acceptodds
Under review as a conference paper at ICLR 2027

RIGS-EEG: Does Reference-Invariant Graph-Spectral Reconstruction Improve Missing EEG Channel Recovery?

Abstract

Whole-channel loss removes local temporal observations from EEG, making reconstruction depend on information from other electrodes. Common-reference components and high-amplitude observations can affect spatial propagation, while temporal fitting alone provides insufficient constraints on missing potentials. We introduce RIGS-EEG, a five-step unrolled network that combines common-reference separation, reliability-weighted spatial initialization, and graph–time-frequency shrinkage. The model represents relative potentials on an electrode graph, progressively refines estimates through lightweight spectral gains, and exactly preserves observed samples at the output, using only 970 learnable parameters. With a fixed mask and reliability weights invariant to common shifts, the architecture is equivariant to additive common-reference signals. We evaluate 25%, 50%, and 75% whole-channel missingness on PhysioNet, SEED-IV, and MCD against 14 baseline implementations. At 75% missingness, mean NRMSE is 0.0723, 0.1274, and 0.0922, respectively. MCD evaluation uses held-out subjects and averages three recording stages within each subject. With a fixed downstream classifier and 25% missing channels, reconstruction improves balanced accuracy by 3.44–8.37 percentage points and Cohen’s kappa by 0.068–0.122 over corrupted inputs across four PhysioNet movement and motor-imagery tasks. Ablations support the contributions of spatial and spectral constraints and the benefit of reliability weighting on SEED-IV. These findings support explicitly organizing reference, geometry, and spectral information within a compact reconstruction model to recover incomplete EEG while retaining information useful for downstream recognition.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG reconstruction, missing-channel imputation, graph-spectral unrolling, reliability weighting, reference equivariance

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Yvc5OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

In this paper, the authors address whole-channel EEG reconstruction, where one or more electrodes lose their entire temporal signal over an analysis window. The authors propose RIGS-EEG, a lightweight, domain-structured pipeline containing only 970 trainable parameters, the overall contributions can be highlighted as follows:

  • The authors downweight abnormally high-amplitude channels and subtract the reliability-weighted common reference across visible electrodes so that missing zero-filled channels do not bias the reference estimate.
  • They also initializes missing channels in the relative-potential domain using a 3D Gaussian distance kernel across all visible electrodes, ensuring an initial estimate even when a sensor's immediate local neighborhood is completely missing.
  • They then use a 5-Step Unrolled Graph-Time-Frequency Refinement where they refine the estimate over five unrolled iterations, each consisting of a data-consistency step, a projection into a 4-NN Graph Fourier + Short-Time Fourier Transform (STFT) basis, magnitude shrinkage via a tiny 194-parameter MLP conditioned on graph/temporal frequency and sensor coverage, and an inverse transform, followed by reference restoration and hard projection of observed samples. Contributions & Validation: The authors prove that under a fixed mask and shift-invariant reliability weights, the architecture is equivariant to additive common-reference shifts. They validate RIGS-EEG against 14 classical and deep learning baselines across three datasets (PhysioNet, SEED-IV, and an in-house MCD dataset) at 25%, 50%, and 75% whole-channel missingness. RIGS-EEG achieves the lowest mean NRMSE across all nine settings, recovers 86%–94% of downstream classification accuracy lost to 25% channel missingness when evaluated with a frozen EEGNet classifier on PhysioNet, and includes component ablations isolating the graph, spectral, reference, and reliability modules.
Critical strengths
  1. I very much enjoy the parameter efficiency of the proposed model. In the paper we see that RIGS-EEG beats 14 baselines across all nine dataset and missing-channel settings while using only 970 trainable parameters (194 per step). This includes beating much heavier Transformer models (like SAITS, PatchTST, and DMAE-EEG) and diffusion models (like CSDI). Combining a full-head Gaussian average for the initial guess (which still works when 75% of channels are gone and all local neighbors are missing) with a 4-nearest-neighbor graph and STFT filter for the refinement steps makes a lot of practical sense. It nicely shows that building the basic physics of EEG (3D sensor locations, reference removal, and space-time-frequency filtering) directly into the pipeline works better for missing channel recovery than simply throwing a massive neural network at the problem. And I applaud the authors for this choice.

  2. In addition to measuring reconstruction error (NRMSE) at 25%, 50%, and 75% missing channels against 14 baselines, the authors test whether the recovered brainwaves are actually usable by a pre-trained EEGNet classifier across four motor tasks on PhysioNet (Table 3). Because it is easy for a reconstruction model to get a low NRMSE just by outputting overly smooth, flat waves, showing that a frozen classifier recovers 86% to 94% of the accuracy lost to missing channels is a great way to prove that useful signal details are preserved.

Critical weaknesses
  1. Now instead of testing one fixed model and loss function across all benchmarks, the paper switches between two different setups depending on the dataset (Sections 3.2, 3.5, 4.1, and Appendix D). On PhysioNet ("v1"), reliability weighting is turned off (q_c = 1), and the model is trained with standard MSE plus extra spectral and spatial loss terms (whose exact equations and weights are not given in the text). On SEED-IV and MCD ("v2"), reliability weighting is turned on, and the model is trained with channel-normalized RMSE (Eq. 12) without those extra loss terms. In the PhysioNet ablation study (Section 4.4, Line 426), the w/o graph and w/o spectral variants remove both the model components and their additional training losses simultaneously, making it impossible to determine how much of the performance drop is attributable to the architecture versus the loss function. Meanwhile, w/o reliability is left out for PhysioNet altogether.

Together this choice makes it quite difficult for me as a reviewer to understand where the main gains are coming from.

  1. Although reference invariance is highlighted in the title and main contributions, the analysis and empirical gains around referencing have clear gaps. The method only accounts for adding a single global reference wave to all channels equally (Appendix D). It does not address bipolar montages (differences between pairs of adjacent electrodes) or general linear re-referencing, which are very common in clinical and real-world EEG. Even for a common reference, removing the reference module (w/o reference in Table 4) makes zero difference on PhysioNet at 25% missingness (0.0509 vs. 0.0509) and almost no difference at 50% and 75% on PhysioNet or at 25% and 50% on SEED-IV, likely because the channel-centering step Π(Z)\Pi(Z) in Eqs. 4 and 9 already subtracts the unweighted channel mean at every iteration. Also, in the frozen classifier test at 25% missing channels (Table 3), RIGS-EEG only beats classic, parameter-free Spherical Splines (SS) by 0.39 to 0.55 percentage points in balanced accuracy, and beats GH and CT by 0.24 to 0.35 points. These gains are much smaller than the standard deviations (±0.86%\pm 0.86\% to ±2.43%\pm 2.43\%), and downstream classification is not tested at 50% or 75% missingness, where simple interpolation breaks down. Finally, all reconstruction models are trained with a single random seed (42), and no statistical significance tests are reported (there are minor notation typos, such as a broken Eq. (??) reference on Line 168).
Questions for authors
  1. Why use different training setups ("v1" vs. "v2") across datasets? Right now, PhysioNet is trained with the "v1" setup (qc=1q_c = 1, standard MSE, plus extra spectral and spatial losses), while SEED-IV and MCD use the "v2" setup (reliability weighting turned on, channel-normalized RMSE, and no extra losses). Could you share the exact equations and weights for those extra v1 losses on PhysioNet? Also, what happens if you use a single consistent setup (for example, v2 with reliability weighting turned on) across all three datasets?

  2. Section 4.4 notes that on PhysioNet, the w/o graph and w/o spectral ablations also remove their matching auxiliary training losses at the same time. To show how much of the drop comes from the model architecture versus the loss function, what are the NRMSE scores on PhysioNet when you keep the full RIGS-EEG model intact and only remove those extra training losses?

  3. What happens to the frozen classifier at 50% and 75% missing channels (Table 3)? At 25% missingness, classic Spherical Splines (SS) gets within 0.39 to 0.55 percentage points of RIGS-EEG on the frozen EEGNet classifier (Table 3), whereas in Table 2, SS breaks down much more than RIGS-EEG at 50% and 75% missingness. Since the frozen classifier is already trained, evaluating it on the 50% and 75% reconstructed test sets requires no retraining. Could you share the Table 3 results at 50% and 75% missingness?

  4. Removing the reference module (w/o reference) makes almost no difference on PhysioNet or on SEED-IV at 25% and 50% missingness in Table 4. Could you provide paired statistical significance tests (such as Wilcoxon signed-rank across test subjects) for the main NRMSE results in Tables 1 and 2, as well as a quick test-time check showing what happens when an additive reference shift is applied to the test inputs? Also, how would you adapt this approach for bipolar montages (differences between adjacent electrodes), where reference changes are not just a single global wave added to all channels?

8 comments

First, thank you for your recognition of our paper and for your valuable feedback. We also thank you for your expert guidance on our work in this field. I will now answer each of your questions in turn.

Regarding your questions

[1] We disabled reliability weighting on PhysioNet because our previous experimental results showed that the unweighted and weighted results were approximately equal at missing rates of 25%, 50%, and 75%, with no clear difference, whereas the weighting module made a considerable contribution on SEED-IV and MCD. This demonstrates that the effects of the weighting module vary across datasets. We therefore ultimately decided to set qc=1q_c = 1 on PhysioNet.

[2] Thank you for pointing out this attribution issue. We have added a controlled experiment that changes only the training objective: we retained the complete RIGS-EEG architecture and the original reliability setting (qc=1q_c = 1), set the weights of both the spectral and spatial auxiliary losses to 0, and used the same data splits, preprocessing, training rules, and validation-based selection criteria. The NRMSE values of the MSE-only model at missing rates of 25%, 50%, and 75% were 0.051128±0.000212, 0.057514±0.000198, and 0.072501±0.000144, respectively. These values are the mean ± sample standard deviation across three independent training runs with seeds 42, 43, and 44. Compared with the original configuration, the mean NRMSE increased by approximately 0.22% at 25% missingness and remained essentially unchanged at 50% and 75% missingness. This indicates that, under the current evaluation setting, the two auxiliary losses make only a limited additional contribution to the mean NRMSE, and that training the complete architecture with only the MSE objective can maintain reconstruction performance close to that of the original configuration. We also explicitly describe the original module-removal ablations as joint ablations of the architecture and the associated losses.

[3] We have added the experiments for Table 3 at missing rates of 50% and 75%, and have finished compiling these results:

Table 3 extension: Frozen-classifier results on PhysioNet at 50% channel missingness (mean ± SD).

MethodBalanced accuracy (%) MM-HBalanced accuracy (%) MM-HFBalanced accuracy (%) MI-HBalanced accuracy (%) MI-HFCohen’s κ MM-HCohen’s κ MM-HFCohen’s κ MI-HCohen’s κ MI-HF
Complete EEG74.54 ± 1.5549.36 ± 1.7675.38 ± 0.3254.89 ± 3.320.492 ± 0.0310.326 ± 0.0240.507 ± 0.0060.397 ± 0.044
Corrupted EEG62.86 ± 0.7934.29 ± 1.6367.29 ± 1.0538.59 ± 1.650.257 ± 0.0160.124 ± 0.0220.345 ± 0.0210.181 ± 0.022
SS72.12 ± 1.1745.53 ± 1.7472.74 ± 0.9650.16 ± 2.910.447 ± 0.0240.268 ± 0.0240.454 ± 0.0190.321 ± 0.039
IDW71.34 ± 1.9244.47 ± 2.3272.41 ± 0.9048.38 ± 2.490.427 ± 0.0390.260 ± 0.0310.448 ± 0.0180.311 ± 0.033
KRIGING71.52 ± 1.5144.19 ± 1.4772.43 ± 0.3949.10 ± 3.040.431 ± 0.0300.257 ± 0.0200.448 ± 0.0080.320 ± 0.041
STC68.71 ± 1.8940.70 ± 2.1170.80 ± 1.4144.93 ± 1.960.374 ± 0.0380.210 ± 0.0290.415 ± 0.0280.265 ± 0.026
STDC62.86 ± 0.7934.29 ± 1.6367.29 ± 1.0538.59 ± 1.650.257 ± 0.0160.124 ± 0.0220.345 ± 0.0210.181 ± 0.022
HALRTC62.86 ± 0.7934.29 ± 1.6367.29 ± 1.0538.59 ± 1.650.257 ± 0.0160.124 ± 0.0220.345 ± 0.0210.181 ± 0.022
GLAE71.38 ± 1.3644.44 ± 2.6772.22 ± 1.2048.99 ± 2.870.428 ± 0.0280.260 ± 0.0360.444 ± 0.0230.319 ± 0.038
CT64.22 ± 0.3934.42 ± 1.4367.92 ± 1.1839.64 ± 2.030.284 ± 0.0080.126 ± 0.0190.358 ± 0.0240.195 ± 0.027
SAITS68.70 ± 1.2645.08 ± 1.2071.20 ± 1.9247.38 ± 2.110.374 ± 0.0250.269 ± 0.0160.423 ± 0.0390.297 ± 0.029
CSDI71.29 ± 1.2644.93 ± 2.0772.05 ± 1.1349.13 ± 2.350.426 ± 0.0250.267 ± 0.0280.440 ± 0.0230.321 ± 0.031
PATCHTST64.26 ± 0.5634.52 ± 1.5168.12 ± 1.1739.62 ± 1.830.285 ± 0.0110.127 ± 0.0200.362 ± 0.0230.194 ± 0.024
GH71.73 ± 1.9744.96 ± 2.1372.74 ± 1.0749.42 ± 2.670.435 ± 0.0400.260 ± 0.0290.454 ± 0.0220.318 ± 0.036
SRL-ITV69.79 ± 1.5042.80 ± 2.0971.12 ± 1.4647.30 ± 1.850.396 ± 0.0300.238 ± 0.0280.422 ± 0.0290.296 ± 0.025
DMAE-EEG69.97 ± 1.1942.78 ± 1.7971.48 ± 1.9446.95 ± 1.640.400 ± 0.0240.238 ± 0.0240.429 ± 0.0380.292 ± 0.022
RIGS-EEG72.43 ± 1.6245.89 ± 2.3673.63 ± 1.0550.43 ± 2.700.450 ± 0.0330.279 ± 0.0320.472 ± 0.0210.338 ± 0.036

Table 3 extension: Frozen-classifier results on PhysioNet at 75% channel missingness (mean ± SD).

MethodBalanced accuracy (%) MM-HBalanced accuracy (%) MM-HFBalanced accuracy (%) MI-HBalanced accuracy (%) MI-HFCohen’s κ MM-HCohen’s κ MM-HFCohen’s κ MI-HCohen’s κ MI-HF
Complete EEG74.54 ± 1.5549.36 ± 1.7675.38 ± 0.3254.89 ± 3.320.492 ± 0.0310.326 ± 0.0240.507 ± 0.0060.397 ± 0.044
Corrupted EEG59.16 ± 2.4429.70 ± 1.6461.68 ± 2.0132.65 ± 0.800.183 ± 0.0490.063 ± 0.0220.233 ± 0.0400.102 ± 0.011
SS66.97 ± 1.2939.55 ± 1.4368.57 ± 0.9542.47 ± 2.190.350 ± 0.0250.195 ± 0.0190.381 ± 0.0190.242 ± 0.029
IDW66.84 ± 2.6039.34 ± 1.6369.46 ± 1.4042.86 ± 1.470.337 ± 0.0520.192 ± 0.0220.389 ± 0.0280.238 ± 0.020
KRIGING67.37 ± 1.2040.11 ± 1.9167.07 ± 0.8541.23 ± 1.720.318 ± 0.0240.186 ± 0.0250.361 ± 0.0170.216 ± 0.023
STC62.33 ± 3.0434.83 ± 3.1266.70 ± 2.2238.47 ± 1.250.246 ± 0.0610.132 ± 0.0420.334 ± 0.0440.179 ± 0.016
STDC59.16 ± 2.4429.70 ± 1.6461.68 ± 2.0132.65 ± 0.800.183 ± 0.0490.063 ± 0.0220.233 ± 0.0400.102 ± 0.011
HALRTC59.16 ± 2.4429.70 ± 1.6461.68 ± 2.0132.65 ± 0.800.183 ± 0.0490.063 ± 0.0220.233 ± 0.0400.102 ± 0.011
GLAE62.94 ± 1.4134.96 ± 1.9565.53 ± 1.9337.61 ± 1.630.259 ± 0.0280.133 ± 0.0260.310 ± 0.0390.168 ± 0.022
CT61.08 ± 1.6329.95 ± 1.7363.45 ± 2.4233.00 ± 1.190.222 ± 0.0320.066 ± 0.0230.268 ± 0.0480.107 ± 0.016
SAITS57.22 ± 1.9733.89 ± 1.9058.57 ± 4.1132.95 ± 1.600.144 ± 0.0400.119 ± 0.0260.171 ± 0.0820.106 ± 0.021
CSDI64.30 ± 1.9139.99 ± 1.2767.81 ± 1.4342.01 ± 0.570.286 ± 0.0390.201 ± 0.0170.356 ± 0.0290.226 ± 0.008
PATCHTST60.90 ± 1.7029.93 ± 1.7663.35 ± 2.1832.95 ± 1.070.218 ± 0.0340.066 ± 0.0240.266 ± 0.0440.106 ± 0.014
GH66.78 ± 2.3939.08 ± 1.6468.30 ± 0.9142.79 ± 1.570.346 ± 0.0480.185 ± 0.0220.386 ± 0.0180.233 ± 0.021
SRL-ITV62.52 ± 1.3032.65 ± 1.6564.46 ± 2.6135.30 ± 1.050.250 ± 0.0260.103 ± 0.0220.288 ± 0.0520.137 ± 0.014
DMAE-EEG61.03 ± 2.0436.27 ± 1.0264.82 ± 1.1438.32 ± 1.060.221 ± 0.0410.151 ± 0.0140.296 ± 0.0230.177 ± 0.014
RIGS-EEG67.97 ± 2.3840.28 ± 1.7969.99 ± 1.5343.80 ± 1.410.359 ± 0.0480.205 ± 0.0240.399 ± 0.0310.250 ± 0.019

As shown, RIGS continues to achieve the highest performance, although the performance of all methods declines as the missing rate increases.

[4] The current theory concerns common additive reference changes and cannot be directly extended to bipolar montages. For bipolar observations Z=BXZ = BX, one potential extension is to reformulate observation consistency as consistency between BX^B\widehat{X} and the observed differences, while addressing the reference degree of freedom and identifiability involved in recovering potentials. This extension has not yet been validated, and we will explicitly list it as a limitation of applicability.

Regarding the weaknesses

[1] This weakness overlaps with your question [1], so I will not elaborate further here.

[2] We agree that the terminology and scope of applicability in the original manuscript need to be more precise. What the paper proves is the following: when the mask is fixed and the reliability weights remain invariant to a common shift, adding a common reference waveform to the input adds the same waveform to the output. We will consistently use the term “additive common-reference equivariance” and explicitly state that this property does not provide a guarantee under bipolar montages or general linear re-referencing.

The cross-reference issues you identified do indeed need to be corrected. We sincerely apologize for the adverse effect these errors had on your reading. They resulted from incomplete updates to the cross-references in Overleaf and our failure to add labels to the equations. The references should be corrected as follows: “Eq. (??)” below Eq. (2) in Section 3.2 should point to the definition of the reliability weights in Eq. (2) on the same page; “Appendix C” in the parameter statistics in Section 3.4 should point to “Appendix D”.

I have added the relevant experiments with additional random seeds. The detailed results are shown below (the experiments in the original paper PDF correspond to random seed 42; only the added experiments with random seeds 43 and 44 are shown here):

Reconstruction results with random seed 43

MethodPhysioNet 25%PhysioNet 50%PhysioNet 75%SEED-IV 25%SEED-IV 50%SEED-IV 75%MCD 25%MCD 50%MCD 75%
SS0.0545 ± 0.01510.0614 ± 0.01690.0792 ± 0.02100.4177 ± 1.25980.4938 ± 1.50380.6734 ± 2.07200.1001 ± 0.02380.1114 ± 0.02760.1417 ± 0.0351
IDW0.0687 ± 0.00860.0711 ± 0.00870.0765 ± 0.00920.2864 ± 0.64570.2948 ± 0.66010.3051 ± 0.65470.0873 ± 0.01770.0891 ± 0.01840.0948 ± 0.0194
KRIGING0.0789 ± 0.02520.0900 ± 0.02910.0862 ± 0.02451.2403 ± 4.27041.3902 ± 4.75751.0059 ± 3.28610.1468 ± 0.03870.1658 ± 0.04440.1501 ± 0.0378
STC0.0927 ± 0.00930.0936 ± 0.00940.0951 ± 0.00960.3291 ± 0.74480.3332 ± 0.74680.3414 ± 0.74370.0956 ± 0.01590.0962 ± 0.01640.1083 ± 0.0170
STDC0.0898 ± 0.00900.0955 ± 0.00930.1030 ± 0.01020.3231 ± 0.62750.3528 ± 0.69420.3835 ± 0.67800.0908 ± 0.01760.0919 ± 0.01790.0942 ± 0.0187
HALRTC0.0893 ± 0.00900.0952 ± 0.00950.1025 ± 0.01030.3238 ± 0.63500.3532 ± 0.67560.3819 ± 0.64500.0897 ± 0.01780.0907 ± 0.01760.0939 ± 0.0182
GLAE0.0520 ± 0.01190.0596 ± 0.01110.0955 ± 0.00790.4155 ± 1.27780.3161 ± 0.86970.2310 ± 0.45360.0953 ± 0.01640.0942 ± 0.01090.1198 ± 0.0142
CT0.0552 ± 0.01540.0691 ± 0.01050.0825 ± 0.01690.1137 ± 0.09150.1320 ± 0.48950.1460 ± 0.12110.0896 ± 0.01770.0905 ± 0.01790.0938 ± 0.0188
SAITS0.0524 ± 0.01160.0636 ± 0.01130.1064 ± 0.02110.1384 ± 0.23890.1391 ± 0.17640.1556 ± 0.10520.0929 ± 0.01760.1017 ± 0.01860.1433 ± 0.0261
CSDI0.0659 ± 0.01030.0725 ± 0.01010.0881 ± 0.01130.1097 ± 0.02710.1165 ± 0.02860.1365 ± 0.03450.1011 ± 0.01980.1301 ± 0.02180.2052 ± 0.0315
PATCHTST0.0592 ± 0.01750.0697 ± 0.02600.0990 ± 0.01510.1145 ± 0.03660.1344 ± 0.02460.1569 ± 0.03970.0896 ± 0.01740.0905 ± 0.01800.0941 ± 0.0182
GH0.0547 ± 0.01150.0603 ± 0.01110.0785 ± 0.01010.2373 ± 0.58850.2632 ± 0.65320.2858 ± 0.65560.0891 ± 0.01950.0920 ± 0.02040.0965 ± 0.0211
SRL-ITV0.0557 ± 0.00950.0704 ± 0.00740.1166 ± 0.00610.0882 ± 0.02150.1007 ± 0.03820.1290 ± 0.04850.0878 ± 0.01460.0993 ± 0.00960.1389 ± 0.0108
DMAE-EEG0.0732 ± 0.01020.0762 ± 0.01040.0934 ± 0.01440.1638 ± 0.11270.1620 ± 0.10210.1770 ± 0.07010.0880 ± 0.01700.0891 ± 0.01640.1029 ± 0.0135
RIGS-EEG0.0513 ± 0.01200.0578 ± 0.01120.0725 ± 0.01030.0858 ± 0.02810.0985 ± 0.03350.1270 ± 0.04400.0861 ± 0.01770.0883 ± 0.01850.0921 ± 0.0192
Reviewer_9LjmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies whole-channel EEG missingness, where an electrode has no samples in the analysis window, so reconstruction must borrow from other electrodes under coupled effects of common reference, spatial coverage, and observation quality. The paper proposes RIGS-EEG, a compact reconstruction pipeline that first separates a common reference from relative potentials, builds a reliability-weighted spatial estimate, and then refines missing channels with five graph–time–frequency steps, while keeping the actually measured electrode values unchanged in the final output. The paper reports that RIGS-EEG effectively recovers fully missing electrode waveforms under 75% channel loss on PhysioNet, SEED-IV, and MCD (NRMSE 0.0723 / 0.1274 / 0.0922), outperforming 14 baselines while using only 970 trainable parameters.

Critical strengths

• Reconstruction performance is strong under severe channel loss, since at 75% missing electrodes RIGS-EEG reports the lowest mean NRMSE on PhysioNet, SEED-IV, and MCD against 14 baselines. • The method is highly parameter-efficient, using only 970 trainable parameters while still outperforming much larger deep reconstruction baselines. • The results generalize across different evaluation protocols, holding under subject-wise splits on PhysioNet and MCD as well as a session-wise split on SEED-IV.

Critical weaknesses

• The reconstruction gains over the strongest baselines are sometimes small. At 75% missing electrodes, the reported advantages are only 1.62% and 1.71% on SEED-IV and MCD, so the practical margin is clearer on PhysioNet than on the other two datasets. • One of the three evaluation datasets, MCD, is in-house and non-public, which makes an important part of the empirical claim harder to verify independently. • The paper trains reconstruction models with a single seed and reports no paired significance test, so the mean NRMSE improvements are not shown to be statistically reliable under training variability.

Questions for authors

The reconstruction models appear to be trained with a single seed. Could the authors report results with two different seeds to increase the reliability of the NRMSE findings? Reliability weighting is disabled on PhysioNet ( q_c =1) but used on SEED-IV and MCD. Could the authors explain this choice and whether enabling it on PhysioNet changes the results. The method uses a symmetric 4-nearest-neighbor electrode graph for refinement. Could the authors explain why 4 neighbours were chosen, and whether the results change with a different neighbourhood size? In Section 3.2, the paper says SEED-IV and MCD use “Eq. (??)” for reliability weighting. Could the authors fix this broken reference.

3 comments

Thank you for your valuable comments on our work, which have helped us improve it. In particular, your questions about public access to the MCD dataset and additional random-seed experiments hit the nail on the head. I will now address each of your questions with complete sincerity:

[1] The reconstruction model was indeed evaluated with only a single random seed, which is insufficient to demonstrate its reliability. We therefore designate the original experiment as seed 42 and have added two experiments with random seeds 43 and 44. The results for seed 42 are in the original PDF; the results for seeds 43 and 44 are shown below:

Reconstruction results with random seed 43

MethodPhysioNet 25%PhysioNet 50%PhysioNet 75%SEED-IV 25%SEED-IV 50%SEED-IV 75%MCD 25%MCD 50%MCD 75%
SS0.0545 ± 0.01510.0614 ± 0.01690.0792 ± 0.02100.4177 ± 1.25980.4938 ± 1.50380.6734 ± 2.07200.1001 ± 0.02380.1114 ± 0.02760.1417 ± 0.0351
IDW0.0687 ± 0.00860.0711 ± 0.00870.0765 ± 0.00920.2864 ± 0.64570.2948 ± 0.66010.3051 ± 0.65470.0873 ± 0.01770.0891 ± 0.01840.0948 ± 0.0194
KRIGING0.0789 ± 0.02520.0900 ± 0.02910.0862 ± 0.02451.2403 ± 4.27041.3902 ± 4.75751.0059 ± 3.28610.1468 ± 0.03870.1658 ± 0.04440.1501 ± 0.0378
STC0.0927 ± 0.00930.0936 ± 0.00940.0951 ± 0.00960.3291 ± 0.74480.3332 ± 0.74680.3414 ± 0.74370.0956 ± 0.01590.0962 ± 0.01640.1083 ± 0.0170
STDC0.0898 ± 0.00900.0955 ± 0.00930.1030 ± 0.01020.3231 ± 0.62750.3528 ± 0.69420.3835 ± 0.67800.0908 ± 0.01760.0919 ± 0.01790.0942 ± 0.0187
HALRTC0.0893 ± 0.00900.0952 ± 0.00950.1025 ± 0.01030.3238 ± 0.63500.3532 ± 0.67560.3819 ± 0.64500.0897 ± 0.01780.0907 ± 0.01760.0939 ± 0.0182
GLAE0.0520 ± 0.01190.0596 ± 0.01110.0955 ± 0.00790.4155 ± 1.27780.3161 ± 0.86970.2310 ± 0.45360.0953 ± 0.01640.0942 ± 0.01090.1198 ± 0.0142
CT0.0552 ± 0.01540.0691 ± 0.01050.0825 ± 0.01690.1137 ± 0.09150.1320 ± 0.48950.1460 ± 0.12110.0896 ± 0.01770.0905 ± 0.01790.0938 ± 0.0188
SAITS0.0524 ± 0.01160.0636 ± 0.01130.1064 ± 0.02110.1384 ± 0.23890.1391 ± 0.17640.1556 ± 0.10520.0929 ± 0.01760.1017 ± 0.01860.1433 ± 0.0261
CSDI0.0659 ± 0.01030.0725 ± 0.01010.0881 ± 0.01130.1097 ± 0.02710.1165 ± 0.02860.1365 ± 0.03450.1011 ± 0.01980.1301 ± 0.02180.2052 ± 0.0315
PATCHTST0.0592 ± 0.01750.0697 ± 0.02600.0990 ± 0.01510.1145 ± 0.03660.1344 ± 0.02460.1569 ± 0.03970.0896 ± 0.01740.0905 ± 0.01800.0941 ± 0.0182
GH0.0547 ± 0.01150.0603 ± 0.01110.0785 ± 0.01010.2373 ± 0.58850.2632 ± 0.65320.2858 ± 0.65560.0891 ± 0.01950.0920 ± 0.02040.0965 ± 0.0211
SRL-ITV0.0557 ± 0.00950.0704 ± 0.00740.1166 ± 0.00610.0882 ± 0.02150.1007 ± 0.03820.1290 ± 0.04850.0878 ± 0.01460.0993 ± 0.00960.1389 ± 0.0108
DMAE-EEG0.0732 ± 0.01020.0762 ± 0.01040.0934 ± 0.01440.1638 ± 0.11270.1620 ± 0.10210.1770 ± 0.07010.0880 ± 0.01700.0891 ± 0.01640.1029 ± 0.0135
RIGS-EEG0.0513 ± 0.01200.0578 ± 0.01120.0725 ± 0.01030.0858 ± 0.02810.0985 ± 0.03350.1270 ± 0.04400.0861 ± 0.01770.0883 ± 0.01850.0921 ± 0.0192

Reconstruction results with random seed 44

MethodPhysioNet 25%PhysioNet 50%PhysioNet 75%SEED-IV 25%SEED-IV 50%SEED-IV 75%MCD 25%MCD 50%MCD 75%
SS0.0552 ± 0.01530.0613 ± 0.01650.0794 ± 0.02130.4171 ± 1.25910.4934 ± 1.50330.6732 ± 2.07210.1005 ± 0.02380.1113 ± 0.02670.1421 ± 0.0351
IDW0.0683 ± 0.00860.0704 ± 0.00810.0770 ± 0.00970.2861 ± 0.64500.2943 ± 0.65920.3047 ± 0.65400.0873 ± 0.01770.0892 ± 0.01720.0946 ± 0.0196
KRIGING0.0786 ± 0.02510.0892 ± 0.02910.0866 ± 0.02451.2400 ± 4.26981.3298 ± 4.75771.0060 ± 3.28580.1468 ± 0.03890.1659 ± 0.04340.1502 ± 0.0374
STC0.0925 ± 0.00930.0931 ± 0.00960.0957 ± 0.00980.3284 ± 0.74420.3324 ± 0.74600.3414 ± 0.74360.0957 ± 0.01590.0961 ± 0.01630.1081 ± 0.0171
STDC0.0895 ± 0.00960.0951 ± 0.00940.1035 ± 0.01040.3229 ± 0.62710.3522 ± 0.69340.3835 ± 0.67710.0907 ± 0.01760.0919 ± 0.01680.0940 ± 0.0186
HALRTC0.0890 ± 0.00920.0947 ± 0.00970.1031 ± 0.01020.3233 ± 0.63460.3530 ± 0.67470.3821 ± 0.64450.0898 ± 0.01740.0908 ± 0.01690.0935 ± 0.0181
GLAE0.0524 ± 0.01190.0590 ± 0.01110.0961 ± 0.00750.4152 ± 1.27730.3158 ± 0.86870.2312 ± 0.45300.0960 ± 0.01600.0942 ± 0.00980.1196 ± 0.0043
CT0.0548 ± 0.01540.0687 ± 0.01050.0830 ± 0.01690.1134 ± 0.09110.1315 ± 0.48880.1461 ± 0.12030.0899 ± 0.01750.0904 ± 0.01700.0934 ± 0.0189
SAITS0.0526 ± 0.01170.0632 ± 0.01160.1069 ± 0.02150.1383 ± 0.23840.1386 ± 0.17540.1558 ± 0.10470.0928 ± 0.01720.1016 ± 0.01750.1431 ± 0.0269
CSDI0.0656 ± 0.01040.0721 ± 0.01010.0886 ± 0.01140.1091 ± 0.02660.1159 ± 0.02790.1363 ± 0.03410.1011 ± 0.01970.1302 ± 0.02300.2050 ± 0.0314
PATCHTST0.0591 ± 0.01750.0693 ± 0.02640.0995 ± 0.01520.1142 ± 0.03640.1339 ± 0.02350.1569 ± 0.04040.0895 ± 0.01750.0905 ± 0.01890.0936 ± 0.0191
GH0.0545 ± 0.01150.0598 ± 0.01120.0792 ± 0.01190.2369 ± 0.58820.2628 ± 0.65240.2858 ± 0.65100.0890 ± 0.01990.0919 ± 0.02120.0963 ± 0.0212
SRL-ITV0.0556 ± 0.00950.0701 ± 0.00790.1172 ± 0.00680.0880 ± 0.02110.1002 ± 0.03780.1292 ± 0.04850.0879 ± 0.01440.0992 ± 0.01040.1386 ± 0.0038
DMAE-EEG0.0723 ± 0.01010.0759 ± 0.01030.0941 ± 0.01450.1636 ± 0.11220.1614 ± 0.10120.1771 ± 0.06980.0870 ± 0.01710.0891 ± 0.01720.1028 ± 0.0135
RIGS-EEG0.0509 ± 0.01190.0574 ± 0.01120.0727 ± 0.01040.0855 ± 0.02760.0980 ± 0.03260.1270 ± 0.04340.0861 ± 0.01780.0883 ± 0.01740.0920 ± 0.0191

[2] Among the three datasets, MCD was indeed not publicly available previously. As our latest update, however, we have now made MCD publicly available anonymously at https://huggingface.co/datasets/anonymous-588/MCD-EEG/tree/main. You can access this URL to view and download our dataset.

[3] We disabled reliability weighting on PhysioNet because our previous experimental results showed that the unweighted and weighted results were approximately equal at missing rates of 25%, 50%, and 75%, with no clear difference, whereas the weighting module made a considerable contribution on SEED-IV and MCD. This demonstrates that the effects of the weighting module vary across datasets. We therefore ultimately decided to set qc=1q_c = 1 on PhysioNet.

[4] We use a symmetric 4-nearest-neighbor graph to encode electrode geometry through sparse local connections. This design is primarily based on previous research (Hammond et al., 2011), but we had indeed not experimentally investigated the optimal value of k. We have therefore added the following experiments comparing k = 2, 4, and 8:

RIGS-EEG Neighborhood Size Sensitivity Experiment

NRMSE: mean ± sample standard deviation across training runs; seeds 42, 43, and 44, n = 3.

Neighbors kPhysioNet 25%PhysioNet 50%PhysioNet 75%SEED-IV 25%SEED-IV 50%SEED-IV 75%MCD 25%MCD 50%MCD 75%
20.054114±0.0000520.061380±0.0001310.076132±0.0002050.091077±0.0000730.105277±0.0002790.133567±0.0004980.087190±0.0000380.089159±0.0000490.092742±0.000065
40.051018±0.0002130.057517±0.0002560.072501±0.0002050.085740±0.0002070.098556±0.0005630.127131±0.0001990.086109±0.0000410.088355±0.0000270.092141±0.000100
80.050867±0.0000800.056716±0.0000430.071011±0.0001870.086063±0.0001060.097341±0.0000720.123812±0.0000810.086380±0.0000400.088569±0.0001180.092345±0.000177

At high missing rates, k = 8 does perform slightly better than k = 4 on PhysioNet and SEED-IV, but slightly worse on MCD. Their practical performance is therefore similar, and both do indeed outperform k = 2.

[5] The cross-reference issues you pointed out do indeed need to be corrected. We are very sorry that our mistakes in this regard adversely affected your reading experience. These arose because we did not fully update the references or add labels to the equations when using Overleaf for typesetting. The relevant references should be corrected as follows: “Eq. (??)” below Eq. (2) in Section 3.2 should refer to the definition of the reliability weights in Eq. (2) on the same page; “Appendix C” in the parameter statistics in Section 3.4 should refer to “Appendix D”.

Finally, thank you very much for taking time out of your busy schedule to read our responses. If you have any questions after reading them, we hope you will raise them promptly, and we will take them seriously. If possible, we sincerely hope that you can raise the score for our paper, for which we would be most grateful.

Reviewer_NqtdOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents RIGS-EEG, a method for reconstructing missing EEG channels. The pipeline separates a common reference component, initializes missing electrodes through reliability-weighted spatial interpolation, performs five unrolled graph and time–frequency refinement steps, and finally applies a hard projection that preserves all observed samples. The learned component contains only 970 trainable parameters. The method is evaluated against 14 baselines on PhysioNet, SEED-IV, and a private MCD dataset at three missing-channel rates. It achieves the lowest reported mean NRMSE across all nine dataset–missingness conditions among the compared methods. The paper also evaluates downstream utility using frozen classifiers trained on complete PhysioNet signals.

The architecture is compact, interpretable, and appropriately structured around electrode geometry and observation consistency. However, the paper’s contribution appears specialized and methodologically incremental. The “reference-invariant” claim is broader than the property actually established, and the evaluation relies on artificially removed clean channels rather than realistic electrode failures. Ambiguities in the method description, insufficiently controlled ablations, and dataset-specific training variants further weaken the empirical conclusions.

Critical strengths
  • The model combines explicit structural priors with learning while using only 970 trainable parameters. Its components have interpretable roles tied to EEG geometry, spectral structure, reliability, and data consistency.

  • The paper evaluates three datasets, three missingness levels, and 14 baselines. RIGS-EEG attains the best reported mean reconstruction result in every tested condition.

  • The paper attempts to formalize its behavior under an additive waveform shared across channels rather than relying solely on an informal robustness claim.

Critical weaknesses
  1. The contribution is specialized and methodologically incremental

The proposed system combines established ideas: spatial interpolation, graph representations, spectral processing, time–frequency shrinkage, reliability weighting, observation projection, and finite algorithm unrolling. The integration is sensible, but the paper does not clearly identify a sufficiently general methodological advance beyond EEG montage reconstruction.

The strongest potential conceptual contribution is the explicit handling of a common additive reference component. However, as discussed below, the corresponding property is narrower than the paper’s terminology suggests. The experiments also do not demonstrate that the method provides a general reconstruction principle that transfers beyond the tested EEG configurations.

  1. “Reference-invariant” overstates the proved property

The derivation appears to establish equivariance to adding the same waveform to all visible channels, assuming that the mask and reliability weights remain fixed or are themselves invariant to this shift. Under this transformation, the reconstructed output shifts by the same waveform. This is equivariance, not invariance.

The result does not establish robustness to general EEG re-referencing operations, such as bipolar derivations, linked references, average-reference changes, or REST. The title and claims should therefore use a more precise term, such as “additive common-reference equivariant.” A direct numerical test of the claimed equivariance would also be useful.

  1. The initial reference subtraction appears algebraically redundant

Under the equations as written, the initial common-reference subtraction has no independent effect on the initialization because the subsequent across-channel centering removes it.

  1. Artificial missingness is a weak proxy for practical electrode failure

The evaluation removes known clean channels from otherwise complete recordings. Real EEG electrode failures are often spatially clustered, intermittent, saturated, contaminated by high-amplitude noise, or accompanied by corruption in channels that remain nominally observed.

  1. Dataset-specific method variants weaken the cross-dataset comparison

The three-dataset evaluation does not appear to use a single fixed training recipe. PhysioNet uses uniform reliability weights and an MSE-based objective with spectral and spatial auxiliary losses, whereas SEED-IV and MCD use reliability weighting and a channel-normalized RMSE objective.

Dataset-specific configurations may be reasonable, but the paper does not adequately explain why these variants were selected or whether test-set performance influenced these choices. This makes it difficult to determine whether the results demonstrate intrinsic robustness or successful dataset-specific adaptation.

  1. The ablation study does not isolate individual contributions

The ablation definitions are too broad to map unambiguously onto specific equations or computational blocks. Several comparisons modify more than one factor simultaneously.

  1. The evaluation does not establish broad applicability

The tested datasets contain 64, 62, and 60 channels. These are similar high-density configurations for a method whose behavior depends directly on electrode geometry, nearest-neighbor graphs, and spatial kernels. Increasing the missing-channel percentage within these layouts does not substitute for evaluation on natively sparse or substantially denser montages.

The downstream evaluation is narrower still: classification is reported only for PhysioNet at 25% missingness. Although multiple motor tasks are included, they share the same acquisition setting and missingness protocol.

The reconstruction study should include different native channel counts and layouts. The frozen-classifier evaluation should cover 50% and 75% missingness and, ideally, another dataset. This would clarify whether the reconstructed signals retain useful information under the more difficult conditions emphasized in the reconstruction experiments.

  1. The method section requires clearer organization and editing

Section 3 contains at least one unresolved reference—“SEED-IV and MCD enables Eq. (??)”—and introduces some definitions before their computational roles are clear.

Questions for authors
  • Can every ablation be repeated while changing only one architectural or training factor?

  • How sensitive is performance to the unrolling depth, graph construction, reliability thresholds, and auxiliary-loss weights?

  • Can frozen-classifier performance be reported at 50% and 75% missingness?

  • How well does the method transfer to natively low-density or substantially different electrode layouts?

3 comments

Thank you for your valuable comments on our paper. Your distinctive insights into this field have helped guide improvements to our paper. We acknowledge some of the issues and are making corrections, while some misunderstandings may require further discussion. Next, I will answer all of your questions one by one:

[1] I guarantee that every ablation is reproducible and changes only one architectural or training factor at a time. This is indeed what we did.

[2] The 4-nearest-neighbor graph provides a sparse, local prior on electrode geometry for graph-spectral refinement. We construct the graph by selecting nearest neighbors based on three-dimensional electrode distances and then taking the union of the edges to obtain a symmetric graph, so the final node degree may exceed 4. This question coincides with the one raised by Reviewer 9Ljm; please see my response to their Question [4] for the specific results. Dense Gaussian initialization can use all electrodes that remain observed, so an initial estimate can still be obtained even when all 4 nearest neighbors are missing; the local graph defines the spectral coordinates for subsequent refinement. Changing k changes the graph edges, weights, and spectral structure, which may affect the results. The graph-removal ablation in the original Table 4 supports the role of the graph component and can indirectly reflect its sensitivity.

[3] Yes, these results have now been compiled:

Table 3 extension: Frozen-classifier results on PhysioNet at 50% channel missingness (mean ± SD).

MethodBalanced accuracy (%) MM-HBalanced accuracy (%) MM-HFBalanced accuracy (%) MI-HBalanced accuracy (%) MI-HFCohen’s κ MM-HCohen’s κ MM-HFCohen’s κ MI-HCohen’s κ MI-HF
Complete EEG74.54 ± 1.5549.36 ± 1.7675.38 ± 0.3254.89 ± 3.320.492 ± 0.0310.326 ± 0.0240.507 ± 0.0060.397 ± 0.044
Corrupted EEG62.86 ± 0.7934.29 ± 1.6367.29 ± 1.0538.59 ± 1.650.257 ± 0.0160.124 ± 0.0220.345 ± 0.0210.181 ± 0.022
SS72.12 ± 1.1745.53 ± 1.7472.74 ± 0.9650.16 ± 2.910.447 ± 0.0240.268 ± 0.0240.454 ± 0.0190.321 ± 0.039
IDW71.34 ± 1.9244.47 ± 2.3272.41 ± 0.9048.38 ± 2.490.427 ± 0.0390.260 ± 0.0310.448 ± 0.0180.311 ± 0.033
KRIGING71.52 ± 1.5144.19 ± 1.4772.43 ± 0.3949.10 ± 3.040.431 ± 0.0300.257 ± 0.0200.448 ± 0.0080.320 ± 0.041
STC68.71 ± 1.8940.70 ± 2.1170.80 ± 1.4144.93 ± 1.960.374 ± 0.0380.210 ± 0.0290.415 ± 0.0280.265 ± 0.026
STDC62.86 ± 0.7934.29 ± 1.6367.29 ± 1.0538.59 ± 1.650.257 ± 0.0160.124 ± 0.0220.345 ± 0.0210.181 ± 0.022
HALRTC62.86 ± 0.7934.29 ± 1.6367.29 ± 1.0538.59 ± 1.650.257 ± 0.0160.124 ± 0.0220.345 ± 0.0210.181 ± 0.022
GLAE71.38 ± 1.3644.44 ± 2.6772.22 ± 1.2048.99 ± 2.870.428 ± 0.0280.260 ± 0.0360.444 ± 0.0230.319 ± 0.038
CT64.22 ± 0.3934.42 ± 1.4367.92 ± 1.1839.64 ± 2.030.284 ± 0.0080.126 ± 0.0190.358 ± 0.0240.195 ± 0.027
SAITS68.70 ± 1.2645.08 ± 1.2071.20 ± 1.9247.38 ± 2.110.374 ± 0.0250.269 ± 0.0160.423 ± 0.0390.297 ± 0.029
CSDI71.29 ± 1.2644.93 ± 2.0772.05 ± 1.1349.13 ± 2.350.426 ± 0.0250.267 ± 0.0280.440 ± 0.0230.321 ± 0.031
PATCHTST64.26 ± 0.5634.52 ± 1.5168.12 ± 1.1739.62 ± 1.830.285 ± 0.0110.127 ± 0.0200.362 ± 0.0230.194 ± 0.024
GH71.73 ± 1.9744.96 ± 2.1372.74 ± 1.0749.42 ± 2.670.435 ± 0.0400.260 ± 0.0290.454 ± 0.0220.318 ± 0.036
SRL-ITV69.79 ± 1.5042.80 ± 2.0971.12 ± 1.4647.30 ± 1.850.396 ± 0.0300.238 ± 0.0280.422 ± 0.0290.296 ± 0.025
DMAE-EEG69.97 ± 1.1942.78 ± 1.7971.48 ± 1.9446.95 ± 1.640.400 ± 0.0240.238 ± 0.0240.429 ± 0.0380.292 ± 0.022
RIGS-EEG72.43 ± 1.6245.89 ± 2.3673.63 ± 1.0550.43 ± 2.700.450 ± 0.0330.279 ± 0.0320.472 ± 0.0210.338 ± 0.036

Table 3 extension: Frozen-classifier results on PhysioNet at 75% channel missingness (mean ± SD).

MethodBalanced accuracy (%) MM-HBalanced accuracy (%) MM-HFBalanced accuracy (%) MI-HBalanced accuracy (%) MI-HFCohen’s κ MM-HCohen’s κ MM-HFCohen’s κ MI-HCohen’s κ MI-HF
Complete EEG74.54 ± 1.5549.36 ± 1.7675.38 ± 0.3254.89 ± 3.320.492 ± 0.0310.326 ± 0.0240.507 ± 0.0060.397 ± 0.044
Corrupted EEG59.16 ± 2.4429.70 ± 1.6461.68 ± 2.0132.65 ± 0.800.183 ± 0.0490.063 ± 0.0220.233 ± 0.0400.102 ± 0.011
SS66.97 ± 1.2939.55 ± 1.4368.57 ± 0.9542.47 ± 2.190.350 ± 0.0250.195 ± 0.0190.381 ± 0.0190.242 ± 0.029
IDW66.84 ± 2.6039.34 ± 1.6369.46 ± 1.4042.86 ± 1.470.337 ± 0.0520.192 ± 0.0220.389 ± 0.0280.238 ± 0.020
KRIGING67.37 ± 1.2040.11 ± 1.9167.07 ± 0.8541.23 ± 1.720.318 ± 0.0240.186 ± 0.0250.361 ± 0.0170.216 ± 0.023
STC62.33 ± 3.0434.83 ± 3.1266.70 ± 2.2238.47 ± 1.250.246 ± 0.0610.132 ± 0.0420.334 ± 0.0440.179 ± 0.016
STDC59.16 ± 2.4429.70 ± 1.6461.68 ± 2.0132.65 ± 0.800.183 ± 0.0490.063 ± 0.0220.233 ± 0.0400.102 ± 0.011
HALRTC59.16 ± 2.4429.70 ± 1.6461.68 ± 2.0132.65 ± 0.800.183 ± 0.0490.063 ± 0.0220.233 ± 0.0400.102 ± 0.011
GLAE62.94 ± 1.4134.96 ± 1.9565.53 ± 1.9337.61 ± 1.630.259 ± 0.0280.133 ± 0.0260.310 ± 0.0390.168 ± 0.022
CT61.08 ± 1.6329.95 ± 1.7363.45 ± 2.4233.00 ± 1.190.222 ± 0.0320.066 ± 0.0230.268 ± 0.0480.107 ± 0.016
SAITS57.22 ± 1.9733.89 ± 1.9058.57 ± 4.1132.95 ± 1.600.144 ± 0.0400.119 ± 0.0260.171 ± 0.0820.106 ± 0.021
CSDI64.30 ± 1.9139.99 ± 1.2767.81 ± 1.4342.01 ± 0.570.286 ± 0.0390.201 ± 0.0170.356 ± 0.0290.226 ± 0.008
PATCHTST60.90 ± 1.7029.93 ± 1.7663.35 ± 2.1832.95 ± 1.070.218 ± 0.0340.066 ± 0.0240.266 ± 0.0440.106 ± 0.014
GH66.78 ± 2.3939.08 ± 1.6468.30 ± 0.9142.79 ± 1.570.346 ± 0.0480.185 ± 0.0220.386 ± 0.0180.233 ± 0.021
SRL-ITV62.52 ± 1.3032.65 ± 1.6564.46 ± 2.6135.30 ± 1.050.250 ± 0.0260.103 ± 0.0220.288 ± 0.0520.137 ± 0.014
DMAE-EEG61.03 ± 2.0436.27 ± 1.0264.82 ± 1.1438.32 ± 1.060.221 ± 0.0410.151 ± 0.0140.296 ± 0.0230.177 ± 0.014
RIGS-EEG67.97 ± 2.3840.28 ± 1.7969.99 ± 1.5343.80 ± 1.410.359 ± 0.0480.205 ± 0.0240.399 ± 0.0310.250 ± 0.019

As can be seen, RIGS still achieves the highest performance, although all methods perform worse as the missingness rate increases.

[4] Your question indeed gets straight to the heart of the issue. We are sorry that we do not have enough time to add a native low-density dataset as an experimental sample to substantiate our method. However, the electrode distribution in SEED-IV actually differs from those in the other two datasets. Moreover, our missingness setting involves random electrode loss; although this is not a native low-density configuration, it can simulate a low-density electrode distribution. These can all be seen in Figure 3 of Appendix A in our paper.

Next, I would like to offer some explanations regarding the weaknesses you identified in our paper

[1] As you note, the experiments do not establish a general reconstruction principle that extends beyond the EEG configurations tested. However, our main contribution is precisely a methodological breakthrough in EEG reconstruction; we have not claimed a contribution to general-purpose signal reconstruction.

[2] Thank you for pointing out this distinction in terminology. We agree that the precise property satisfied by the model output is "additive common-reference equivariance," while the internal relative-potential representation has the corresponding invariance. We will consider revising the title, abstract, and main text consistently and explicitly stating the conditions under which this conclusion holds, to avoid generalizing it into a guarantee for arbitrary re-referencing operations.

[3] This may be redundant at the level of the equations. In practical processing, however, we subtract the common reference to separate the observed data into a common-reference component and a feature component, process them differently, and then combine them again. The purpose is to control variables.

[4] Indeed, this evaluation removes known clean channels from recordings that should otherwise be complete. However, this also happens to answer your Question [4]: at high missingness rates, its effect is equivalent to a native low-density electrode distribution.

[5] We acknowledge that we indeed failed to explain these variants adequately. This was because we took the widespread use of the PhysioNet dataset in this field for granted. The dataset exhibits inherent cleanliness, meaning that its noise density is extremely low, which is also why we do not apply weighting. Your question closely parallels the one raised by Reviewer 9Ljm; please see my response to their Question [3].

[6] We are sorry that, due to space limitations, we did not provide the specific formulas for the ablation experiments. However, we guarantee that all of our ablation experiments ablate core modules, and that each experiment removes only the corresponding core module; no experiment removes multiple modules at once.

[7] The weakness you identified is indeed an issue on our part, and it corresponds to both your Questions [3] and [4]. We have already acknowledged it and provided additional explanations above, so we will not elaborate further here.

[8] The cross-reference issues you pointed out do indeed need to be corrected. We are very sorry that our mistakes in this regard adversely affected your reading experience. These arose because we did not fully update the references or add labels to the equations when using Overleaf for typesetting. The relevant references should be corrected as follows: "Eq. (??)" below Eq. (2) in Section 3.2 should refer to the definition of the reliability weights in Eq. (2) on the same page; "Appendix C" in the parameter statistics in Section 3.4 should refer to "Appendix D".

Finally, thank you for your sincere suggestions on our work. Your insights into this field have indeed contributed to improving our work. We have therefore treated your comments with care and responded seriously. We hope that, after reading our responses, you will gain a deeper understanding of our work. If you have any questions, please feel free to ask. This work has not been easy. If possible, we hope you can raise our score.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of the Rebuttal

We sincerely thank all three reviewers(9Ljm, Yvc5 and Nqtd) for their careful reading and valuable comments. We have addressed the concerns individually, added experiments, and clarified the method, theoretical conditions, and scope. Your feedback has helped us strengthen the experimental support and presentation.

We summarize our responses to the three reviewers below:

  • Reviewer 9Ljm: (1) Reproducibility and data access:\color{blue}{\text{(1) Reproducibility and data access:}} We added reconstruction results for seeds 43 and 44, with the original paper corresponding to seed 42, and made the MCD dataset anonymously available. (2) Reliability and graph design:\color{blue}{\text{(2) Reliability and graph design:}} We explain why PhysioNet uses qc=1q_c=1, discuss the differing contribution of weighting across datasets, and provide comparisons for k = 2, 4, and 8. (3) Presentation:\color{blue}{\text{(3) Presentation:}} We identify the required equation and appendix cross-reference corrections.

  • Reviewer Nqtd: (1) Contribution and interpretation:\color{blue}{\text{(1) Contribution and interpretation:}} We clarify the contribution to EEG reconstruction and the distinction between output equivariance and invariance of the internal relative-potential representation. (2) Experimental design:\color{blue}{\text{(2) Experimental design:}} We explain the symmetric graph, dense initialization, and ablation settings, and add frozen-classifier evaluations at 50% and 75% channel missingness. (3) Scope:\color{blue}{\text{(3) Scope:}} We acknowledge the absence of additional native low-density dataset experiments and welcome further discussion of the method’s applicability.

  • Reviewer Yvc5: (1) Controlled loss ablation:\color{blue}{\text{(1) Controlled loss ablation:}} We retain the full architecture and original reliability setting while setting both auxiliary-loss weights to zero. The MSE-only NRMSE values are 0.051128 ± 0.000212,\color{green}{\text{The MSE-only NRMSE values are 0.051128 ± 0.000212,}} 0.057514 ± 0.000198, and 0.072501 ± 0.000144 at 25%, 50%,\color{green}{\text{0.057514 ± 0.000198, and 0.072501 ± 0.000144 at 25\%, 50\%,}} and 75% missingness, respectively.\color{green}{\text{and 75\% missingness, respectively.}} These are means ± sample SDs over three independent training runs with seeds 42, 43, and 44, indicating limited additional contributions from the auxiliary losses under the current evaluation. We identify the original module-removal experiments as joint ablations of architecture and associated losses. (2) Additional evidence and theoretical limits:\color{blue}{\text{(2) Additional evidence and theoretical limits:}} We add high-missingness classification and random-seed results and specify the conditions for additive common-reference equivariance, which does not guarantee bipolar or general linear rereferencing robustness.

Thank you again for your time and effort. We respectfully ask you to reconsider your scores in light of our responses. We expect to upload the revised PDF before November 16, 00:00 AoE. Further questions are welcome. We wish you every success in your work and all the best.

Best wishes,
The Authors

AuthorsOpenReview ↗

Updated Manuscript: Summary of Revisions and Additional Experiments

Dear Area Chair and Reviewers,

Thank you for the time and effort you have devoted to reviewing our submission and for your valuable feedback. We have revised and updated the latest PDF. To make the changes easier to identify, in the version with highlighted revisions, red indicates corrected or replaced content, including terminology changes and revised wording, while blue indicates newly added content, including additional experimental results and accompanying explanations. The main revisions are as follows:

  1. Title and abstract. We revised the title and abstract to more accurately reflect the scope of our contributions and avoid overclaiming.
  2. Terminology and wording. We updated several technical terms and refined the wording to make the methodological descriptions and discussion clearer and more precise.
  3. Additional experiments. We consolidated the additional experimental results provided in response to the reviewers' comments in Appendix E, so that they can be conveniently reviewed in one place.
  4. Research background and related work. In response to questions about the research background and its relationship to prior work, we warmly invite you to consult Appendix B, if time permits, for a more detailed discussion and explanation of the related literature.

We sincerely hope that you can find the time to read our responses and revised manuscript; this would mean a great deal to us. Even if the paper is ultimately not accepted, your feedback will still provide valuable guidance for improving this work. We have taken this research seriously throughout and invested substantial time and effort in it. We would be very grateful for any further comments or guidance you may be able to offer.

Thank you again for your time and effort. We wish you all the best in your research and work!

Best wishes,
The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.