acceptodds
Under review as a conference paper at ICLR 2027

MExFormer: Rethinking Early Channel Fusion for Multichannel EEG Representation Learning

Abstract

Learning informative EEG representations for identifying neurological states requires capturing channel-specific temporal dynamics and dependencies across spatially distinct electrodes. Existing multivariate time-series Transformers employ diverse strategies for channel interaction, including mixing information across channels during tokenization or embedding rather than retaining individual channels as explicit attention tokens, potentially limiting the direct modeling of inter-electrode interactions. We therefore introduce MExFormer, a Transformer that independently tokenizes each channel into temporally localized patches at multiple resolutions and defers channel fusion to explicit inter-channel self-attention, progressively integrating channel interactions, temporal structure, and multi-resolution information through a three-stage hierarchy. We evaluate the proposed approach on EEG classification across two public datasets under a subject-independent protocol. Beyond classification performance, we assess the quality of learned representations through robustness to corrupted EEG inputs and post-hoc evaluation of intermediate-layer representations using a weak classifier. Across the datasets and six evaluation metrics, MExFormer achieves the strongest overall performance on APAVA dataset and remains competitive on TDBRAIN dataset, while exhibiting greater robustness to channel corruption and more discriminative intermediate representations.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
EEG Representation Learning, Multichannel EEG, Time-Series Transformers, Cross-Channel Attention, Multi-Resolution Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_QTmtOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

MExFormer argues that multivariate time-series Transformers fuse channel information too early (during tokenization/embedding), preventing explicit modeling of inter-electrode interactions at the point where channel-specific temporal structure is still localized. It tokenizes each channel independently into multi-resolution temporal patches and organizes attention in three stages: inter-channel self-attention at each temporal position (with a learned "container" token aggregating and redistributing cross-channel information), temporal attention over containers plus a per-granularity "router", and inter-granularity attention over routers. Evaluation uses two binary clinical datasets under subject-independent splits — APAVA (Alzheimer's, 23 subjects) and TDBRAIN (Parkinson's, 72 subjects) — with 1-second segments, against 11 time-series Transformer baselines. MExFormer is best on APAVA (79.71 accuracy) and competitive on TDBRAIN (90.23 vs Informer's 90.50); it shows greater robustness to channel exclusion (10–30% of channels zeroed) and more discriminative intermediate layers under a weak logistic-regression probe.

Critical strengths
  • The design question — at which stage channel information should be fused — is well articulated, and the container/router mechanism is a clean way to make cross-channel interaction an explicit attention computation. Evaluating robustness to channel corruption and probing intermediate layers with a weak classifier are thoughtful complements to accuracy that directly test the representation claim.
  • Experiments are transparent and honestly scoped: fixed subject-level partitions shared across models, five seeds with mean ± SD, and the paper openly states it leads on only one of two datasets and lists the missing architectural ablation as future work.
Critical weaknesses
  • The central claim is never isolated. There is no ablation that changes only the fusion ordering (e.g., the same blocks with channel fusion deferred to after temporal aggregation), which the paper's own limitations acknowledge — so the results cannot be attributed to "channel-first interaction" rather than capacity, multi-resolution patching, or the container mechanism itself. For a paper whose thesis is about ordering, this is the decisive missing experiment.
  • The evaluation is too narrow for an EEG representation-learning claim: two small binary clinical datasets (23 and 72 subjects, 1-second segments), no EEG-specific baselines (EEGNet, EEG Conformer, LaBraM/CBraMod), and no standard EEG benchmarks (motor imagery, emotion, sleep). On TDBRAIN the strongest baseline actually edges MExFormer, and most margins are within one seed standard deviation; the intermediate-layer probe covers only four models and layers 1/3/6.
Questions for authors
  • Can you add the ordering ablation identified as future work (identical blocks, channel fusion after temporal aggregation, and vice versa) before the review period ends?
  • How do EEG-specific models and at least two standard EEG benchmarks change the picture?
  • Is one-second segmentation appropriate for AD/PD detection here, and do the conclusions hold across segment lengths?
11 comments

We sincerely thank the reviewer for the thoughtful review and constructive suggestions.
Following the publication of the reviews, we conducted additional experiments to address the concerns raised, including an architectural ordering ablation, an additional EEG-specific baseline, an evaluation on a motor imagery dataset, and experiments with different EEG segmentation of the samples.
We address each point below.

Q1: Can you add the ordering ablation identified as future work (identical blocks, channel fusion after temporal aggregation, and vice versa) before the review period ends?

We fully agree that this is an important experiment for isolating the contribution of the proposed channel-first interaction. In response, we conducted a controlled ordering ablation in which we swap only the first two attention stages while keeping the remaining architectural configuration, model capacity, and training hyperparameters unchanged.

Specifically, the original MExFormer first performs inter-channel attention between channel-specific temporal representations and subsequently aggregates information across time through temporal attention over the resulting container representations. In the ablated configuration, the same operations are applied in the reverse order: temporal attention is first performed independently within each channel, and the resulting temporally contextualized channel representations are subsequently processed through inter-channel attention. The inter-granularity stage and all remaining components are kept unchanged.

The results are reported in Tables below as mean ±\pm STD across the same five random seeds used in the main experiments.

On APAVA, reversing the ordering decreases macro-F1 from 77.3677.36 to 76.1676.16 and substantially increases the variability across runs. Although the reversed configuration achieves higher AUROC and AUPRC, the original channel-first configuration provides more stable and consistent classification performance. This suggests that resolving inter-channel interactions before temporal aggregation is particularly beneficial for obtaining robust decision boundaries on APAVA, which is consistent with the importance of altered inter-regional EEG interactions in Alzheimer's disease.

The effect is substantially clearer on TDBRAIN. Reversing the ordering decreases macro-F1 from 90.2090.20 to 80.5580.55, with corresponding decreases in accuracy, AUROC, and AUPRC. This controlled comparison therefore shows that the sequence in which channel and temporal interactions are performed can have a substantial effect even when the underlying blocks, model capacity, and training configuration are kept fixed.

Overall, these results provide direct evidence that the ordering of channel and temporal interactions is an important architectural design choice. In particular, the strong degradation of the temporal-first variant on TDBRAIN supports the motivation for resolving inter-channel interactions while channel-specific temporal representations remain explicit, before subsequent temporal aggregation.

APAVA (2-Classes)

ModelsAccuracyPrecisionRecallF1 scoreAUROCAUPRC
MExFormer (full)79.71±1.7179.71_{\pm1.71}82.34±1.5882.34_{\pm1.58}76.42±2.1676.42_{\pm2.16}77.36±2.2277.36_{\pm2.22}89.79±2.8389.79_{\pm2.83}89.36±3.0689.36_{\pm3.06}
Ordering ablation79.43±3.4579.43_{\pm3.45}84.87±2.4284.87_{\pm2.42}75.34±4.3075.34_{\pm4.30}76.16±4.8076.16_{\pm4.80}91.98±1.9191.98_{\pm1.91}91.95±2.1291.95_{\pm2.12}

TDBRAIN (2-Classes)

ModelsAccuracyPrecisionRecallF1 scoreAUROCAUPRC
MExFormer (full)90.23±0.9290.23_{\pm0.92}90.64±0.8690.64_{\pm0.86}90.23±0.9290.23_{\pm0.92}90.20±0.9290.20_{\pm0.92}97.28±0.8197.28_{\pm0.81}97.35±0.8897.35_{\pm0.88}
Ordering ablation80.98±2.2580.98_{\pm2.25}83.76±1.5683.76_{\pm1.56}80.98±2.2580.98_{\pm2.25}80.55±2.6480.55_{\pm2.64}89.05±3.4489.05_{\pm3.44}88.59±3.8788.59_{\pm3.87}

Q2: How do EEG-specific models and at least two standard EEG benchmarks change the picture?

We agree that including EEG-specific architectures strengthens the scope of the evaluation. In response, we incorporated EEG Conformer as an additional EEG-specific baseline under the same subject-independent partitions, training environment, and evaluation protocol used for the models in the original submission.

We selected EEG Conformer as the EEG-specific baseline because it provides the closest architectural comparison to MExFormer under the same supervised training setting, combining EEG-oriented processing with Transformer-based attention. EEGNet is a CNN-based model, while LaBraM and CBraMod rely on large-scale EEG pretraining and external data, making them less directly comparable for isolating architectural differences within our controlled evaluation.

The results show that MExFormer substantially outperforms EEG Conformer on APAVA across the main classification metrics. On clean TDBRAIN, the two models are closely matched in accuracy and macro-F1, while EEG Conformer achieves higher AUROC and AUPRC. This again suggests that the benefit of a particular EEG architecture can depend on the characteristics of the task and evaluation metric, rather than one architecture being uniformly superior under all conditions.

The difference becomes clearer under channel corruption. MExFormer consistently retains higher accuracy and macro-F1 than EEG Conformer across all corruption levels on both datasets, with the gap becoming more pronounced as additional channels are excluded. In particular, at 30% channel exclusion on TDBRAIN, MExFormer retains 86.60% accuracy compared with 83.77% for EEG Conformer. This additional EEG-specific comparison therefore reinforces the robustness result observed against the original time-series baselines and supports the proposed motivation for preserving channel-specific representations and explicitly modeling their interactions before temporal aggregation.

APAVA (2-Classes)

ModelsAccuracyPrecisionRecallF1 scoreAUROCAUPRC
EEG Conformer74.16±3.4374.16_{\pm3.43}76.72±6.3776.72_{\pm6.37}71.02±1.6671.02_{\pm1.66}71.39±2.0671.39_{\pm2.06}78.46±5.8778.46_{\pm5.87}77.81±6.2977.81_{\pm6.29}

TDBRAIN (2-Classes)

ModelsAccuracyPrecisionRecallF1 scoreAUROCAUPRC
EEG Conformer90.06±3.9490.06_{\pm3.94}91.34±2.6591.34_{\pm2.65}90.06±3.9490.06_{\pm3.94}89.95±4.0889.95_{\pm4.08}98.49±0.5798.49_{\pm0.57}98.55±0.5598.55_{\pm0.55}

APAVA (2-Classes)

Models10% Accuracy10% F1 score20% Accuracy20% F1 score30% Accuracy30% F1 score
EEG Conformer72.27±2.0772.27_{\pm2.07}68.68±1.0168.68_{\pm1.01}70.50±1.3270.50_{\pm1.32}65.77±2.3865.77_{\pm2.38}67.78±1.8567.78_{\pm1.85}60.93±5.9060.93_{\pm5.90}

TDBRAIN (2-Classes)

Models10% Accuracy10% F1 score20% Accuracy20% F1 score30% Accuracy30% F1 score
EEG Conformer88.33±2.9288.33_{\pm2.92}88.26±3.0288.26_{\pm3.02}86.81±0.8986.81_{\pm0.89}86.78±0.9386.78_{\pm0.93}83.77±1.5183.77_{\pm1.51}83.72±1.5583.72_{\pm1.55}

Due to the limited review-period timeframe, we were not able to complete an additional large-scale EEG benchmark evaluation. We therefore prioritized experiments that directly test the central representation-learning claims of the paper, including the ordering ablation, aggregation ablation, multi-resolution analysis, EEG-specific baseline comparison, and intermediate-representation probing. Extending the evaluation to larger and more diverse EEG datasets is an immediate next step, and we plan to investigate whether the same representation-learning behavior and robustness trends hold at substantially larger scale.

Q3: Is one-second segmentation appropriate for AD/PD detection here, and do the conclusions hold across segment lengths?

We thank the reviewer for raising this point. To examine whether the conclusions depend on the one-second segmentation used in the original evaluation, we extended the experiments to two- and four-second EEG segments while retaining the same subject-independent evaluation protocol.

The additional results show that segmentation length affects absolute performance and that the preferred temporal context varies across datasets. On APAVA, MExFormer achieves the strongest performance across both additional segment lengths, reaching 83.77%83.77\% accuracy and 81.92%81.92\% macro-F1 with two-second segments and 80.38%80.38\% accuracy and 79.89%79.89\% macro-F1 with four-second segments. Notably, its advantage over the competing models is preserved and becomes more pronounced than in the original one-second setting.

On TDBRAIN, the results remain consistent with the original one-second evaluation. Informer achieves the highest accuracy and macro-F1 at both two and four seconds, while MExFormer remains competitive across both segmentation lengths. This consistency across temporal contexts indicates that the relative behavior of the models is not an artifact of the one-second segmentation choice.

Overall, these experiments indicate that the main conclusions are not specific to one-second segmentation. Across three temporal contexts, MExFormer consistently shows a particularly strong advantage on APAVA, while the relative advantage of early-fusion architectures on clean TDBRAIN remains visible. We will include these additional segmentation-length experiments and clarify this variation across datasets in the revised manuscript.

APAVA (2-Classes)

Models2s Accuracy2s F1 score2s AUPRC4s Accuracy4s F1 score4s AUPRC
Autoformer70.72±8.0470.72_{\pm8.04}69.62±7.9969.62_{\pm7.99}74.73±11.2674.73_{\pm11.26}62.26±4.1562.26_{\pm4.15}59.31±3.7159.31_{\pm3.71}63.40±5.4363.40_{\pm5.43}
Crossformer74.84±1.5174.84_{\pm1.51}70.94±2.3170.94_{\pm2.31}78.56±1.8078.56_{\pm1.80}53.08±8.2053.08_{\pm8.20}51.64±8.0751.64_{\pm8.07}58.22±9.2358.22_{\pm9.23}
FEDformer71.98±1.6971.98_{\pm1.69}70.74±1.4570.74_{\pm1.45}80.18±0.9980.18_{\pm0.99}62.14±2.6862.14_{\pm2.68}59.97±3.8459.97_{\pm3.84}69.73±2.7869.73_{\pm2.78}
Informer69.47±5.4569.47_{\pm5.45}64.91±6.0964.91_{\pm6.09}65.40±7.1265.40_{\pm7.12}56.23±4.7856.23_{\pm4.78}49.13±3.4049.13_{\pm3.40}55.29±4.4155.29_{\pm4.41}
iTransformer75.31±2.4375.31_{\pm2.43}74.44±2.3874.44_{\pm2.38}82.41±2.6082.41_{\pm2.60}60.75±3.9260.75_{\pm3.92}58.43±4.7858.43_{\pm4.78}60.20±3.3560.20_{\pm3.35}
MTST71.67±1.9371.67_{\pm1.93}64.82±3.3764.82_{\pm3.37}71.06±1.9271.06_{\pm1.92}65.41±3.8065.41_{\pm3.80}61.93±4.4461.93_{\pm4.44}65.92±2.9665.92_{\pm2.96}
Nonformer74.09±3.3674.09_{\pm3.36}71.50±2.5971.50_{\pm2.59}71.96±1.3871.96_{\pm1.38}61.64±4.8161.64_{\pm4.81}55.73±2.4755.73_{\pm2.47}57.74±5.3357.74_{\pm5.33}
PatchTST67.36±2.6567.36_{\pm2.65}61.02±6.0661.02_{\pm6.06}73.65±5.4273.65_{\pm5.42}67.80±7.7367.80_{\pm7.73}65.43±6.9365.43_{\pm6.93}70.05±4.6870.05_{\pm4.68}
Reformer73.62±0.5573.62_{\pm0.55}68.62±1.6268.62_{\pm1.62}71.13±3.3371.13_{\pm3.33}61.51±2.9361.51_{\pm2.93}51.54±7.5651.54_{\pm7.56}59.97±4.6959.97_{\pm4.69}
Transformer71.92±5.0871.92_{\pm5.08}66.72±6.5966.72_{\pm6.59}68.05±4.1768.05_{\pm4.17}58.11±3.5758.11_{\pm3.57}51.04±5.6051.04_{\pm5.60}53.38±3.0153.38_{\pm3.01}
Medformer75.97±1.9675.97_{\pm1.96}73.31±1.7173.31_{\pm1.71}83.43±2.2183.43_{\pm2.21}74.34±3.6574.34_{\pm3.65}70.56±4.3470.56_{\pm4.34}79.52±3.7979.52_{\pm3.79}
EEG Conformer75.03±4.9175.03_{\pm4.91}72.90±5.2172.90_{\pm5.21}79.35±7.0379.35_{\pm7.03}78.36±3.8278.36_{\pm3.82}75.35±4.9075.35_{\pm4.90}80.50±2.0680.50_{\pm2.06}
MExFormer83.77±2.27‾\underline{83.77_{\pm2.27}}81.92±2.92‾\underline{81.92_{\pm2.92}}94.88±2.32‾\underline{94.88_{\pm2.32}}80.38±3.70‾\underline{80.38_{\pm3.70}}79.89±3.94‾\underline{79.89_{\pm3.94}}89.42±2.60‾\underline{89.42_{\pm2.60}}
Reviewer_s4qmOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

MExFormer is a supervised transformer for multichannel EEG classification. Each channel is patched separately at several temporal resolutions. Within each layer, channel tokens and a learnable container first attend to each other at every time position, the containers then attend across time together with a per-resolution router, and the routers finally attend across resolutions. It is tested on two binary clinical datasets, APAVA and TDBRAIN, with channel-dropout robustness and layer-wise linear probes.

Critical strengths
  1. The question of when channel information should be fused is clear and relevant, and the related work lays out the design space well, from early mixing in Informer-style tokens to channel-independent PatchTST and router-based Crossformer. The three-stage ordering is simple and easy to follow.

  2. The robustness test is a useful addition. Under 30 percent channel dropout on TDBRAIN, MExFormer keeps 86.6 percent accuracy while the best baseline, Medformer, falls to 80.2, which fits the idea that keeping channels as separate tokens helps when electrodes are missing.

Critical weaknesses
  1. The central claim is never tested directly. The paper argues that resolving channel interactions before temporal aggregation is what matters, but there is no ablation that swaps the stage order, removes the container, or uses a single resolution, and the limitations section admits this. Without those controls, the gains could come from the multi-resolution patching or simply from model size, which is not reported.

  2. The evaluation is too narrow for general claims. There are two small binary datasets with one fixed subject split, and with the Medformer split the APAVA test set has only four subjects. Medformer is reported at 76.6 accuracy on APAVA and 87.7 on TDBRAIN, while its own paper reports 78.7 and 89.6 with the same seeds. On TDBRAIN, Informer is slightly ahead. EEG models such as EEGNet, EEGConformer or CBraMod are missing.

  3. Key details for reproduction are missing. The patch lengths, number of layers, hidden size, number of heads, training schedule, parameter count and how baseline hyperparameters were chosen are not given, and no code is provided. The layer-wise probe is also hard to read, because it is unclear which tokens are probed, and on TDBRAIN at layer 1 MExFormer scores 80.8 with a standard deviation of 6.6, against 81.4 for Medformer.

Questions for authors

see weaknesses.

7 comments

We thank the reviewer for the careful evaluation and for highlighting several aspects of the experimental design, comparison scope, and reproducibility that required further clarification and additional evidence. We address these points below.

Q1: The central claim is never tested directly. The paper argues that resolving channel interactions before temporal aggregation is what matters, but there is no ablation that swaps the stage order, removes the container, or uses a single resolution, and the limitations section admits this. Without those controls, the gains could come from the multi-resolution patching or simply from model size, which is not reported.

We agree that the original submission did not isolate the central architectural claim as directly as it should have. In particular, the observed gains could not be cleanly separated from the effects of the aggregation tokens or the use of multiple temporal resolutions. We therefore conducted additional controlled ablations targeting these alternative explanations.

First, we directly test the proposed channel-before-temporal ordering by keeping the same attention blocks, model dimensions, training hyperparameters, and overall capacity, and changing only the order of the first two stages. In the original MExFormer, inter-channel attention is performed before temporal aggregation, whereas in the reversed variant temporal attention is first applied independently within each channel and the resulting representations are subsequently processed by inter-channel attention.

This controlled comparison provides direct evidence that the ordering itself matters. On TDBRAIN, reversing the stages reduces accuracy from 90.23%90.23\% to 80.98%80.98\% and macro-F1 from 90.20%90.20\% to 80.55%80.55\%, together with substantial reductions in AUROC and AUPRC. On APAVA, the effect is more moderate: the original ordering achieves higher accuracy and macro-F1 and substantially lower variability across runs, although the reversed configuration achieves higher AUROC and AUPRC. Thus, the fixed-capacity experiment shows that the proposed ordering cannot be explained simply by differences in model size or by the choice of attention blocks.

APAVA (2-Classes)

ModelsAccuracyPrecisionRecallF1 scoreAUROCAUPRC
MExFormer (full)79.71±1.7179.71_{\pm1.71}82.34±1.5882.34_{\pm1.58}76.42±2.1676.42_{\pm2.16}77.36±2.2277.36_{\pm2.22}89.79±2.8389.79_{\pm2.83}89.36±3.0689.36_{\pm3.06}
Ordering ablation79.43±3.4579.43_{\pm3.45}84.87±2.4284.87_{\pm2.42}75.34±4.3075.34_{\pm4.30}76.16±4.8076.16_{\pm4.80}91.98±1.9191.98_{\pm1.91}91.95±2.1291.95_{\pm2.12}
Mean-pool ablation80.15±1.3080.15_{\pm1.30}82.92±1.0082.92_{\pm1.00}76.87±1.6876.87_{\pm1.68}77.86±1.7077.86_{\pm1.70}89.60±3.9589.60_{\pm3.95}89.15±3.9789.15_{\pm3.97}

TDBRAIN (2-Classes)

ModelsAccuracyPrecisionRecallF1 scoreAUROCAUPRC
MExFormer (full)90.23±0.9290.23_{\pm0.92}90.64±0.8690.64_{\pm0.86}90.23±0.9290.23_{\pm0.92}90.20±0.9290.20_{\pm0.92}97.28±0.8197.28_{\pm0.81}97.35±0.8897.35_{\pm0.88}
Ordering ablation80.98±2.2580.98_{\pm2.25}83.76±1.5683.76_{\pm1.56}80.98±2.2580.98_{\pm2.25}80.55±2.6480.55_{\pm2.64}89.05±3.4489.05_{\pm3.44}88.59±3.8788.59_{\pm3.87}
Mean-pool ablation75.42±5.1375.42_{\pm5.13}80.84±1.9780.84_{\pm1.97}75.42±5.1375.42_{\pm5.13}74.09±6.3674.09_{\pm6.36}87.28±4.2087.28_{\pm4.20}86.82±4.9286.82_{\pm4.92}

We additionally replace the container and router aggregation with mean pooling. This modification has only a small effect on APAVA, but causes a substantial degradation on TDBRAIN, where macro-F1 decreases from 90.20%90.20\% to 74.09%74.09\%. This indicates that the learned aggregation mechanism provides an additional benefit, particularly on TDBRAIN, but is distinct from the ordering effect isolated above.

To test whether the gains are instead explained by multi-resolution patching, we also evaluate single-resolution MExFormer variants using patch lengths of 8, 16, 32, and 64. The full multi-resolution configuration performs best on both datasets, confirming that multiple resolutions contribute to the final performance. However, the ordering ablation changes neither the multi-resolution configuration nor the model capacity, so the large degradation observed when the first two stages are reversed cannot be attributed to multi-resolution processing.

APAVA (2-Classes)

Patch lengthAccuracyPrecisionRecallF1 scoreAUROCAUPRC
872.77±4.2872.77_{\pm4.28}72.76±4.6472.76_{\pm4.64}70.64±3.2270.64_{\pm3.22}70.97±3.6370.97_{\pm3.63}75.13±2.6875.13_{\pm2.68}75.44±3.8675.44_{\pm3.86}
1673.11±1.6173.11_{\pm1.61}74.91±3.2674.91_{\pm3.26}69.59±2.0169.59_{\pm2.01}69.83±2.3469.83_{\pm2.34}75.21±4.4575.21_{\pm4.45}76.11±4.3276.11_{\pm4.32}
3277.65±1.0277.65_{\pm1.02}80.13±1.3180.13_{\pm1.31}74.19±1.4674.19_{\pm1.46}74.96±1.5474.96_{\pm1.54}78.77±2.9478.77_{\pm2.94}79.59±3.1379.59_{\pm3.13}
6475.35±1.3275.35_{\pm1.32}77.37±1.8077.37_{\pm1.80}71.67±1.4971.67_{\pm1.49}72.27±1.6372.27_{\pm1.63}76.21±5.8876.21_{\pm5.88}75.38±5.9175.38_{\pm5.91}
full79.71±1.7179.71_{\pm1.71}82.34±1.5882.34_{\pm1.58}76.42±2.1676.42_{\pm2.16}77.36±2.2277.36_{\pm2.22}89.79±2.8389.79_{\pm2.83}89.36±3.0689.36_{\pm3.06}

TDBRAIN (2-Classes)

Patch lengthAccuracyPrecisionRecallF1 scoreAUROCAUPRC
886.38±2.7786.38_{\pm2.77}87.69±3.1387.69_{\pm3.13}86.38±2.7786.38_{\pm2.77}86.26±2.7886.26_{\pm2.78}95.60±3.0495.60_{\pm3.04}95.64±3.1795.64_{\pm3.17}
1680.58±17.1480.58_{\pm17.14}75.96±28.5275.96_{\pm28.52}80.58±17.1480.58_{\pm17.14}77.22±24.5677.22_{\pm24.56}94.18±3.6694.18_{\pm3.66}93.91±3.9493.91_{\pm3.94}
3283.77±3.3083.77_{\pm3.30}84.60±2.9684.60_{\pm2.96}83.77±3.3083.77_{\pm3.30}83.66±3.3683.66_{\pm3.36}93.25±2.1293.25_{\pm2.12}93.62±1.9893.62_{\pm1.98}
6476.62±2.5476.62_{\pm2.54}76.79±2.5376.79_{\pm2.53}76.63±2.5476.63_{\pm2.54}76.59±2.5676.59_{\pm2.56}85.96±2.5885.96_{\pm2.58}85.81±3.0785.81_{\pm3.07}
full90.23±0.9290.23_{\pm0.92}90.64±0.8690.64_{\pm0.86}90.23±0.9290.23_{\pm0.92}90.20±0.9290.20_{\pm0.92}97.28±0.8197.28_{\pm0.81}97.35±0.8897.35_{\pm0.88}

Taken together, these additional experiments separate the main factors raised by the reviewer: the channel-before-temporal ordering has an independent effect under a controlled fixed-capacity comparison, while the aggregation tokens and multi-resolution processing provide additional benefits. We will revise the corresponding discussion and limitations accordingly.

Q2: The evaluation is too narrow for general claims. There are two small binary datasets with one fixed subject split, and with the Medformer split the APAVA test set has only four subjects. Medformer is reported at 76.6 accuracy on APAVA and 87.7 on TDBRAIN, while its own paper reports 78.7 and 89.6 with the same seeds. On TDBRAIN, Informer is slightly ahead. EEG models such as EEGNet, EEGConformer or CBraMod are missing.

We thank the reviewer for raising these concerns regarding the evaluation protocol and the scope of the baseline comparison.

Regarding the subject-independent evaluation, we intentionally use the subject partitions provided in the final version of the official Medformer codebase in order to maintain a controlled comparison with the closest prior work. This includes the APAVA partition with four subjects in the test set. Rather than modifying this established partition, which would reduce direct comparability with the reference implementation, we retain it consistently across all evaluated models.

Due to the limited review-period timeframe, we were not able to complete an additional large-scale EEG benchmark evaluation. We therefore prioritized experiments that directly test the central representation-learning claims of the paper, including the ordering ablation, aggregation ablation, multi-resolution analysis, EEG-specific baseline comparison, and intermediate-representation probing. Extending the evaluation to larger and more diverse EEG datasets is an immediate next step, and we plan to investigate whether the same representation-learning behavior and robustness trends hold at substantially larger scale.

We would also like to clarify the comparison with the Medformer results reported in the original paper. We use the same random seeds as the Medformer evaluation, but our experiments are based on the final version of the official Medformer codebase, which was updated by the authors after publication. Therefore, our experiments do not correspond exactly to the implementation state associated with the originally published results. In addition, we use a stricter early-stopping criterion across all models in our experimental pipeline. Specifically, patience is set to 10 epochs and the counter is reset only when the validation macro-F1 improves by at least (0.001) over the current best value. In the Medformer implementation, the corresponding improvement threshold is (0.00001). We found this criterion to be overly sensitive to very small validation fluctuations and therefore adopted the stricter threshold consistently across all evaluated models. All baseline results in our paper are obtained by re-running the models from the final Medformer codebase under this common training and evaluation pipeline, using the same subject partitions, random seeds, and checkpoint-selection procedure as MExFormer. We will make these implementation differences explicit in the revised manuscript.

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

This paper asks when channel information should be fused in Transformer based multichannel EEG classification. MExFormer tokenizes each channel independently into multi-resolution patches and fuses information through three attention stages: inter-channel attention at each patch position via container tokens, temporal attention over the container tokens, and cross-resolution attention over router tokens. Under subject independent splits with five seeds on two clinical EEG datasets, APAVA for Alzheimer detection and TDBRAIN for Parkinson detection, it outperforms eleven generic time-series Transformers on APAVA, ranks second on clean TDBRAIN behind Informer, degrades most gracefully under test-time channel zeroing, and yields more discriminative intermediate representations under a linear probe.

Critical strengths

The paper reframes architecture design around the ordering of channel, temporal, and resolution fusion, and supports it with evaluations beyond accuracy. Under zeroing of 10 to 30 percent of channel values at test time on TDBRAIN, MExFormer drops about 3.6 points from 90.23 to 86.60 while Reformer drops about 11.9 points from 90.22 to 78.35 according to Table 3, and layer-wise probes show progressively more discriminative representations in Table 4. This is practically relevant for deployments with faulty electrodes.

Critical weaknesses

The central hypothesis that channel-first fusion before temporal fusion causes the gains is never isolated. There is no ablation that swaps the first two stages within MExFormer, removes the container and router tokens, or varies the number of granularities, and the comparisons with Crossformer and Medformer confound ordering with capacity, tuning, and router approximation differences. Section 5 defers controlled architectural ablations to future work, which is an explicit admission that the core claim is untested.

Questions for authors
  1. Can you report an ablation that keeps the capacity fixed and only swaps the first two stages, replaces the container and router tokens with mean pooling, and varies the number of granularities? Does the ordering effect survive? This directly tests the central claim.
  2. Is the improvement over Reformer on APAVA statistically significant under a paired test across seeds, and how robust are the conclusions to redrawing the subject level splits, given that the paper uses one fixed split with seeds varying only the initialization?
  3. What exactly does randomly zeroing each input channel whole values by 10 percent mean? Is a random subset of channels fully zeroed per sample, or are 10 to 30 percent of time points zeroed within every channel? Is the corrupted channel set identical across models and seeds, and would baselines trained with channel dropout augmentation close the robustness gap?
  4. How were the eleven baselines tuned, in terms of search space, budget, and selection metric, and do the twelve models have comparable parameter counts and FLOPs? In particular, why does the early fusing Informer beat MExFormer on clean TDBRAIN, and how does this square with the motivation against early fusion?
  5. On TDBRAIN, accuracy and recall are exactly equal for all twelve models in Table 2, unlike APAVA. Is this a table construction artifact, or is the test set exactly class balanced? Please also specify the layer counts used for the layer 1, 3, and 6 probes, and whether all baselines had at least six layers.
9 comments

We thank the reviewer for the feedback and for identifying several aspects of the experimental evaluation that required further clarification. We address the questions below.

Q1: Can you report an ablation that keeps the capacity fixed and only swaps the first two stages, replaces the container and router tokens with mean pooling, and varies the number of granularities? Does the ordering effect survive? This directly tests the central claim.

We thank the reviewer for suggesting these controlled ablations. We conducted additional experiments to separately examine the contribution of the channel--temporal ordering, the aggregation tokens, and the use of multiple temporal resolutions.

For the ordering ablation, we retained the same attention blocks, model dimensions, training hyperparameters, and overall capacity, and changed only the order of the first two stages. In the original MExFormer, inter-channel attention is performed before temporal aggregation. In the reversed variant, temporal attention is first applied independently within each channel, after which the resulting temporally contextualized channel representations are processed by inter-channel attention.

The ordering effect is particularly clear on TDBRAIN (see tables below). Reversing the first two stages reduces accuracy from 90.23%90.23\% to 80.98%80.98\% and macro-F1 from 90.20%90.20\% to 80.55%80.55\%, with corresponding reductions in AUROC and AUPRC. Because the underlying blocks and model capacity are unchanged, this substantial degradation directly shows that the sequence in which channel and temporal interactions are performed can have a major effect on the learned representation.

On APAVA, the effect is more moderate. The original channel-first ordering achieves higher accuracy and macro-F1 (79.71%79.71\% and 77.36%77.36\%) than the reversed configuration (79.43%79.43\% and 76.16%76.16\%), while also exhibiting substantially lower variability across runs. The reversed configuration obtains only higher AUROC and AUPRC. Overall, the ablation supports the proposed channel-first ordering, with a particularly strong effect on TDBRAIN.

We next replaced the container and router aggregation with mean pooling. On APAVA, this modification has only a small effect on classification performance, suggesting that the representations produced by the preceding attention stages can still be summarized effectively using a simple aggregation operation for this dataset. In contrast, the effect is substantial on TDBRAIN: macro-F1 decreases from 90.20%90.20\% to 74.09%74.09\%, and the variability across runs also increases considerably. These results indicate that the aggregation tokens can play an important role by selectively propagating information between the successive channel, temporal, and granularity stages.

APAVA (2-Classes)

ModelsAccuracyPrecisionRecallF1 scoreAUROCAUPRC
MExFormer (full)79.71±1.7179.71_{\pm1.71}82.34±1.5882.34_{\pm1.58}76.42±2.1676.42_{\pm2.16}77.36±2.2277.36_{\pm2.22}89.79±2.8389.79_{\pm2.83}89.36±3.0689.36_{\pm3.06}
Ordering ablation79.43±3.4579.43_{\pm3.45}84.87±2.4284.87_{\pm2.42}75.34±4.3075.34_{\pm4.30}76.16±4.8076.16_{\pm4.80}91.98±1.9191.98_{\pm1.91}91.95±2.1291.95_{\pm2.12}
Mean-pool ablation80.15±1.3080.15_{\pm1.30}82.92±1.0082.92_{\pm1.00}76.87±1.6876.87_{\pm1.68}77.86±1.7077.86_{\pm1.70}89.60±3.9589.60_{\pm3.95}89.15±3.9789.15_{\pm3.97}

TDBRAIN (2-Classes)

ModelsAccuracyPrecisionRecallF1 scoreAUROCAUPRC
MExFormer (full)90.23±0.9290.23_{\pm0.92}90.64±0.8690.64_{\pm0.86}90.23±0.9290.23_{\pm0.92}90.20±0.9290.20_{\pm0.92}97.28±0.8197.28_{\pm0.81}97.35±0.8897.35_{\pm0.88}
Ordering ablation80.98±2.2580.98_{\pm2.25}83.76±1.5683.76_{\pm1.56}80.98±2.2580.98_{\pm2.25}80.55±2.6480.55_{\pm2.64}89.05±3.4489.05_{\pm3.44}88.59±3.8788.59_{\pm3.87}
Mean-pool ablation75.42±5.1375.42_{\pm5.13}80.84±1.9780.84_{\pm1.97}75.42±5.1375.42_{\pm5.13}74.09±6.3674.09_{\pm6.36}87.28±4.2087.28_{\pm4.20}86.82±4.9286.82_{\pm4.92}

Finally, we evaluated single-resolution variants using patch lengths of 8, 16, 32, and 64 and compared them with the full multi-resolution MExFormer. The full configuration achieves the strongest performance across all reported metrics on both APAVA and TDBRAIN. On APAVA, the best single-resolution configuration reaches 77.65%77.65\% accuracy and 74.96%74.96\% macro-F1, compared with 79.71%79.71\% and 77.36%77.36\% for the full model. On TDBRAIN, the best single-resolution configuration reaches 86.38%86.38\% accuracy and 86.26%86.26\% macro-F1, compared with 90.23%90.23\% and 90.20%90.20\% for the full model. The improvements are particularly evident in AUROC and AUPRC, supporting the benefit of combining information across multiple temporal resolutions rather than relying on a single patch scale.

Taken together, these ablations separate the three architectural factors raised by the reviewer. Most importantly, the fixed-capacity ordering experiment shows that the advantage of performing inter-channel interaction before temporal aggregation cannot be explained solely by the use of aggregation tokens or multiple temporal resolutions. We will incorporate these controlled comparisons and clarify their respective contributions in the revised manuscript.

APAVA (2-Classes)

Patch lengthAccuracyPrecisionRecallF1 scoreAUROCAUPRC
872.77±4.2872.77_{\pm4.28}72.76±4.6472.76_{\pm4.64}70.64±3.2270.64_{\pm3.22}70.97±3.6370.97_{\pm3.63}75.13±2.6875.13_{\pm2.68}75.44±3.8675.44_{\pm3.86}
1673.11±1.6173.11_{\pm1.61}74.91±3.2674.91_{\pm3.26}69.59±2.0169.59_{\pm2.01}69.83±2.3469.83_{\pm2.34}75.21±4.4575.21_{\pm4.45}76.11±4.3276.11_{\pm4.32}
3277.65±1.0277.65_{\pm1.02}80.13±1.3180.13_{\pm1.31}74.19±1.4674.19_{\pm1.46}74.96±1.5474.96_{\pm1.54}78.77±2.9478.77_{\pm2.94}79.59±3.1379.59_{\pm3.13}
6475.35±1.3275.35_{\pm1.32}77.37±1.8077.37_{\pm1.80}71.67±1.4971.67_{\pm1.49}72.27±1.6372.27_{\pm1.63}76.21±5.8876.21_{\pm5.88}75.38±5.9175.38_{\pm5.91}
full79.71±1.7179.71_{\pm1.71}82.34±1.5882.34_{\pm1.58}76.42±2.1676.42_{\pm2.16}77.36±2.2277.36_{\pm2.22}89.79±2.8389.79_{\pm2.83}89.36±3.0689.36_{\pm3.06}

TDBRAIN (2-Classes)

Patch lengthAccuracyPrecisionRecallF1 scoreAUROCAUPRC
886.38±2.7786.38_{\pm2.77}87.69±3.1387.69_{\pm3.13}86.38±2.7786.38_{\pm2.77}86.26±2.7886.26_{\pm2.78}95.60±3.0495.60_{\pm3.04}95.64±3.1795.64_{\pm3.17}
1680.58±17.1480.58_{\pm17.14}75.96±28.5275.96_{\pm28.52}80.58±17.1480.58_{\pm17.14}77.22±24.5677.22_{\pm24.56}94.18±3.6694.18_{\pm3.66}93.91±3.9493.91_{\pm3.94}
3283.77±3.3083.77_{\pm3.30}84.60±2.9684.60_{\pm2.96}83.77±3.3083.77_{\pm3.30}83.66±3.3683.66_{\pm3.36}93.25±2.1293.25_{\pm2.12}93.62±1.9893.62_{\pm1.98}
6476.62±2.5476.62_{\pm2.54}76.79±2.5376.79_{\pm2.53}76.63±2.5476.63_{\pm2.54}76.59±2.5676.59_{\pm2.56}85.96±2.5885.96_{\pm2.58}85.81±3.0785.81_{\pm3.07}
full90.23±0.9290.23_{\pm0.92}90.64±0.8690.64_{\pm0.86}90.23±0.9290.23_{\pm0.92}90.20±0.9290.20_{\pm0.92}97.28±0.8197.28_{\pm0.81}97.35±0.8897.35_{\pm0.88}

Q2: Is the improvement over Reformer on APAVA statistically significant under a paired test across seeds, and how robust are the conclusions to redrawing the subject level splits, given that the paper uses one fixed split with seeds varying only the initialization?

We acknowledge that the use of a single fixed subject-level partition limits the extent to which variability across different subject splits can be assessed. We intentionally followed the subject partitions provided by the original Medformer codebase in order to maintain a directly comparable experimental protocol with the closest prior work.

To provide additional evidence beyond this fixed configuration, we evaluated the method under different signal segmentation lengths (2 s and 4 s instead of 1 s). These experiments do not replace evaluation across multiple subject partitions, but they provide complementary evidence that the observed behavior is not specific to a single segmentation setting.

The additional results on different segment lengths are shown below. On APAVA, MExFormer achieves the strongest performance across both additional segment lengths, reaching 83.77%83.77\% accuracy and 81.92%81.92\% macro-F1 with two-second segments and 80.38%80.38\% accuracy and 79.89%79.89\% macro-F1 with four-second segments. Notably, its advantage over the competing models is preserved and becomes more pronounced than in the original one-second setting.

On TDBRAIN, the results remain consistent with the original one-second evaluation. Informer achieves the highest accuracy and macro-F1 at both two and four seconds, while MExFormer remains competitive across both segmentation lengths. This consistency across temporal contexts indicates that the relative behavior of the models is not an artifact of the one-second segmentation choice.

Overall, these experiments indicate that the main conclusions are not specific to one-second segmentation. Across three temporal contexts, MExFormer consistently shows a particularly strong advantage on APAVA, while the relative advantage of early-fusion architectures on clean TDBRAIN remains visible. We will include these additional segmentation-length experiments in the revised manuscript.

APAVA (2-Classes)

Models2s Accuracy2s F1 score2s AUPRC4s Accuracy4s F1 score4s AUPRC
Autoformer70.72±8.0470.72_{\pm8.04}69.62±7.9969.62_{\pm7.99}74.73±11.2674.73_{\pm11.26}62.26±4.1562.26_{\pm4.15}59.31±3.7159.31_{\pm3.71}63.40±5.4363.40_{\pm5.43}
Crossformer74.84±1.5174.84_{\pm1.51}70.94±2.3170.94_{\pm2.31}78.56±1.8078.56_{\pm1.80}53.08±8.2053.08_{\pm8.20}51.64±8.0751.64_{\pm8.07}58.22±9.2358.22_{\pm9.23}
FEDformer71.98±1.6971.98_{\pm1.69}70.74±1.4570.74_{\pm1.45}80.18±0.9980.18_{\pm0.99}62.14±2.6862.14_{\pm2.68}59.97±3.8459.97_{\pm3.84}69.73±2.7869.73_{\pm2.78}
Informer69.47±5.4569.47_{\pm5.45}64.91±6.0964.91_{\pm6.09}65.40±7.1265.40_{\pm7.12}56.23±4.7856.23_{\pm4.78}49.13±3.4049.13_{\pm3.40}55.29±4.4155.29_{\pm4.41}
iTransformer75.31±2.4375.31_{\pm2.43}74.44±2.3874.44_{\pm2.38}82.41±2.6082.41_{\pm2.60}60.75±3.9260.75_{\pm3.92}58.43±4.7858.43_{\pm4.78}60.20±3.3560.20_{\pm3.35}
MTST71.67±1.9371.67_{\pm1.93}64.82±3.3764.82_{\pm3.37}71.06±1.9271.06_{\pm1.92}65.41±3.8065.41_{\pm3.80}61.93±4.4461.93_{\pm4.44}65.92±2.9665.92_{\pm2.96}
Nonformer74.09±3.3674.09_{\pm3.36}71.50±2.5971.50_{\pm2.59}71.96±1.3871.96_{\pm1.38}61.64±4.8161.64_{\pm4.81}55.73±2.4755.73_{\pm2.47}57.74±5.3357.74_{\pm5.33}
PatchTST67.36±2.6567.36_{\pm2.65}61.02±6.0661.02_{\pm6.06}73.65±5.4273.65_{\pm5.42}67.80±7.7367.80_{\pm7.73}65.43±6.9365.43_{\pm6.93}70.05±4.6870.05_{\pm4.68}
Reformer73.62±0.5573.62_{\pm0.55}68.62±1.6268.62_{\pm1.62}71.13±3.3371.13_{\pm3.33}61.51±2.9361.51_{\pm2.93}51.54±7.5651.54_{\pm7.56}59.97±4.6959.97_{\pm4.69}
Transformer71.92±5.0871.92_{\pm5.08}66.72±6.5966.72_{\pm6.59}68.05±4.1768.05_{\pm4.17}58.11±3.5758.11_{\pm3.57}51.04±5.6051.04_{\pm5.60}53.38±3.0153.38_{\pm3.01}
Medformer75.97±1.9675.97_{\pm1.96}73.31±1.7173.31_{\pm1.71}83.43±2.2183.43_{\pm2.21}74.34±3.6574.34_{\pm3.65}70.56±4.3470.56_{\pm4.34}79.52±3.7979.52_{\pm3.79}
EEG Conformer75.03±4.9175.03_{\pm4.91}72.90±5.2172.90_{\pm5.21}79.35±7.0379.35_{\pm7.03}78.36±3.8278.36_{\pm3.82}75.35±4.9075.35_{\pm4.90}80.50±2.0680.50_{\pm2.06}
MExFormer83.77±2.27‾\underline{83.77_{\pm2.27}}81.92±2.92‾\underline{81.92_{\pm2.92}}94.88±2.32‾\underline{94.88_{\pm2.32}}80.38±3.70‾\underline{80.38_{\pm3.70}}79.89±3.94‾\underline{79.89_{\pm3.94}}89.42±2.60‾\underline{89.42_{\pm2.60}}

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.