establish a mapping model between frontal and occipital EEG signals. Technical Pathway ②: Cross-subject and cross-device transfer of the established cross-brain-region neural signal reconstruction model from a healthy subject to patient data. Breaking the constraints of electrode placement in specific brain regions: a cross-brain-region neural signal reconstruction approach. Schematic illustration of a novel framework enabling cross-brain-region reconstruction of occipital lobe neural activity from frontal EEG signals when direct occipital electrode placement is infeasible for patients with brain injury. Technical Pathway ①: Simultaneous placement of EEG electrodes on the occipital and frontal regions of a healthy subject to establish a mapping model between frontal and occipital EEG signals. Technical Pathway ②: Cross-subject and cross-device transfer of the established cross-brain-region neural signal reconstruction model from a healthy subject to patient data. To realize reliable cross-brain-region SSVEP reconstruction and overcome the limitations of direct frontal EEG decoding, we propose an innovative cross-brain-region reconstruction framework, Dynamic sliding point-wise reconstruction with triple-band cross-fusion network (DSTF-Net), that departs from conventional end-to-end m-to-n mapping paradigms (which directly reconstruct n occipital EEG samples from m frontal counterparts). The core innovation lies in a time-window sliding point-wise reconstruction strategy, where temporal windows are sequentially slid across the frontal EEG time series to predict occipital neural activity at each individual time point. This design is specifically tailored to tackle the inherent high time-variability of EEG signals, enabling capture of fine-grained, dynamic temporal dependencies between frontal and occipital regions—dependencies that static reconstruction methods typically overlook and that are critical for retrieving weak SSVEP signals attenuated by skull and scalp tissues. Underpinning this reconstruction strategy is a custom triple-band cross-fusion convolutional neural network (CNN) tailored to the spatiotemporal and spectral properties of EEG signals, integrated with four additional complementary innovations to ensure high-fidelity reconstruction. First, the CNN architecture features three parallel convolutional streams, each dedicated to processing low-, mid-, or high-frequency bands of frontal EEG. Each stream integrates temporal convolution blocks (for time-domain dynamics), spatial convolution blocks (for inter-electrode spatial patterns), and depthwise separable convolutions (for efficient feature refinement). Second, a novel triple-band cross-fusion module integrates these frequency-specific features via cyclic cross-attention (mid-frequency features modulate low- and high-frequency weights via sigmoid gating), depth concatenation, and residual connections—ensuring robust inter-frequency interaction and feature integrity. Third, a stage-wise hierarchical training mechanism mitigates gradient interference by sequentially training the three single-band streams before optimizing the fusion module, guaranteeing stable convergence and reliable feature extraction. Fourth, a neurophysiologically constrained loss function incorporates Granger causality-based temporal constraints to enforce the biological plausibility of reconstruction results. The framework is trained on paired frontal-occipital EEG data from healthy participants, then transferred to 20 new users (including 8 brain-injured patients who could only maintain a supine position) via cross-subject and cross-device transfer—enabling occipital SSVEP reconstruction using only the new users’ frontal EEG. Critical to validating the theoretical feasibility of this cross-brain-region mapping approach, causal analysis49–52 of simultaneously acquired frontal and occipital EEG data from a healthy subject revealed that SSVEP frequency components in frontal signals are not generated locally but are primarily driven by SSVEP signals originating in the occipital lobe. Specifically, our analysis confirmed a unidirectional signal-level directional association of SSVEP features propagating from the occipital to the frontal lobe, with no detectable reverse directional dependency. This signal-level pattern supports the biological plausibility of our cross-region mapping framework and provides a theoretical basis for reconstructing occipital-like SSVEP activity from frontal EEG recordings. As demonstrated later, this approach achieves a maximum 33.32% improvement in SSVEP decoding accuracy over baselines in the transferred new user cohort, particularly overcoming the clinical barrier of occipital electrode inaccessibility for the 8 supine brain-injured patients. Beyond BCI applications, this work establishes a generalizable paradigm for cross-brain-region neural signal mapping—supporting neural signal acquisition at easily accessible electrode sites (e.g., frontal lobe). This capability lays a preliminary foundation for the development of wearable EEG devices by addressing key barriers to BCI miniaturization and portability, holding broad implications for neuroengineering, clinical neuroscience, and the translation of neural interface technologies to real-world clinical settings. Results Frontal EEG demonstrates inadequate capacity for decoding SSVEP information When the occipital lobe is unavailable for EEG electrode placement, the frontal region—characterized by its hairless surface and sufficient contact area—represents an optimal alternative site. However, whether SSVEP information can be decoded from frontal EEG recordings requires experimental validation. Simultaneous SSVEP signals were recorded from occipital and frontal regions in a healthy subject. Figure 2A displays power spectra of occipital POz, PO4, PO6, and Oz channels, revealing prominent amplitude peaks at the fundamental frequency (9 Hz) and its harmonics. In contrast, Fig. 2B shows frontal lobe power spectra with no discernible peaks at the stimulation frequency. Figure 2C presents canonical correlation analysis (CCA) results: occipital EEG demonstrated significant feature coefficients at the fundamental and harmonic frequencies for 8–11 Hz stimulations, whereas frontal EEG lacked frequency-specific responses. Figure 2D compares decoding accuracies using CCA for different sample lengths: occipital EEG outperformed frontal EEG across all lengths (**p < 0.01, ***p < 0.001). Figure 2E illustrates decoding trends: occipital accuracy stabilized at 100% beyond 1300 samples, while frontal accuracy remained below 45% without monotonic improvement. Figure 2F aggregates results across all sample lengths, confirming near-perfect occipital decoding versus <40% frontal accuracy. Filter bank canonical correlation analysis (FBCCA) validation in Fig. 2G–I replicates these findings: occipital decoding accuracy remained significantly higher across all sample lengths (***p < 0.001), with occipital signals achieving near-complete accuracy and frontal signals failing to exceed 50%. Collectively, these results confirm that frontal EEG lacks sufficient SSVEP information for reliable decoding.Fig. 2Comparison of decoding performance between frontal and occipital EEG.A Power spectra of SSVEP signals recorded from occipital electrodes POz, PO4, PO6, and Oz at a stimulation frequency of f = 9 Hz. Blue shaded regions highlight power spectral amplitudes at the fundamental, second-, and third-harmonic frequencies of the stimulation. B Power spectra of frontal EEG signals from four electrode channels under f = 9 Hz stimulation. Pink shaded regions indicate power spectral amplitudes at the fundamental, second-, and third-harmonic frequencies of the stimulation. C Canonical correlation coefficient spectra of frontal and occipital EEG during SSVEP stimulation at 8, 9, 10, and 11 Hz. The x-axis represents frequency (Hz), and the y-axis shows the canonical correlation coefficient calculated against frequency-matched reference signals. Dotted-dashed lines denote characteristic coefficients at the fundamental, second-, third-, and fourth-harmonic frequencies of the stimulation. D Decoding accuracies of frontal and occipital EEG using canonical correlation analysis (CCA) across different sample lengths. Accuracy values represent the mean ± s.e.m. of 13 blocks (each containing four trials) per sample length. One-way ANOVA revealed significant differences (**p < 0.01) and ***p < 0.001). E Decoding accuracy (CCA) as a function of sample length for occipital and frontal EEG. F Mean decoding accuracy across all sample lengths shown in (E). G Decoding accuracies of frontal and occipital EEG using filter bank canonical correlation analysis (FBCCA) across different sample lengths. Accuracy values represent the mean ± s.e.m. of 13 blocks (each containing four trials) per sample length. H Decoding accuracy (FBCCA) as a function of sample length for occipital and frontal EEG. I Mean decoding accuracy across all sample lengths shown in (H). Comparison of decoding performance between frontal and occipital EEG. A Power spectra of SSVEP signals recorded from occipital electrodes POz, PO4, PO6, and Oz at a stimulation frequency of f = 9 Hz. Blue shaded regions highlight power spectral amplitudes at the fundamental, second-, and third-harmonic frequencies of the stimulation. B Power spectra of frontal EEG signals from four electrode channels under f = 9 Hz stimulation. Pink shaded regions indicate power spectral amplitudes at the fundamental, second-, and third-harmonic frequencies of the stimulation. C Canonical correlation coefficient spectra of frontal and occipital EEG during SSVEP stimulation at 8, 9, 10, and 11 Hz. The x-axis represents frequency (Hz), and the y-axis shows the canonical correlation coefficient calculated against frequency-matched reference signals. Dotted-dashed lines denote characteristic coefficients at the fundamental, second-, third-, and fourth-harmonic frequencies of the stimulation. D Decoding accuracies of frontal and occipital EEG using canonical correlation analysis (CCA) across different sample lengths. Accuracy values represent the mean ± s.e.m. of 13 blocks (each containing four trials) per sample length. One-way ANOVA revealed significant differences (**p < 0.01) and ***p < 0.001). E Decoding accuracy (CCA) as a function of sample length for occipital and frontal EEG. F Mean decoding accuracy across all sample lengths shown in (E). G Decoding accuracies of frontal and occipital EEG using filter bank canonical correlation analysis (FBCCA) across different sample lengths. Accuracy values represent the mean ± s.e.m. of 13 blocks (each containing four trials) per sample length. H Decoding accuracy (FBCCA) as a function of sample length for occipital and frontal EEG. I Mean decoding accuracy across all sample lengths shown in (H). Causal analysis supports frontal-to-occipital reconstruction of EEG signals To address the insufficient SSVEP decoding from frontal EEG, we propose a cross-brain-region reconstruction algorithm that maps frontal EEG to occipital neural activity for SSVEP decoding. Prior to clinical implementation, the theoretical feasibility of reconstructing occipital EEG from frontal recordings must first be established. We analyzed the SSVEP causality between the frontal and occipital lobes using EEG data obtained from a healthy subject. The first analysis focused on frequency-domain Granger causality computed from the averaged signals of the occipital region (channels PO3, POz, and PO4) and the frontal region (frontal electrode channels 1, 2, and 3) (Fig. 3A and B). In Fig. 3A, under four distinct stimulation frequencies, significant causality peaks (p < 0.01) were observed at either the stimulation frequency or its harmonics, whereas in Fig. 3B no such peaks were detected. These results indicate that under SSVEP stimulation, the unidirectional propagation of signals from the occipital to the frontal lobes exhibits significant frequency specificity. Notably, the significant causality peaks varied in their corresponding harmonic components across the different stimulation frequencies. For lower stimulation frequencies (8 and 9 Hz), significant peaks were observed at the second and third harmonics, whereas for higher stimulation frequencies (10 and 11 Hz), peaks emerged at the fundamental frequency and the second harmonic. These findings suggest that long-range inter-regional SSVEP transmission exhibits an optimal frequency range, in which harmonic components contribute substantially to cross-region signal coupling.Fig. 3SSVEP causality analysis between the occipital and frontal lobes.A Frequency-domain causality from the occipital to frontal lobes under four distinct stimulation frequencies. Error shading indicates the mean ± s.e.m. (n = 10 trials), and blue dashed lines indicate the SSVEP fundamental frequency (f) and its harmonics (2f and 3f). Statistical significance was assessed using the Wilcoxon rank-sum test (*p < 0.05, **p < 0.01). B Frequency-domain causality from the frontal to occipital lobes under the same stimulation conditions as in (A). C Pairwise-conditional time-domain causality computed between occipital and frontal channels. D and E Band-limited time-domain causality between occipital and frontal channels, computed by integrating the spectral causality over specified frequency ranges. In panels C–E, black-shaded regions indicate significant differences (p = 0.01), with statistical significance adjusted for multiple comparisons using the false discovery rate. SSVEP causality analysis between the occipital and frontal lobes. A Frequency-domain causality from the occipital to frontal lobes under four distinct stimulation frequencies. Error shading indicates the mean ± s.e.m. (n = 10 trials), and blue dashed lines indicate the SSVEP fundamental frequency (f) and its harmonics (2f and 3f). Statistical significance was assessed using the Wilcoxon rank-sum test (*p < 0.05, **p < 0.01). B Frequency-domain causality from the frontal to occipital lobes under the same stimulation conditions as in (A). C Pairwise-conditional time-domain causality computed between occipital and frontal channels. D and E Band-limited time-domain causality between occipital and frontal channels, computed by integrating the spectral causality over specified frequency ranges. In panels C–E, black-shaded regions indicate significant differences (p = 0.01), with statistical significance adjusted for multiple comparisons using the false discovery rate. The second analysis involved computing pairwise Granger causality between occipital and frontal channels using 10 Hz stimulation as a representative example. Examining the full-frequency domain Granger causality (equivalent to time-domain Granger causality) in Fig. 3C, we observed significant causal relationships (p < 0.01) among occipital channels, among most frontal channels, and from each occipital channel to frontal channels. Conversely, no causality was detected from frontal to occipital channels, which is consistent with the frequency-domain Granger causality results. Focusing on the fundamental frequency and its second harmonic, we integrated the frequency-domain Granger causality over a 2 Hz bandwidth to obtain band-limited pairwise Granger causality (Fig. 3D and E). Notably, the causality among frontal channels became non-significant, whereas the causality from occipital to frontal channels was enhanced. These observations demonstrate that the SSVEP frequency components recorded in the frontal lobe are not locally generated but are predominantly derived from occipital-originated SSVEP signals via cross-region transmission. The significant directional Granger causality observed across task-relevant frequency bands reflects consistent temporal dependency between the two brain regions induced by SSVEP stimulation, which provides a valid signal-level basis for reconstructing occipital SSVEP features from frontal EEG signals. Dynamic sliding point-wise reconstruction with triple-band cross-fusion network (DSTF-Net) decoded the temporal mapping of frontal-occipital EEG A frontal-occipital EEG mapping model was established using our proposed DSTF-Net, trained on simultaneous EEG data from a single healthy subject (30 trials, 5 frontal/8 occipital channels). Reconstruction efficacy was quantified by the correlation between reconstructed and ground-truth occipital time-domain waveforms. Averaging signals across respective channels yielded three 1-D signal sets (test frontal, reconstructed occipital, ground-truth occipital). Under 8 Hz stimulation (Fig. 4A), simultaneously acquired frontal and ground-truth occipital signals showed weak correlation (Pearson correlation coefficient r = −0.255, coefficient of determination R² ≈ 0). In contrast, DSTF-Net-reconstructed occipital signals correlated with ground truth (r = 0.4388, R² = 0.1925), with waveforms approximately distributed along the 45° reference line. Figure 4B shows correlation coefficients of the three signal sets across sample lengths [500:500:4800] (averaged over 12 trials per length). Reconstruction via DSTF-Net significantly enhanced frontal-occipital correlation across all sample lengths. These time-domain correlation results confirm that DSTF-Net effectively establishes the frontal-occipital EEG mapping.Fig. 4Effect analysis of DSTF-Net for cross-brain region reconstruction of occipital lobe signals via prefrontal lobe signals.A Results of correlation analysis of the waveforms between occipital lobe signals of a healthy subject and simultaneously acquired frontal lobe signals, as well as frontal lobe reconstructed signals. Among them, the red dash-dot line represents the 45° fitting line, the green solid line represents the first-order polynomial data fitting line, the r value denotes the Pearson correlation coefficient between the two groups of signals, and the R² value is the coefficient of determination for the two groups of signals. B Pearson correlation coefficients of occipital vs. frontal/reconstructed frontal signals in a healthy subject at sample lengths [500:500:4800] (averaged over 12 trials). C and D Canonical correlation coefficients between prefrontal EEG, prefrontal-reconstructed occipital EEG, and sine–cosine template signals with frequencies [6:0.5:25] Hz at sample length [500:50:4800], calculated by cross-subject and cross-device transfer of the mapping model established based on prefrontal and occipital EEG data of a healthy subject to new users’ EEG data; C shows the analysis results of prefrontal-reconstructed occipital EEG under ssvep stimulation frequencies of 8, 9, and 10 Hz, and D shows the analysis results of original prefrontal EEG. E Signal-to-noise ratios of prefrontal EEG and its reconstructed occipital EEG at a sample length of 4000, calculated using canonical correlation coefficients as the feature index, with 105 trials collected from new users. F Normalized MSE, signal smoothness, cross-spectral coherence, and individual difference compensation rate indices for evaluating the cross-brain region reconstruction performance of DSTF-Net, calculated from 15 blocks collected from new users (where the index for each block is the result of averaging over 7 trials). Effect analysis of DSTF-Net for cross-brain region reconstruction of occipital lobe signals via prefrontal lobe signals. A Results of correlation analysis of the waveforms between occipital lobe signals of a healthy subject and simultaneously acquired frontal lobe signals, as well as frontal lobe reconstructed signals. Among them, the red dash-dot line represents the 45° fitting line, the green solid line represents the first-order polynomial data fitting line, the r value denotes the Pearson correlation coefficient between the two groups of signals, and the R² value is the coefficient of determination for the two groups of signals. B Pearson correlation coefficients of occipital vs. frontal/reconstructed frontal signals in a healthy subject at sample lengths [500:500:4800] (averaged over 12 trials). C and D Canonical correlation coefficients between prefrontal EEG, prefrontal-reconstructed occipital EEG, and sine–cosine template signals with frequencies [6:0.5:25] Hz at sample length [500:50:4800], calculated by cross-subject and cross-device transfer of the mapping model established based on prefrontal and occipital EEG data of a healthy subject to new users’ EEG data; C shows the analysis results of prefrontal-reconstructed occipital EEG under ssvep stimulation frequencies of 8, 9, and 10 Hz, and D shows the analysis results of original prefrontal EEG. E Signal-to-noise ratios of prefrontal EEG and its reconstructed occipital EEG at a sample length of 4000, calculated using canonical correlation coefficients as the feature index, with 105 trials collected from new users. F Normalized MSE, signal smoothness, cross-spectral coherence, and individual difference compensation rate indices for evaluating the cross-brain region reconstruction performance of DSTF-Net, calculated from 15 blocks collected from new users (where the index for each block is the result of averaging over 7 trials). The frontal-occipital mapping model trained on a healthy subject′s data was transferred to new users′ EEG data via cross-subject and cross-device transfer, to reconstruct occipital activity from their frontal EEG. We analyzed the reconstruction performance post-transfer by calculating the canonical correlation coefficients (CCCs) between: (1) frontal EEG, (2) reconstructed occipital EEG (from frontal signals), and sine-cosine template signals with frequencies ranging from 6 to 25 Hz (step size 0.5 Hz), across sample lengths of [500:50:4800]. Figure 4C presents the analysis results of reconstructed occipital EEG under SSVEP stimulation frequencies of 8, 9, and 10 Hz, while Fig. 4D shows the results of raw frontal EEG. Compared with raw frontal EEG, the CCCs between reconstructed occipital EEG and template signals at the corresponding stimulation frequencies were significantly enhanced. Specifically, the true frequency components that were not prominent in raw frontal EEG achieved effects similar to occipital EEG after reconstruction, and this improvement was consistent across all sample lengths of [500:50:4800]. Furthermore, we collected 105 trials from new users and calculated the signal-to-noise ratio (SNR) of raw frontal EEG and reconstructed occipital EEG at a sample length of 4000, using CCCs as the feature metric. As shown in Fig. 4E, DSTF-Net reconstruction significantly improved the SNR of effective signal components in frontal EEG. Notably, since the frontal-occipital mapping model was trained on a healthy subject′s data and transferred to new users, the occipital signals reconstructed from new users′ frontal EEG are benchmarked against the healthy subject′ occipital signals. We quantified reconstruction fidelity using the normalized mean square error (detailed in the “Methods” subsection “Evaluation metrics”), which normalizes against the MSE of a healthy subject′s frontal-occipital signals (reflecting inherent physiological differences). This normalization eliminates intrinsic frontal-occipital disparities, focusing exclusively on whether new users′ reconstructed occipital signals are closer to healthy subjects′ occipital signals than healthy subjects′ frontal signals are. As shown in Fig. 4F, NMSE scores were consistently above 0.8, indicating that the reconstructed occipital signals were nearly congruent with healthy subjects′ occipital signals (minimal amplitude deviation) and closer to the latter than healthy subjects′ frontal signals were. Additionally, valid physiological EEG components are predominantly distributed in the low-frequency range, whereas high-frequency bands generally contain non-physiological artifacts and noise rather than effective SSVEP neural features. We computed the signal smoothness metric (detailed in the “Methods” subsection “Evaluation metrics”) to verify whether the reconstruction algorithm introduced non-physiological high-frequency artifacts. Results in Fig. 4F showed that the differential RMS of reconstructed occipital signals was significantly smaller than that of healthy subjects’ occipital signals (used as the baseline), demonstrating that the DSTF-Net reconstruction algorithm did not introduce high-frequency noise and was consistent with EEG physiological characteristics. Alpha waves are the core physiological rhythm of the occipital lobe. We calculated the inter-spectral coherence of alpha waves between the new users′ frontal EEG and reconstructed occipital signals. Results in Fig. 4F showed that coherence scores (detailed in the “Methods” subsection “Evaluation metrics”) were overall distributed around 0.25, indicating that the alpha waves in the reconstructed occipital signals are likely independent of the frontal lobe. This implies that the effectiveness of the reconstructed occipital signals stems almost entirely from the robust signal mapping capability of DSTF-Net, rather than sufficient information for occipital reconstruction inherent in the raw frontal signals. Individual differences exist in frontal alpha wave power across subjects. An excellent reconstruction algorithm should compensate for such differences, making the alpha wave power of reconstructed occipital signals close to that of healthy subjects′ occipital lobe rather than remaining at the level of new users′ frontal lobe. We computed the individual difference compensation rate metric (detailed in the “Methods” subsection “Evaluation metrics”). As shown in Fig. 4F, the individual difference compensation scores were mainly distributed around 0.8, indicating that DSTF-Net largely compensated for individual differences, and the deviation of reconstructed occipital signals was much smaller than the inherent frontal lobe differences between new users and healthy subjects. Collectively, the above analysis results demonstrate the competitive performance and feasibility of DSTF-Net in establishing cross-brain-region signal-mapping relationships and the robustness of its adaptation to cross-subject and cross-device transfer. DSTF-Net yields up to 33.47% accuracy improvement Recognition accuracy is the core metric to validate the efficacy of occipital EEG reconstruction from frontal signals. We evaluated the proposed DSTF-Net using data from 20 new users, including 8 brain-injured patients who could only remain in a supine position and 12 healthy subjects. Each subject completed 4 blocks, with 21 or 24 trials per block. For accuracy calculation, we adopted subject-level statistical analysis: we first calculated trial-level decoding accuracy within each block, then averaged the four block-level accuracy values to obtain the individual accuracy for each subject. Finally, we averaged accuracy across all 20 subjects for pairwise comparisons among different methods. All accuracy results visualized in Fig. 5A correspond to such subject-level values. We compared DSTF-Net + CCA with three baselines (direct CCA decoding of raw frontal EEG, GAN + CCA, and diffusion model + CCA), with consistent CCA decoding frameworks to ensure fair comparison.Fig. 5SSVEP decoding accuracy of DSTF-Net: method comparison, ablation studies, and robustness under noisy conditions.A SSVEP recognition accuracy of different methods for supine brain-injured patients, healthy subjects, and all users. B SSVEP decoding accuracy of DSTF-Net with different cross-attention modulation strategies in ablation studies. C SSVEP decoding accuracy of DSTF-Net with ablation of core architectural and training components. D Model recognition accuracy under Gaussian noise interference with noise levels ranging from 0.01 to 0.07. E Model recognition accuracy under impulse noise interference with noise levels ranging from 0.01 to 0.07. (***p < 0.001, **p < 0.05, ns indicates no significance). SSVEP decoding accuracy of DSTF-Net: method comparison, ablation studies, and robustness under noisy conditions. A SSVEP recognition accuracy of different methods for supine brain-injured patients, healthy subjects, and all users. B SSVEP decoding accuracy of DSTF-Net with different cross-attention modulation strategies in ablation studies. C SSVEP decoding accuracy of DSTF-Net with ablation of core architectural and training components. D Model recognition accuracy under Gaussian noise interference with noise levels ranging from 0.01 to 0.07. E Model recognition accuracy under impulse noise interference with noise levels ranging from 0.01 to 0.07. (***p < 0.001, **p < 0.05, ns indicates no significance). As shown in Fig. 5A, subject-level SSVEP recognition accuracy (Mean ± SD) was evaluated across subgroups. For the 8 supine brain-injured patients, the accuracies were 40.52 ± 10.17% (95% CI [31.93, 49.11]) for CCA, 72.66 ± 17.17% (95% CI [58.17, 87.15]) for DSTF-Net + CCA, 41.67 ± 6.14% (95% CI [36.48, 46.86]) for GAN + CCA, and 34.20 ± 3.88% (95% CI [30.93, 37.47]) for diffusion model + CCA. For the 12 healthy subjects, the accuracies were 50.89 ± 6.54% (95% CI [46.75, 55.03]) for CCA, 85.24 ± 8.01% (95% CI [80.17, 90.31]) for DSTF-Net + CCA, 52.74 ± 7.04% (95% CI [48.29, 57.19]) for GAN + CCA, and 32.14 ± 4.86% (95% CI [29.06, 35.22]) for diffusion model + CCA. Across all 20 users, the overall accuracies were 46.74 ± 9.36% (95% CI [42.37, 51.11]) for CCA, 80.21 ± 13.59% (95% CI [73.85, 86.57]) for DSTF-Net + CCA, 48.31 ± 8.99% (95% CI [44.10, 52.52]) for GAN + CCA, and 32.97 ± 5.21% (95% CI [30.52, 35.42]) for diffusion model + CCA. Notably, paired t-tests revealed that DSTF-Net achieved substantial and statistically significant accuracy improvements across all subgroups (all p < 0.001, Bonferroni corrected). In patients, the improvements were 32.14% over CCA (Cohen’s d = 1.67), 30.99% over GAN (Cohen’s d = 1.89), and 38.46% over the diffusion model (Cohen’s d = 2.15). In healthy subjects, the improvements were 34.35% over CCA (Cohen’s d = 3.24), 32.50% over GAN (Cohen’s d = 3.10), and 53.10% over the diffusion model (Cohen’s d = 4.90). Overall, the improvements were 33.47% over CCA (Cohen’s d = 2.79), 31.90% over GAN (Cohen’s d = 2.90), and 47.24% over the diffusion model (Cohen’s d = 3.37). The >80% accuracy and significant improvements are clinically meaningful: it enables effective intent decoding using only frontal EEG, overcoming the occipital electrode inaccessibility barrier for supine or cranially injured patients. Baselines lack practical utility, highlighting DSTF-Net’s translational value. Ablation studies of DSTF-Net To validate the necessity of key architectural and training components in our cross-brain-region neural signal reconstruction network, we conducted systematic ablation experiments on 15 independent data blocks (each containing 21 trials)—ensuring statistical robustness of the results. All experiments were evaluated using the decoding accuracy of reconstructed signals as the sole metric (consistent with our core logic that a reduction in decoding accuracy directly indicates degraded reconstruction precision and loss of valid neural features from source subjects’ occipital regions). The baseline model (full architecture with mid-frequency-focused cross-attention, residual connections across all frequency bands, four-stage training, and triple-band inputs) achieved a decoding accuracy of 0.79126. We tested five cross-attention variants to verify the optimality of mid-frequency-focused modulation (results summarized in Fig. 5B). (1) No cross-attention: Accuracy dropped to 0.76 (−3.95% relative change), confirming inter-band attention is essential for suppressing noise and preserving discriminative neural features across all frequency streams. (2) Low-frequency-focused: Accuracy plummeted to 0.69 (−12.80%), as the low-frequency band (6–90 Hz) lacks the targeted physiological specificity of mid-frequency signals, diluting core occipital neural features in cross-band fusion. (3) Mid-frequency-focused (baseline): Maintained the highest accuracy (0.79126), validating that the mid-frequency band (14–90 Hz)—a physiologically meaningful range for occipital signal discrimination—is the optimal attention hub for balancing noise suppression and feature preservation across low/mid/high streams. (4) High-frequency-focused: Accuracy decreased to 0.73 (−7.74%), due to amplified non-physiological artifacts in the high-frequency band (22–90 Hz) overriding valid neural features in cross-band modulation. (5) Full cross-attention: Accuracy reached 0.77 (−2.69%), with redundancy-induced gradient interference weakening the dominant regulatory role of mid-frequency features in multi-band interaction. Ablation experiments on residual connections, training strategy, and input stream configuration (results summarized in Fig. 5C) further validated the rationality of our network design: Residual connections (all bands) removed: Accuracy declined to 0.7666 (−3.12%). Notably, the residual connections in our network are designed to preserve feature integrity across all low/mid/high frequency streams during cross-band fusion. Their ablation disrupts the direct signal paths for all three frequency bands, leading to gradient vanishing in the deep fusion module and dilution of discriminative neural features across the entire frequency spectrum, which directly reduces the fidelity of reconstructed signals. No staged training (end-to-end): Accuracy plummeted to 0.6535 (−17.41%)—the most severe degradation. This highlights that naive end-to-end training causes unresolvable gradient interference between low/mid/high frequency streams and uneven convergence, failing to extract valid occipital neural features from any single band. Dual-band inputs (randomly removing one stream): To validate the rationality of triple-band inputs, we randomly reduced the network to dual-band inputs (removing one stream arbitrarily). Accuracy dropped to 0.7131 (−9.88%), confirming that triple-band inputs provide complementary feature information that cannot be replaced by any two bands—randomly losing one stream disrupts the integrity of feature representation across the frequency spectrum, leading to reduced reconstruction precision. All ablation results (corresponding to Fig. 5B and C) confirm three key conclusions: Mid-frequency-focused cross-attention is the core architectural design, aligning with EEG physiology to maximize discriminative feature preservation across low/mid/high streams; Four-stage hierarchical training is critical for avoiding gradient interference between frequency bands, with its ablation causing catastrophic performance loss; Residual connections(across all frequency bands) and triple-band inputs complement the architecture—residual links stabilize feature transmission for low/mid/high streams during fusion, while triple-band inputs ensure comprehensive feature representation (validated by performance degradation from random single-stream removal). All results (derived from 15 blocks × 21 trials) validate that our network is a physiologically informed design, where each component synergistically ensures high-fidelity cross-brain-region signal reconstruction, reflected in superior decoding accuracy. Robustness analysis of DSTF-Net under noisy EEG conditions To evaluate the robustness of the proposed DSTF-Net in practical scenarios, we conducted noise-adding experiments on the test set (15 blocks × 21 trials), selecting Gaussian and impulse noise and five gradient noise levels (0.01–0.07). Algorithm performance was quantified by recognition accuracy, with a baseline accuracy (without noise) of 0.791. Noise types were specifically selected to mimic real-world EEG acquisition interferences: Gaussian noise reflects intrinsic system interference, following a zero-mean normal distribution, and was added as xnoise = xclean + σ·N(0,1), where σ is the normalized noise level (0.01–0.07) to ensure physiological relevance. Impulse (spike) noise mimics transient interferences (e.g., poor electrode contact, sudden head movements); the noise level (0.01–0.07, normalized to the amplitude range of clean EEG signals) is directly reflected by the amplitude of the added spike pulses—specifically, for its implementation, five discrete time points per EEG segment were randomly selected, with single-sampling-point spikes added, where spike amplitude was proportional to the noise level (higher noise levels correspond to larger spike amplitudes), and polarity was random, consistent with real transient disturbances. Recognition accuracy under different noise conditions is summarized in Fig. 5D and E. Under Gaussian noise, the algorithm’s accuracy decreased slightly to 0.782 at σ = 0.01 (a 1.14% reduction from the baseline) and to 0.556 at σ = 0.07 (a 30.97% reduction); even at the strongest Gaussian noise, accuracy remained well above the random guess level (0.333), confirming effective suppression of continuous intrinsic interference. The algorithm exhibited superior robustness to impulse noise: accuracy was 0.772 at σ = 0.01 (a 2.40% reduction from the baseline) and 0.610 at σ = 0.07 (a 22.88% reduction)—8.09% higher than the accuracy under the same level of Gaussian noise. This advantage arises from the algorithm’s ability to effectively exclude discrete spike pulses while preserving the intrinsic temporal features of EEG signals. Collectively, these results demonstrate that the proposed algorithm possesses satisfactory robustness. Specifically, accuracy remained above 0.711 at noise levels ≤0.04 (with a maximum reduction of ≤10% from the baseline), and even at the strongest noise level (σ = 0.07), accuracy remained above 0.556, maintaining effective discriminative ability. Additionally, the algorithm showed better robustness to impulse noise than to Gaussian noise, which supports its reliable application in real-world EEG acquisition scenarios where transient interferences are common. Discussion This work addresses a partially unmet need in SSVEP-BCI technology: the inaccessibility of occipital electrode placement for patients with cranial injuries, supine positioning requirements, or skull defects. We propose DSTF-Net, a physiologically informed cross-brain-region reconstruction framework that enables accurate SSVEP decoding exclusively from frontal EEG signals—overcoming the major limitation of weak, attenuated SSVEP components in frontal recordings. The core innovations of DSTF-Net—dynamic sliding point-wise reconstruction, triple-band cross-fusion with mid-frequency-focused attention, stage-wise hierarchical training, and neurophysiologically constrained loss—synergistically capture fine-grained frontal-occipital temporal dependencies, integrate frequency-specific neural features, and enforce biological plausibility. Granger causality analysis validates the signal-level theoretical foundation of our approach, confirming unidirectional occipital-to-frontal SSVEP signal dependency without reverse directional coupling. This inter-regional signal propagation pattern supports the mechanistic rationality of our frontal-to-occipital reconstruction strategy, distinguishing it from generic generative models (GAN and diffusion model) that lack EEG-specific optimization. Quantitative evaluations on 20 new users (including 8 supine brain-injured patients) demonstrate the framework’s superiority: DSTF-Net achieves an overall SSVEP decoding accuracy of 0.8021, representing a 33.47% improvement over direct frontal EEG decoding and substantial gains over GAN (31.90%) and diffusion model (47.24%) baselines. Notably, the framework maintains robust performance across subgroups, delivering 0.7266 accuracy in patients (32.14% improvement over CCA) and 0.8524 in healthy subjects (34.35% improvement over CCA). Complementary metrics—normalized MSE (>0.8), signal smoothness, alpha wave coherence (~0.25), and individual difference compensation rate (~0.8)—confirm high-fidelity reconstruction with minimal artifacts, physiological consistency, and effective mitigation of inter-subject variability. Ablation studies further validate that each core component contributes indispensably to performance, with mid-frequency-focused cross-attention and stage-wise training being the most critical for preserving discriminative occipital neural features. Beyond immediate BCI applications, this work establishes a generalizable paradigm for cross-brain-region neural signal mapping. By enabling neural activity reconstruction from easily accessible electrode sites (e.g., frontal lobe), DSTF-Net lays a foundation for the development of wearable, miniaturized BCI devices—addressing key barriers to portability and clinical translation. For neuroengineering and clinical neuroscience, the framework provides a novel tool to study inter-regional neural communication, with potential extensions to other neural signals (e.g., motor imagery EEG, event-related potentials) and clinical populations (e.g., stroke patients, spinal cord injury survivors). Despite the promising results, several limitations should be acknowledged. First, our validation was carried out with a limited range of stimulation frequencies, and the performance under more frequency settings remains to be verified. Second, this study only evaluated short-term task-related performance. Long-term wearable stability and practical clinical applicability require further validation in future work. Third, the neurophysiological mechanisms underlying cross-brain-region signal reconstruction need to be explored in depth in follow-up research. Future directions include expanding the framework to multi-modal neural data (e.g., fNIRS-EEG fusion) for enhanced robustness, optimizing model efficiency for real-time clinical deployment, and validating long-term performance in larger cohorts of critically ill patients. Additionally, integrating subject-specific fine-tuning strategies may further improve adaptation to extreme physiological variability. Collectively, our findings advance the translation of BCI technology to clinically constrained populations, offering a viable solution for restoring communication and mobility in patients with limited access to traditional neural recording sites. Methods Participants We simultaneously recorded occipital and prefrontal EEG signals from a 33-year-old healthy male subject, referred to as the source subject. Only prefrontal EEG signals were recorded from 8 brain-injured patients (5 males, 3 females; mean age, 48.2 ± 12.2 years) confined to the supine position and 12 healthy subjects (10 males, 2 females). All of them had normal eyesight and hearing, and there were no reported cognitive impairments. Experimental equipment For the source subject, EEG recordings were obtained using a 1000 Hz sampling system with Cz as the reference electrode and FPz as the ground electrode (ZhenTec NT1, ZhenTec Intelligence, China53). Occipital channel acquisition included POz, PO3, PO4, PO5, PO6, Oz, O1, and O2. The system’s modular design permitted supplementary prefrontal monitoring via five adhesive patch electrodes placed over the frontal lobe. For the frontal signal acquisition group, we used a custom-built 1000 Hz EEG acquisition system designed and constructed in-house (Brain-Apex, Xi’an XINYUESHENG Technology Co., Ltd., China). The ZhenTec NT1 system adopts a rigid standard EEG cap, which hinders stable prefrontal electrode placement for supine brain-injured patients. By contrast, our custom-built system supports adhesive patch electrodes tailored for forehead placement, which is compatible with supine clinical postures. Reference and ground electrodes were placed on the left and right mastoids, respectively, with five patch electrodes positioned over the prefrontal lobe for signal recording. Offline experiment SSVEP visual stimulation was delivered via a 144 Hz refresh rate monitor, with the source subject seated at a 1 m viewing distance. The display presented four circular stimuli (120-pixel radius) flickering at frequencies of 8, 9, 10, and 11 Hz. During testing, the participant maintained fixation on individual targets for 5 seconds, followed by a 5-s rest interval before progressing to the next stimulus in sequence. Thirteen complete cycles of all four targets were recorded in an offline experimental session. For the prefrontal signal acquisition group, the display presented flickering light patterns of 8, 9, and 10 Hz with a radius of 120 pixels. Considering the limited experimental time in clinical scenarios and irregular inpatient nursing arrangements, prolonged data collection with four frequencies would increase patient burden and compromise signal quality. Therefore, only three stable and clinically friendly stimulation frequencies were adopted to balance experimental validity and patient comfort. Subjects were required to fixate on a single target for 5 s, followed by a 5-s rest period, before proceeding to the next stimulus in sequence. During the offline experimental phase, we recorded seven complete cycles of all three stimulus targets. For the patient cohort, the monitor was positioned via a bracket at an ~60 cm distance from the participant. The remaining procedures were consistent with those used for collecting EEG data from the source subject. Online experiment DSTF-Net was trained on offline experimental data from the source subject and fine-tuned on offline data from the prefrontal signal acquisition group. Online experiments were conducted on 20 subjects in this group (8 stroke patients, 12 healthy subjects) to validate the model′s online signal reconstruction capability. Consistent with the offline clinical validation setting, the display presented three flickering light stimuli at 8, 9, and 10 Hz; a red cross icon randomly appeared at the center of one of the three stimuli to indicate the target for fixation. Following 5 s of maintained fixation, the system′s recognition result was displayed as a green cross icon at the center of the corresponding target. The display then turned black for 5 s, allowing a brief rest before the commencement of the next fixation trial. Dynamic sliding point-wise reconstruction with triple-band cross-fusion network (DSTF-Net) To address the clinical limitation of occipital electrode inaccessibility for SSVEP-BCI, we propose a dynamic sliding point-wise reconstruction with a triple-band cross-fusion network (DSTF-Net), which enables high-fidelity occipital SSVEP reconstruction from frontal EEG signals via three core innovations: a dynamic sliding point-wise reconstruction strategy, a triple-band cross-fusion convolutional sub-network, and a stage-wise hierarchical training mechanism. The overall architecture of DSTF-Net is illustrated in Fig. 6 and Supplementary Table 2, comprising four functional modules: multi-band EEG preprocessing, parallel spectral-spatial feature extraction, cyclic cross-attention fusion, and neurophysiologically constrained optimization.Fig. 6Overall architecture of the dynamic sliding point-wise reconstruction with triple-band cross-fusion network (DSTF-Net).Schematic of the DSTF-Net architecture for high-fidelity occipital SSVEP reconstruction from frontal EEG (addressing occipital electrode inaccessibility). The framework comprises four interconnected modules integrating four core innovations: (1) Multi-band EEG preprocessing: Preprocesses raw frontal EEG (e.g., filtering, normalization) and decomposes it into low-, mid-, and high-frequency bands for subsequent feature extraction. (2) Parallel spectral-spatial feature extraction: Uses three parallel convolutional streams (one per frequency band) and implements the dynamic sliding point-wise reconstruction strategy to capture frontal-occipital temporal dependencies. (3) Cyclic cross-attention fusion: Acts as the triple-band cross-fusion sub-network to integrate frequency-specific features from the three parallel streams through cyclic cross-attention (mid-frequency features modulate low- and high-frequency feature weights), depth concatenation, and residual connections to ensure robust inter-frequency interaction and feature integrity. (4) Neurophysiologically constrained optimization: Incorporates a Granger causality-based neurophysiologically constrained loss function, enforcing the biological plausibility of reconstructed occipital signals and cooperating with the stage-wise hierarchical training mechanism to mitigate gradient interference and ensure stable model convergence. Arrows indicate the direction of signal flow and feature transmission between modules. Color coding (consistent with the figure) distinguishes different frequency bands and functional sub-modules for clarity. Overall architecture of the dynamic sliding point-wise reconstruction with triple-band cross-fusion network (DSTF-Net). Schematic of the DSTF-Net architecture for high-fidelity occipital SSVEP reconstruction from frontal EEG (addressing occipital electrode inaccessibility). The framework comprises four interconnected modules integrating four core innovations: (1) Multi-band EEG preprocessing: Preprocesses raw frontal EEG (e.g., filtering, normalization) and decomposes it into low-, mid-, and high-frequency bands for subsequent feature extraction. (2) Parallel spectral-spatial feature extraction: Uses three parallel convolutional streams (one per frequency band) and implements the dynamic sliding point-wise reconstruction strategy to capture frontal-occipital temporal dependencies. (3) Cyclic cross-attention fusion: Acts as the triple-band cross-fusion sub-network to integrate frequency-specific features from the three parallel streams through cyclic cross-attention (mid-frequency features modulate low- and high-frequency feature weights), depth concatenation, and residual connections to ensure robust inter-frequency interaction and feature integrity. (4) Neurophysiologically constrained optimization: Incorporates a Granger causality-based neurophysiologically constrained loss function, enforcing the biological plausibility of reconstructed occipital signals and cooperating with the stage-wise hierarchical training mechanism to mitigate gradient interference and ensure stable model convergence. Arrows indicate the direction of signal flow and feature transmission between modules. Color coding (consistent with the figure) distinguishes different frequency bands and functional sub-modules for clarity. The input preprocessing and dynamic sliding point-wise reconstruction strategy of the model are as follows: (a) Frontal and occipital EEG signals were synchronously recorded from the source subject (5 frontal channels, 8 occipital channels) at a sampling rate of 1000 Hz. For preprocessing, frontal EEG signals were first decomposed into three overlapping frequency bands (low: 6–90 Hz, mid: 14–90 Hz, high: 22–90 Hz) using a finite impulse response (FIR) filter bank. This broadband range follows the classic filter-bank configuration of mainstream SSVEP analysis54 to accommodate diverse stimulation frequencies and their higher-order harmonics for improved method generalizability, followed by normalization to the range [0,1] using min-max scaling to mitigate inter-subject amplitude variations. (b) A pivotal innovation of DSTF-Net lies in its dynamic sliding point-wise reconstruction strategy, which is specifically tailored to tackle the inherent high time-variability of EEG signals—a core challenge that has long limited the accuracy of cross-brain-region neural signal reconstruction. EEG signals, particularly attenuated SSVEP components propagating from the occipital to the frontal lobe, exhibit rapid, fine-scale temporal fluctuations that encode critical neural coupling information. To capture these dynamic temporal details, we employ a fixed-length temporal window (200 ms, 200 time points) that is slid sequentially along the preprocessed frontal EEG time series, yielding (T−200) overlapping input segments (where T denotes the total number of time points in the EEG sequence). For each segment, the framework focuses on predicting the corresponding occipital EEG signal at the terminal time point of the window—this design represents a fundamental departure from the conventional static n-to-m mapping paradigm (which directly reconstructs n occipital samples from m frontal samples). By shifting the focus to point-wise reconstruction anchored on sliding temporal contexts, the strategy enables ultra-fine-grained modeling of the time-varying frontal-occipital neural coupling dynamics, a capability that is indispensable for retrieving weak SSVEP signals that are severely attenuated by skull and scalp tissues. The architecture of the triple-band cross-fusion convolutional sub-network is as follows: The triple-band cross-fusion convolutional sub-network of DSTF-Net consists of three parallel convolutional streams (one for each frequency band) and a cyclic cross-attention fusion module, tailored to extract and integrate spectral–spatial–temporal features from frontal EEG signals. As shown in Fig. 6 and Supplementary Table 2, each parallel convolutional stream follows an identical structure but is independently initialized to adapt to the unique characteristics of its target frequency band. The stream architecture includes three cascaded blocks: (1) A temporal convolution block with a 1 × C kernel (where C is the number of frontal channels) to capture time-domain neural oscillations; (2) A spatial convolution block with a C×1 kernel to extract inter-electrode spatial patterns; (3) A depthwise separable convolution block (grouped convolution followed by point-wise convolution) to refine frequency-specific features. Each block is followed by batch normalization, an ELU activation function, and a 20% dropout layer to prevent overfitting. The output of each stream is a 1-D feature vector mapped to the dimension of occipital EEG signals via a fully connected layer (fc_1, fc_2, fc_3 for low, mid, and high bands, respectively). The cyclic cross-attention fusion module dynamically integrates features from the three streams, with mid-frequency features serving as the attention hub. Specifically, mid-frequency features are fed into two fully connected layers followed by sigmoid gating to generate attention weights for low- and high-frequency features, respectively. The weighted low/high-frequency features are then depth-concatenated with their original features, and mid-frequency features are retained as residuals to preserve core information. Finally, the refined low-, mid-, and high-frequency features are concatenated and processed via a 1 × 3 convolution layer and a fully connected layer (fc_final in Supplementary Table 2) to output the reconstructed occipital EEG signal. The process of the stage-wise hierarchical training mechanism is as follows: To address gradient interference and uneven convergence in multi-band fusion networks, we designed a four-stage hierarchical training strategy. In Stage 1–3, we sequentially unlocked and trained the low-, mid-, and high-frequency convolutional streams (corresponding to the Multi-Band Feature Extraction Module in Supplementary Table 2) while freezing other layers, ensuring each stream independently converges to optimal single-band feature extraction. In Stage 4, we unlocked the entire fusion module (corresponding to the Prediction Generator in Supplementary Table 2) and increased the training epochs from 4 to 8 to enhance dynamic feature integration. The loss function of the model is formulated as follows: The total loss function for model training is a weighted combination of three complementary components, designed to balance reconstruction accuracy, model generalization, and neurophysiological plausibility:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{L}}_{{\rm{total}}}={{L}}_{{\rm{Huber}}}+{\lambda }\cdot {{L}}_{{L}2}+{{\lambda }}_{{\rm{causal}}}\cdot {{L}}_{{\rm{causal}}}$$\end{document}Ltotal=LHuber+λ⋅LL2+λcausal⋅Lcausal Huber loss (robust to outliers in EEG signals):2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${L}_{{{\rm {Huber}}}}=\left\{\begin{array}{c}\frac{1}{2}|\left|{\hat{y}}_{t}-{y}_{t}\right|{|}_{2}^{2}{if||}{\hat{y}}_{t}-{y}_{t}|{|}_{2}\le \delta ,\\ \delta (\Vert {\hat{y}}_{t}-{\hat{y}}_{t}{\Vert }_{2}-\frac{1}{2}\delta )\,{{\rm {otherwise}}},\end{array}\right.$$\end{document}LHuber={12||y^t−yt||22if||y^t−yt||2≤δ,δ(∥y^t−y^t∥2−12δ)otherwise, with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\hat{y}}_{t}$$\end{document}yˆt denoting the reconstructed occipital EEG signal at time t, yt the ground-truth occipital EEG signal, and δ = 1 (a empirically optimized threshold for EEG data). L2 regularization (to prevent overfitting):3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{L}}_{\mathrm{L}2}=\frac{1}{2}{\sum }_{{l}\in {u}}{\sum }_{{w}\in {{W}}_{{l}}}{{w}}^{2}$$\end{document}LL2=12∑l∈u∑w∈Wlw2where u is the set of unlocked layers in each training stage, and Wl represents the trainable weights (and biases) of layer l. The regularization coefficient λ = 0.01 was selected via grid search over [0.001, 0.01, 0.1]. Temporal causal constraint (to enforce neurophysiological plausibility derived from frontal-occipital causality analysis):4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{L}}_{{\rm{causal}}}=\max\left(0,{{f}}_{{\rm{pre}}\to {\rm{occ}}}-{{f}}_{{\rm{occ}}\to {\rm{pre}}}\right)$$\end{document}Lcausal=max(0,fpre→occ−focc→pre)where fpre→occ and focc→pre are the Granger causality F-statistics for frontal-to-occipital and occipital-to-frontal signal propagation, respectively. We designed this causal constraint based on our prior analysis (see the analysis results in Fig. 3) that confirmed unidirectional SSVEP propagation from the occipital to the frontal lobe (i.e., no frontal-to-occipital causality was detected), with frontal SSVEP components dominated by occipital signals. The constraint is intended to enforce this neurophysiological mechanism in the reconstructed signals by penalizing any deviation where frontal-to-occipital causality exceeds the reverse direction, ensuring the model′s outputs adhere to the biological basis that underpins the feasibility of cross-brain-region reconstruction. The causal strength coefficient λcausal = 0.005 was selected to balance constraint strength and reconstruction accuracy. For Stages 1–3 (single-band stream training), only LHuber and LL2 were used; for Stage 4 (fusion module training), all three loss components were integrated to ensure both reconstruction accuracy and biological interpretability of DSTF-Net. After training the DSTF-Net on fronto-occipital EEG data from the healthy source subject, the procedure for transferring it to new subjects is as follows: For cross-subject and cross-device transfer to new users’ data, we adopted a layer-wise fine-tuning strategy for the pre-trained DSTF-Net. This fine-tuning process serves as a domain adaptation mechanism to calibrate the model to the dual distribution shifts introduced by new individuals and the new EEG device. Specifically, most layers of the pre-trained model were frozen to preserve the pre-learned cross-brain-region mapping features and prevent overfitting to individual-specific or device-specific noise, while only the fully connected layers were unfrozen for fine-tuning to absorb the data distribution shifts caused by both inter-subject variability and cross-device discrepancies. The detailed fine-tuning process and subsequent inference/decoding pipeline are described as follows: Fine-tuning of pre-trained model: First, preprocessed frontal EEG signals from new users (including clinical patients) were used as the fine-tuning dataset. The pre-trained model parameters were loaded, and the learnable layers were categorized into frozen and unfrozen subsets: all layers except the fully connected layers were frozen to preserve the pre-learned cross-brain-region mapping features, while only the fully connected layers were unfrozen for parameter updating. The fine-tuning was conducted in four sequential stages, with specific settings as follows: (1) Unfreeze fully connected layer ‘fc_1’ (see details in Supplementary Table 2) of the low-frequency stream, freeze all other layers; train for 4 epochs with an initial learning rate of 0.005, decay rate of 0.1, and mini-batch size of 64. (2) Unfreeze fully connected layer ‘fc_2’ (see details in Supplementary Table 2) of the mid-frequency stream, freeze all other layers; adopt the same training parameters as Stage 1. (3) Unfreeze fully connected layer ‘fc_3’ (see details in Supplementary Table 2) of the high-frequency stream, freeze all other layers; training parameters remain consistent with previous stages. (4) Unfreeze fully connected layer ‘fc_final’ (see details in Supplementary Table 2), freeze all other layers; extend the training epochs to 16 while keeping other parameters unchanged, to enhance the integration of multi-band features adapted to new users. Inference and SSVEP decoding: After fine-tuning, only the preprocessed frontal EEG signals from new users were fed into the optimized DSTF-Net to generate reconstructed occipital EEG signals. SSVEP decoding was implemented using canonical correlation analysis (CCA) with the following steps: (1) Construct reference templates: Sine and cosine waves at the target stimulation frequencies (8, 9, 10 Hz) and their first to fourth harmonics were generated to form the CCA reference templates. (2) Canonical correlation calculation: Compute the canonical correlation coefficient (CCC) between the reconstructed occipital EEG signals and each reference template. (3) Weighted correlation and decoding decision: The CCCs corresponding to different harmonics of each stimulation frequency were weighted (weight coefficients: 1, 0.67, 0.21, 0.11 for the 1st–4th harmonics, respectively). The stimulation frequency with the highest weighted CCC was selected as the final decoded result. Canonical correlation analysis (CCA) For EEG data X recorded from multiple electrodes and the reference signal Y, the goal of CCA32 is to find two projection vectors wX and wY so that the two groups of linear combination signals wXTX and wYTY have the largest correlation coefficients. The canonical correlation coefficient was calculated by5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rho =\max \frac{{E}({{w}}_{{X}}^{\mathrm{T}}{X}{{Y}}^{\mathrm{T}}{{w}}_{{Y}})}{\sqrt{{E}({{w}}_{{X}}^{\mathrm{T}}{X}{{X}}^{\mathrm{T}}{{w}}_{{X}}){E}({{w}}_{{Y}}^{\mathrm{T}}{Y}{{Y}}^{\mathrm{T}}{{w}}_{{Y}})}}$$\end{document}ρ=maxE(wXTXYTwY)E(wXTXXTwX)E(wYTYYTwY) The reference signals Y were constructed at the stimulation frequency f:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{Y}}_{{f}}=\left\{\begin{array}{c}\cos 2{{\pi }}{ft}\\ \begin{array}{c}\sin 2{{\pi }}{ft}\\ \begin{array}{c}\mathrm{...}\\ \begin{array}{c}\cos 2{k}{{\pi }}{ft}\\ \sin 2{k}{{\pi }}{ft}\end{array}\end{array}\end{array}\end{array}\right\},{t}=1/{{f}}_{{\mathrm{s}}}{,...,l/}{{f}}_{{\text {s}}}$$\end{document}Yf={cos2πftsin2πft...cos2kπftsin2kπft},t=1/fs,...,l/fswhere k is the number of harmonics and this study set it to 3, fs is the sampling rate, and l represents the number of sample points. By calculating the canonical correlation coefficients between X and the reference signals at all stimulus frequencies, the corresponding target with the maximum correlation coefficient is identified as the focused target. Filter bank canonical correlation analysis (FBCCA) FBCCA54 first filters the EEG signal X using d sub-band filters with different upper and lower cutoff frequencies, respectively, to obtain the sub-band filtered signals XSB,1, XSB,2, …, XSB,d. Then, it calculates the maximum canonical correlation coefficient of each sub-band signal at frequency f through CCA:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\rho }_{\mathrm{SB},{i}}=\max \frac{{{w}}_{{X}}^{\mathrm{T}}{{X}}_{{\mathrm{SB}},{i}}{{{Y}}_{{f}}}^{\mathrm{T}}{{w}}_{{Y}}}{\sqrt{({{w}}_{{X}}^{\mathrm{T}}{{X}}_{{\mathrm{SB}},{i}}{{X}}_{{\mathrm{SB}},{i}}^{\mathrm{T}}{{w}}_{{X}})({{w}}_{{Y}}^{\mathrm{T}}{{Y}}_{{f}}{{Y}}_{{f}}^{\mathrm{T}}{{w}}_{{Y}})}}$$\end{document}ρSB,i=maxwXTXSB,iYfTwY(wXTXSB,iXSB,iTwX)(wYTYfYfTwY)Here, XSB,i denotes the ith subband signal. Finally, FBCCA integrates the maximum canonical correlation coefficients of the d sub-band signals using a weighting function to form the feature discriminant coefficient at frequency f:8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{p}}_{{f}}=\mathop{\sum }\limits_{{i}=1}^{{d}}{w}({i})\cdot {({\rho }_{{\mathrm{SB}},{i}})}^{2}$$\end{document}pf=∑i=1dw(i)⋅(ρSB,i)2Here, w(i) represents the weight corresponding to the canonical correlation coefficient of the ith subband signal, which is calculated as9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${w}({i})={{i}}^{-{a}}+{b},{i}\in [1{d}]$$\end{document}w(i)=i−a+b,i∈[1d] In ref. 54, parameters a = 1.25 and b = 0.25 were used for filter design. This study set the filter subband parameter d = 3, with cutoff frequencies defined as [6–90 Hz], [14–90 Hz], and [22–90 Hz] for the three subband filters. After computing integrated coefficients across all stimulus frequencies, FBCCA determines the frequency corresponding to the maximum coefficient as the user’s target gaze frequency. Signal-to-noise ratio (SNR) The SNR was defined as the ratio of the canonical correlation coefficient of the signal at its stimulation frequency f to those at the four adjacent frequencies (f−2·Δf、f-Δf、f + 2·Δf、f + Δf), where Δf denotes the frequency resolution of the analyzed signal. Causal analysis between the frontal and occipital Lobes To investigate the spatiotemporal correlations and causal relationships between signals transmitted from the occipital region to the frontal lobe in a healthy subject, we performed two analyses: (1) the computation of frequency-domain causality from the occipital region to the frontal region, and (2) the computation of band-limited pairwise Granger causality (G-causality) between channels. The analyses were conducted using the multivariate Granger causality (MVGC) toolbox49. Initially, the raw signal X was downsampled from 1000 to 64 Hz in order to reduce computational complexity while preserving the frequency range of interest. A vector autoregressive (VAR) model was subsequently fitted to the multi-trial, multivariate time series data using an empirically determined model order of 9. The estimated VAR parameters were used to compute the autocovariance sequence. Based on the autocovariance sequence, we estimated pairwise Granger causality in both the time and frequency domains. Time-domain causality can be decomposed into its constituent frequency components. Conversely, integrating the spectral causality over the entire frequency range yields the corresponding time-domain causality. Thus, band-limited time-domain causality was computed by integrating the spectral causality over specified frequency intervals. In this study, integration was performed over two frequency bands corresponding to the SSVEP stimulus: the fundamental frequency (f0) and its harmonic (2f0). Finally, p-values for the sample MVGC were computed based on the theoretical asymptotic null distribution, and statistical significance was assessed at a significance level of a = 0.01. Evaluation metrics To comprehensively assess the quality of cross-subject occipital EEG reconstruction (from new users′ prefrontal signals to source subjects′ occipital-like signals), we designed four complementary metrics covering waveform similarity, physiological plausibility, and cross-subject compatibility (all normalized to a 0–1 scale, with higher scores indicating better reconstruction performance). All signals were preprocessed with linear detrending (detrend, MATLAB R2025a) to remove low-frequency baseline drift before metric calculation. MSE-based waveform similarity: This metric quantifies the amplitude matching degree between the new users′ reconstructed occipital signal (Srecon) and the ground-truth occipital signal of the source subject (Ssrc-occ), normalized by the inherent difference between the source subject′s prefrontal (Ssrc-fron) and occipital signals (to eliminate the physiological gap between brain regions). Step 1: Temporal alignment. EEG signals naturally exhibit minor temporal drift (±50 ms) due to physiological artifacts (e.g., respiration, heart rate). We first aligned Srecon with Ssrc-occ using cross-correlation to ensure the MSE reflects amplitude differences rather than temporal misalignment:10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\text {offset}}={\text {arg}}{\max }_{\tau }\left\{{\text {xcorr}}({S}_{{\text {src}}-\text {{occ}}},{S}_{{\text {recon}}})\right\}-{N}_{{\text {src}}-{\text {occ}}}$$\end{document}offset=argmaxτ{xcorr(Ssrc−{occ},Srecon)}−Nsrc−occwhere xcorr(·) denotes the cross-correlation function, τ is the time lag (in samples), and Nsrc-occ is the length of Ssrc-occ. The aligned signals \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${S}_{{\text {recon}}}^{{\text {align}}}$$\end{document}Sreconalign and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${S}_{{\text {src}}-{\text {occ}}}^{{\text {align}}}$$\end{document}Ssrc−occalign were truncated to the same length based on offset. Step 2: MSE calculation and normalization. The raw MSE between aligned signals was computed as:11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{MSE}}_{\mathrm{recon}}=\frac{1}{{N}}{\sum }_{{i}=1}^{{N}}{({{S}}_{\mathrm{recon}}^{\mathrm{align}}({i})-{{S}}_{\mathrm{src}-\mathrm{occ}}^{\mathrm{align}}({i}))}^{2}$$\end{document}MSErecon=1N∑i=1N(Sreconalign(i)−Ssrc−occalign(i))2where N is the length of aligned signals. A baseline MSE (MSEbase) was calculated using the same alignment and MSE procedure for Ssrc-fron and Ssrc-occ (capturing the inherent prefrontal-occipital difference of the source subject). The final similarity score was:12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{score}}_{\mathrm{MSE}}=\max\left(\mathrm{0,1}-\frac{{\mathrm{MSE}}_{\mathrm{recon}}}{{\mathrm{MSE}}_{\mathrm{base}}}\right)$$\end{document}scoreMSE=max(0,1−MSEreconMSEbase) Physiological smoothness: This metric validates whether the reconstructed signal adheres to the low-frequency physiological characteristics of EEG, by quantifying the absence of high-frequency artifacts (e.g., muscle noise, algorithm-induced overfitting). Smoothness was measured using the root mean square (RMS) of first-order differences (reflecting amplitude change rate):13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\text {RMS}}(\Delta S)=\sqrt{\frac{1}{N-1}{\sum }_{i=1}^{N-1}{(S(i+1)-S(i))}^{2}}$$\end{document}RMS(ΔS)=1N−1∑i=1N−1(S(i+1)−S(i))2where ΔS denotes the first-order difference of signal S. The smoothness score was normalized to Ssrc-occ (the gold standard for physiological smoothness):14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{score}}_{\mathrm{smooth}}=\max\left(\mathrm{0,1}-\frac{\mathrm{RMS}(\Delta {{S}}_{\mathrm{recon}})}{\mathrm{RMS}(\Delta {{S}}_{\mathrm{src}-\mathrm{occ}})}\right)$$\end{document}scoresmooth=max(0,1−RMS(ΔSrecon)RMS(ΔSsrc−occ)) A score of 1 indicates the reconstructed signal has identical smoothness to the ground-truth occipital EEG of the source subject; lower scores indicate increased high-frequency noise. Alpha-band coherence: This metric evaluates the physiological relevance between the new user’s prefrontal input (Snew-fron) and the reconstructed signal, focusing on the alpha band (8–13 Hz)—the most prominent rhythmic feature of occipital EEG. Coherence (a frequency-domain measure of linear synchronization between two signals) was computed using the multi-taper spectral estimation method:15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{C}}_{{xy}}({f})=\frac{{\left|{{P}}_{{xy}}({f})\right|}^{2}}{{{P}}_{{xx}}({f})\cdot {{P}}_{{yy}}({f})}$$\end{document}Cxy(f)=|Pxy(f)|2Pxx(f)⋅Pyy(f)where Pxy(f) is the cross-spectral density of Snew-fron and Srecon, and Pxx(f), Pyy(f) are their auto-spectral densities. The alpha-band coherence score was the mean coherence over 8–13 Hz, normalized to 0–1:16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{score}}_{\mathrm{coh}}=\left(\frac{1}{{{N}}_{\alpha }}{\sum }_{{f}\in [\mathrm{8,13}]}{{C}}_{{xy}}({f})\right)$$\end{document}scorecoh=(1Nα∑f∈[8,13]Cxy(f))where Nα is the number of frequency bins in the alpha band. Individual difference compensation rate: This metric quantifies the algorithm’s ability to compensate for inter-individual differences in prefrontal alpha power between new users and source subjects, ensuring the reconstructed signal matches the source subject’s occipital alpha power (rather than retaining the new user’s prefrontal alpha characteristics). Step 1: Alpha power calculation. Alpha-band signals were isolated using a 2nd-order Butterworth bandpass filter (8–13 Hz, zero-phase filtering via filtfilt to avoid phase distortion). Alpha power was computed as the square of the RMS of the filtered signal:17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{P}}_{\alpha }({S})={\left(\sqrt{\frac{1}{{N}}{\sum }_{{i}=1}^{{N}}{{S}}_{\alpha }{({i})}^{2}}\right)}^{2}$$\end{document}Pα(S)=(1N∑i=1NSα(i)2)2where Sα is the alpha-band component of signal S. Step 2: Compensation rate calculation. We first defined the inter-individual prefrontal alpha difference (between the source subject and the new user):18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\delta }}_{\mathrm{fron}}=\frac{\left|{{P}}_{\alpha }({{S}}_{\mathrm{src}-\mathrm{fron}})-{{P}}_{\alpha }({{S}}_{\mathrm{new}-\mathrm{fron}})\right|}{{{P}}_{\alpha }({{S}}_{\mathrm{src}-\mathrm{fron}})}$$\end{document}δfron=|Pα(Ssrc−fron)−Pα(Snew−fron)|Pα(Ssrc−fron)and the reconstruction error in alpha power:19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\delta }}_{\mathrm{recon}}=\frac{\left|{{P}}_{\alpha }({{S}}_{\mathrm{recon}})-{{P}}_{\alpha }({\mathrm{S}}_{\mathrm{src}-\mathrm{occ}})\right|}{{{P}}_{\alpha }({{S}}_{\mathrm{src}-\mathrm{occ}})}$$\end{document}δrecon=|Pα(Srecon)−Pα(Ssrc−occ)|Pα(Ssrc−occ) The compensation rate (normalized to 0–1) was20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{score}}_{\mathrm{comp}}=\max\left(\mathrm{0,1}-\frac{{{\delta }}_{\mathrm{recon}}}{{{\delta }}_{\mathrm{fron}}}\right)$$\end{document}scorecomp=max(0,1−δreconδfron) Ethics All experimental protocols of this study were approved by the Ethics Committee of the Second Affiliated Hospital of Xi’an Jiaotong University (approval number: 2024YS237). A total of 20 participants (8 patients and 12 healthy subjects) were recruited, and written informed consent was obtained from all participants (or their legal guardians if necessary). EEG data were collected using non-invasive procedures without harmful stimulations, and participants’ privacy was protected through data anonymization. No ethical conflicts or commercial interests were involved in this study, and all research procedures were strictly followed according to the approved ethical protocols. Supplementary information Supplementary information Supplementary information