- 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
-
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.
-
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
- 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.
- 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 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 ( to ), 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
-
Why use different training setups ("v1" vs. "v2") across datasets? Right now, PhysioNet is trained with the "v1" setup (, 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?
-
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?
-
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?
-
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
Table 3 extension: Frozen-classifier results on PhysioNet at 75% channel missingness (mean ± SD).
| Method | Balanced accuracy (%) MM-H | Balanced accuracy (%) MM-HF | Balanced accuracy (%) MI-H | Balanced accuracy (%) MI-HF | Cohen’s κ MM-H | Cohen’s κ MM-HF | Cohen’s κ MI-H | Cohen’s κ MI-HF |
|---|---|---|---|---|---|---|---|---|
| Complete EEG | 74.54 ± 1.55 | 49.36 ± 1.76 | 75.38 ± 0.32 | 54.89 ± 3.32 | 0.492 ± 0.031 | 0.326 ± 0.024 | 0.507 ± 0.006 | 0.397 ± 0.044 |
| Corrupted EEG | 59.16 ± 2.44 | 29.70 ± 1.64 | 61.68 ± 2.01 | 32.65 ± 0.80 | 0.183 ± 0.049 | 0.063 ± 0.022 | 0.233 ± 0.040 | 0.102 ± 0.011 |
| SS | 66.97 ± 1.29 | 39.55 ± 1.43 | 68.57 ± 0.95 | 42.47 ± 2.19 | 0.350 ± 0.025 | 0.195 ± 0.019 | 0.381 ± 0.019 | 0.242 ± 0.029 |
| IDW | 66.84 ± 2.60 | 39.34 ± 1.63 | 69.46 ± 1.40 | 42.86 ± 1.47 | 0.337 ± 0.052 | 0.192 ± 0.022 | 0.389 ± 0.028 | 0.238 ± 0.020 |
| KRIGING | 67.37 ± 1.20 | 40.11 ± 1.91 | 67.07 ± 0.85 | 41.23 ± 1.72 | 0.318 ± 0.024 | 0.186 ± 0.025 | 0.361 ± 0.017 | 0.216 ± 0.023 |
| STC | 62.33 ± 3.04 | 34.83 ± 3.12 | 66.70 ± 2.22 | 38.47 ± 1.25 | 0.246 ± 0.061 | 0.132 ± 0.042 | 0.334 ± 0.044 | 0.179 ± 0.016 |
| STDC | 59.16 ± 2.44 | 29.70 ± 1.64 | 61.68 ± 2.01 | 32.65 ± 0.80 | 0.183 ± 0.049 | 0.063 ± 0.022 | 0.233 ± 0.040 | 0.102 ± 0.011 |
| HALRTC | 59.16 ± 2.44 | 29.70 ± 1.64 | 61.68 ± 2.01 | 32.65 ± 0.80 | 0.183 ± 0.049 | 0.063 ± 0.022 | 0.233 ± 0.040 | 0.102 ± 0.011 |
| GLAE | 62.94 ± 1.41 | 34.96 ± 1.95 | 65.53 ± 1.93 | 37.61 ± 1.63 | 0.259 ± 0.028 | 0.133 ± 0.026 | 0.310 ± 0.039 | 0.168 ± 0.022 |
| CT | 61.08 ± 1.63 | 29.95 ± 1.73 | 63.45 ± 2.42 | 33.00 ± 1.19 | 0.222 ± 0.032 | 0.066 ± 0.023 | 0.268 ± 0.048 | 0.107 ± 0.016 |
| SAITS | 57.22 ± 1.97 | 33.89 ± 1.90 | 58.57 ± 4.11 | 32.95 ± 1.60 | 0.144 ± 0.040 | 0.119 ± 0.026 | 0.171 ± 0.082 | 0.106 ± 0.021 |
| CSDI | 64.30 ± 1.91 | 39.99 ± 1.27 | 67.81 ± 1.43 | 42.01 ± 0.57 | 0.286 ± 0.039 | 0.201 ± 0.017 | 0.356 ± 0.029 | 0.226 ± 0.008 |
| PATCHTST | 60.90 ± 1.70 | 29.93 ± 1.76 | 63.35 ± 2.18 | 32.95 ± 1.07 | 0.218 ± 0.034 | 0.066 ± 0.024 | 0.266 ± 0.044 | 0.106 ± 0.014 |
| GH | 66.78 ± 2.39 | 39.08 ± 1.64 | 68.30 ± 0.91 | 42.79 ± 1.57 | 0.346 ± 0.048 | 0.185 ± 0.022 | 0.386 ± 0.018 | 0.233 ± 0.021 |
| SRL-ITV | 62.52 ± 1.30 | 32.65 ± 1.65 | 64.46 ± 2.61 | 35.30 ± 1.05 | 0.250 ± 0.026 | 0.103 ± 0.022 | 0.288 ± 0.052 | 0.137 ± 0.014 |
| DMAE-EEG | 61.03 ± 2.04 | 36.27 ± 1.02 | 64.82 ± 1.14 | 38.32 ± 1.06 | 0.221 ± 0.041 | 0.151 ± 0.014 | 0.296 ± 0.023 | 0.177 ± 0.014 |
| RIGS-EEG | 67.97 ± 2.38 | 40.28 ± 1.79 | 69.99 ± 1.53 | 43.80 ± 1.41 | 0.359 ± 0.048 | 0.205 ± 0.024 | 0.399 ± 0.031 | 0.250 ± 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 , one potential extension is to reformulate observation consistency as consistency between 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
| Method | PhysioNet 25% | PhysioNet 50% | PhysioNet 75% | SEED-IV 25% | SEED-IV 50% | SEED-IV 75% | MCD 25% | MCD 50% | MCD 75% |
|---|---|---|---|---|---|---|---|---|---|
| SS | 0.0545 ± 0.0151 | 0.0614 ± 0.0169 | 0.0792 ± 0.0210 | 0.4177 ± 1.2598 | 0.4938 ± 1.5038 | 0.6734 ± 2.0720 | 0.1001 ± 0.0238 | 0.1114 ± 0.0276 | 0.1417 ± 0.0351 |
| IDW | 0.0687 ± 0.0086 | 0.0711 ± 0.0087 | 0.0765 ± 0.0092 | 0.2864 ± 0.6457 | 0.2948 ± 0.6601 | 0.3051 ± 0.6547 | 0.0873 ± 0.0177 | 0.0891 ± 0.0184 | 0.0948 ± 0.0194 |
| KRIGING | 0.0789 ± 0.0252 | 0.0900 ± 0.0291 | 0.0862 ± 0.0245 | 1.2403 ± 4.2704 | 1.3902 ± 4.7575 | 1.0059 ± 3.2861 | 0.1468 ± 0.0387 | 0.1658 ± 0.0444 | 0.1501 ± 0.0378 |
| STC | 0.0927 ± 0.0093 | 0.0936 ± 0.0094 | 0.0951 ± 0.0096 | 0.3291 ± 0.7448 | 0.3332 ± 0.7468 | 0.3414 ± 0.7437 | 0.0956 ± 0.0159 | 0.0962 ± 0.0164 | 0.1083 ± 0.0170 |
| STDC | 0.0898 ± 0.0090 | 0.0955 ± 0.0093 | 0.1030 ± 0.0102 | 0.3231 ± 0.6275 | 0.3528 ± 0.6942 | 0.3835 ± 0.6780 | 0.0908 ± 0.0176 | 0.0919 ± 0.0179 | 0.0942 ± 0.0187 |
| HALRTC | 0.0893 ± 0.0090 | 0.0952 ± 0.0095 | 0.1025 ± 0.0103 | 0.3238 ± 0.6350 | 0.3532 ± 0.6756 | 0.3819 ± 0.6450 | 0.0897 ± 0.0178 | 0.0907 ± 0.0176 | 0.0939 ± 0.0182 |
| GLAE | 0.0520 ± 0.0119 | 0.0596 ± 0.0111 | 0.0955 ± 0.0079 | 0.4155 ± 1.2778 | 0.3161 ± 0.8697 | 0.2310 ± 0.4536 | 0.0953 ± 0.0164 | 0.0942 ± 0.0109 | 0.1198 ± 0.0142 |
| CT | 0.0552 ± 0.0154 | 0.0691 ± 0.0105 | 0.0825 ± 0.0169 | 0.1137 ± 0.0915 | 0.1320 ± 0.4895 | 0.1460 ± 0.1211 | 0.0896 ± 0.0177 | 0.0905 ± 0.0179 | 0.0938 ± 0.0188 |
| SAITS | 0.0524 ± 0.0116 | 0.0636 ± 0.0113 | 0.1064 ± 0.0211 | 0.1384 ± 0.2389 | 0.1391 ± 0.1764 | 0.1556 ± 0.1052 | 0.0929 ± 0.0176 | 0.1017 ± 0.0186 | 0.1433 ± 0.0261 |
| CSDI | 0.0659 ± 0.0103 | 0.0725 ± 0.0101 | 0.0881 ± 0.0113 | 0.1097 ± 0.0271 | 0.1165 ± 0.0286 | 0.1365 ± 0.0345 | 0.1011 ± 0.0198 | 0.1301 ± 0.0218 | 0.2052 ± 0.0315 |
| PATCHTST | 0.0592 ± 0.0175 | 0.0697 ± 0.0260 | 0.0990 ± 0.0151 | 0.1145 ± 0.0366 | 0.1344 ± 0.0246 | 0.1569 ± 0.0397 | 0.0896 ± 0.0174 | 0.0905 ± 0.0180 | 0.0941 ± 0.0182 |
| GH | 0.0547 ± 0.0115 | 0.0603 ± 0.0111 | 0.0785 ± 0.0101 | 0.2373 ± 0.5885 | 0.2632 ± 0.6532 | 0.2858 ± 0.6556 | 0.0891 ± 0.0195 | 0.0920 ± 0.0204 | 0.0965 ± 0.0211 |
| SRL-ITV | 0.0557 ± 0.0095 | 0.0704 ± 0.0074 | 0.1166 ± 0.0061 | 0.0882 ± 0.0215 | 0.1007 ± 0.0382 | 0.1290 ± 0.0485 | 0.0878 ± 0.0146 | 0.0993 ± 0.0096 | 0.1389 ± 0.0108 |
| DMAE-EEG | 0.0732 ± 0.0102 | 0.0762 ± 0.0104 | 0.0934 ± 0.0144 | 0.1638 ± 0.1127 | 0.1620 ± 0.1021 | 0.1770 ± 0.0701 | 0.0880 ± 0.0170 | 0.0891 ± 0.0164 | 0.1029 ± 0.0135 |
| RIGS-EEG | 0.0513 ± 0.0120 | 0.0578 ± 0.0112 | 0.0725 ± 0.0103 | 0.0858 ± 0.0281 | 0.0985 ± 0.0335 | 0.1270 ± 0.0440 | 0.0861 ± 0.0177 | 0.0883 ± 0.0185 | 0.0921 ± 0.0192 |
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=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=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).