acceptodds
Under review as a conference paper at ICLR 2027

DualBrain: Modeling Bidirectional Spatiotemporal Dependencies in EEG Foundation Models

Abstract

Electroencephalography (EEG) foundation models have recently emerged to overcome the limitations of task-specific decoders, pretraining general-purpose representations that transfer across diverse downstream tasks. However, these architectures do not explicitly maintain both spatial-to-temporal and temporal-to-spatial processing orders as separate pathways within each stage. DualBrain makes this distinction explicit and combines the resulting representations. As a result, existing models fall short of capturing both aspects of spatiotemporal dependency: the temporal evolution of spatial patterns, and the spatial distribution and interaction of temporal patterns. We introduce DualBrain, an EEG foundation model that explicitly models these bidirectional spatiotemporal dependencies through a Multi-view Joint Dynamics Encoder (MJDE), comprising a topographic dynamics branch and a temporal dynamics branch that respectively capture the temporal evolution of spatial patterns and the spatial interactions of temporal patterns, with their complementary representations adaptively integrated through gated fusion. To further ground the model in the physical geometry of the scalp, DualBrain incorporates Spherical Harmonic Positional Encoding (SHPE), which jointly represents electrode positions across both angular dimensions using spherical harmonic bases, enabling structured, montage-independent positional grounding at multiple spatial scales. We pretrain DualBrain on TUH-EEG, the largest currently available single EEG dataset, using a masked autoencoding objective, and evaluate it on 7 tasks across 11 datasets spanning heterogeneous channel configurations and recording conditions. DualBrain achieves the best average performance compared to existing EEG foundation models, with particularly strong gains on seizure detection and mental state identification. These results suggest that explicitly modeling bidirectional spatiotemporal dependencies is a key ingredient for building general-purpose EEG representations.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Electroencephalography (EEG), Foundation model, Self-supervised learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_BwmFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

DualBrain is an EEG foundation model built from two components. (1) The Multi-view Joint Dynamics Encoder (MJDE) runs two parallel attention pathways per stage — Temp→Topo (temporal attention followed by electrode-axis attention) and Topo→Temp (the reverse order) — with their outputs combined by a learned, input-dependent gate, explicitly modeling both the temporal evolution of spatial patterns and the spatial distribution of temporal patterns. (2) Spherical Harmonic Positional Encoding (SHPE) represents electrode positions on the scalp sphere via real spherical harmonic bases across degrees, replacing channel IDs or Cartesian encodings. The model is pretrained with masked autoencoding on TUH (~9,246 h) and evaluated by full fine-tuning on 11 datasets spanning 7 tasks against 3 task-specific models and 6 EEG foundation models. DualBrain reports the best mean balanced accuracy (0.7368 vs 0.7134 for REVE), with the largest gains on seizure detection (+8.1 B-Acc on CHB-MIT, +5.3 on Siena), at 6.33M parameters. Ablations compare encoders (vs LaBraM/CBraMod/CSBrain), processing orders (single-path, fixed-average dual, gated dual), and positional encodings (channel ID, ACPE, 4D PE).

Critical strengths
  • The ablation design is genuinely controlled and matches the claims: Table 3 compares encoder architectures under a common pretraining recipe, and Table 4 isolates the two processing orders (single-path 0.7239/0.7223, fixed dual 0.7221, gated dual 0.7368), so the dual-pathway claim is directly supported by matched comparisons rather than cross-paper numbers. For an empirical architecture paper this discipline is above the norm, and the parameter efficiency (6.33M vs REVE-Base's 69M) is a real practical benefit.
  • The evaluation is broad and honestly calibrated: 11 datasets × 7 tasks, five seeds with standard deviations, reproduced baselines explicitly flagged with †, and the authors openly report where they lose (TUAB, ISRUC B-Acc) rather than only where they win.
Critical weaknesses
  • The SHPE contribution is not established. Table 5 shows no positional encoding dominates per dataset (4D PE leads on TUAB, TUEV, Siena, P-MI; channel-ID on M-AT; ACPE on FACED, ISRUC), and SHPE's mean advantage over 4D PE (~1.1 points) is driven largely by a single small, high-variance dataset (M-AT: 1,707 samples, where 4D PE loses 9.3 points). SHPE's claimed montage-independence is also never directly tested — there is no cross-montage transfer experiment (e.g., pretrain at 19 channels, evaluate on denser/sparser layouts) even though that is the stated motivation. One of the two headline contributions therefore rests on a fragile aggregate.
  • Novelty and gains are both incremental relative to the strength of the claims. The components recombine known ideas (axis-wise/factorized attention orderings from Crossformer/Medformer/CBraMod; spherical Fourier features as in SPE/REVE), and the average lead over REVE is concentrated on seizure datasets while many per-dataset differences fall within one seed standard deviation; Table 2 additionally mixes published numbers (from different pretraining corpora and protocols) with reproduced ones, so the "state-of-the-art" average needs a matched-corpus comparison to be meaningful.
Questions for authors
  • If M-AT is excluded (or with matched seeds), does SHPE still rank first on average over the 11 datasets? Please report the average ordering with and without it.
  • Can you add a direct montage-transfer evaluation to substantiate the montage-independence claim (pretrain 19-ch, fine-tune/test on denser or nonstandard montages)?
  • For Table 2, which entries are reproduced under your preprocessing and splits versus copied from the original papers? A per-row provenance flag would make the average comparable.
6 comments
AuthorsOpenReview ↗

Response to Reviewer BwmF (1/4)

Q1. If M-AT is excluded (or with matched seeds), does SHPE still rank first on average over the 11 datasets? Please report the average ordering with and without it.

Yes. SHPE still ranks first without M-AT, but its margin over 4D-PE is small and not statistically significant, and we have revised our claims accordingly.

Table:PE Performance With and Without M-AT

PEMean B-AccMean RankPaired t-test, Adj. p
SHPE0.7368 / 0.73111.73 / 1.70n/a
4D-PE0.7255 / 0.72682.27 / 2.100.162 / 0.170
Channel-ID0.7270 / 0.71913.09 / 3.300.043 / 0.013
ACPE0.7261 / 0.72182.91 / 2.900.039 / 0.059

Values: all datasets / without M-AT. Two-sided paired t-tests across datasets, Holm-corrected over three comparisons.

Without M-AT, SHPE’s lead over 4D-PE falls from +1.13 to +0.43 pp, so M-AT contributes to the aggregate gap, but the ordering does not change. Paired t-tests give the following results. Against 4D-PE, the difference is not significant, with all 11 datasets (p = 0.162) or without M-AT (p = 0.170). Against Channel-ID, it is significant in both cases (p = 0.043 and p = 0.013). Against ACPE, it is significant with all 11 datasets (p = 0.039) but not without M-AT (p = 0.059). With only 11 datasets, we treat these tests as an approximate summary, not strong evidence, and we do not claim that SHPE significantly outperforms all alternatives on average.

The evidence supports consistency, not per-dataset superiority. SHPE ranks 1st on four datasets, 2nd on six, and 3rd on one, never last, and is never more than 1.18 pp below the best alternative on any dataset. 4D-PE ranks 3rd on three datasets and 4th on two. We now describe SHPE as competitive across heterogeneous tasks without large task-specific degradation.

Clarification of our contribution. We agree that the evidence does not show SHPE significantly outperforming all alternatives, and we will revise the abstract, introduction, and Section 4.2 to replace language implying general superiority. Our intended contribution is robustness across heterogeneous tasks, which matters for a general-purpose EEG model. SHPE attains the best mean B-Acc and mean rank (1.73 vs. 2.27 for the next best, 4D-PE), is never ranked last on any of the 11 datasets, and is never more than 1.18 pp below the best alternative on any dataset. By contrast, 4D-PE ranks 3rd or 4th on five datasets, including a large drop on M-AT (−8.13 pp). We therefore describe SHPE as competitive on every task and the most reliable on average, not as the best encoding on each benchmark. Its practical advantage, that it does not depend on a fixed channel set, is tested directly in Q2.

AuthorsOpenReview ↗

Response to Reviewer BwmF (2/4)

Q2: Direct test of montage independence.

Q2-1. Montage Expansion

We agree that montage independence should be tested directly, since fine-tuning lets the network adapt to each layout. To isolate montage transfer from fine-tuning, we froze each encoder pretrained on 19-channel TUEG and trained only a channel-count-independent linear head per task. We compared SHPE with Channel-ID, ACPE, and 4D-PE, keeping the encoder and head fixed when the test montage changed.

Montage expansion. We trained linear probes on the 19 electrodes shared with TUEG and tested the same frozen encoders and probes on expanded montages, without refitting: FACED (30 ch), PhysioNet-MI (64), SEED-V (62), and Siena (29). SHPE and 4D-PE compute positions from electrode coordinates, so added electrodes receive embeddings automatically. Channel-ID used a predefined embedding table covering all evaluated electrode names; added electrodes had valid IDs, but their embeddings were untrained table entries and were not learned during 19-channel pretraining or probing. This is the natural behavior of a fixed-table method on unseen electrodes, and no method received further adaptation.

DatasetEvaluationChannel-IDACPE4D-PESHPE
FACED (30ch)19ch0.2934±0.00430.3098±0.00120.2502±0.00760.2497±0.0043
Expanded0.2546±0.00750.2303±0.00590.2403±0.00650.2337±0.0045
Change (%)−13.2−25.7−4.0−6.4
P-MI (64ch)19ch0.4306±0.00500.4137±0.00780.3543±0.00790.3442±0.0087
Expanded0.3425±0.00230.2777±0.00450.2833±0.00400.3310±0.0052
Change (%)−20.5−32.9−20.0−3.9
SEED-V (62ch)19ch0.2253±0.00190.2144±0.00400.2203±0.00220.2231±0.0010
Expanded0.2113±0.00270.2103±0.00110.2244±0.00260.2192±0.0024
Change (%)−6.2−1.9+1.9−1.8
Siena (29ch)19ch0.8562±0.00500.8446±0.00400.8479±0.01020.8282±0.0060
Expanded0.8626±0.01180.8476±0.00320.8773±0.00120.8483±0.0022
Change (%)+0.7+0.4+3.5+2.4
Avg.Change (%)−9.8−15.0−4.7−2.4

Linear probes were trained on the 19 channels shared with TUEG using frozen pretrained encoders, then tested on expanded channel sets (FACED: 30, P-MI: 64, SEED-V: 62, Siena: 29) without further adaptation. Each dataset reports B-Acc before and after expansion (mean ± SD), followed by relative change (%).

SHPE has the smallest average relative change (−2.4%), followed by 4D-PE (−4.7%). Compared with Channel-ID and ACPE, SHPE degrades less or improves more on all four datasets (on SEED-V the gap to ACPE is only 0.1 pp). Its advantage over 4D-PE comes from PhysioNet-MI alone (−3.9% vs. −20.0%); 4D-PE degrades less or improves more on FACED, SEED-V, and Siena. Relative change also depends on each method’s 19-channel baseline, so we report absolute B-Acc: after expansion it is highest for Channel-ID on FACED (0.2546) and PhysioNet-MI (0.3425), and for 4D-PE on SEED-V (0.2244) and Siena (0.8773). SEED-V accuracies are close to chance (0.20) and only weakly informative. With four datasets and no significance tests, we treat this as supporting evidence, not a conclusive demonstration.

AuthorsOpenReview ↗

Response to Reviewer BwmF (3/4)

Q2: Direct test of montage independence.

Q2-2. Electrode Dropout and Summary

Electrode dropout. Across all 11 datasets, we trained linear probes on the original electrode configurations and removed 25% of electrodes at test time, keeping encoders and probes fixed. Electrode subsets were matched across PE methods and seeds.

DatasetChannel-IDACPE4D-PESHPE
TUAB0.7878±0.00050.7981±0.00050.8072±0.00130.8100±0.0013
0.7657±0.01320.6458±0.02410.7657±0.02230.7667±0.0293
−2.8%−19.1%−5.1%−5.3%
TUEV0.5848±0.01410.5398±0.01290.5968±0.01180.5658±0.0310
0.5299±0.03530.5344±0.02610.5103±0.05150.5237±0.0515
−9.4%−1.0%−14.5%−7.4%
TUSZ0.8542±0.00250.8738±0.00420.8434±0.00250.8435±0.0027
0.8210±0.01190.8233±0.01440.7898±0.03290.7745±0.0267
−3.9%−5.8%−6.4%−8.2%
CHB-MIT0.7880±0.01030.7843±0.03560.8408±0.00770.8059±0.0272
0.7812±0.02530.7336±0.05220.7224±0.09870.7987±0.0429
−0.9%−6.5%−14.1%−0.9%
Siena0.8562±0.00500.8446±0.00400.8479±0.01020.8282±0.0060
0.8102±0.02770.8413±0.00730.8434±0.01650.8293±0.0144
−5.4%−0.4%−0.5%+0.1%
SEED-V0.2253±0.00190.2144±0.00400.2203±0.00220.2231±0.0010
0.2180±0.00240.2062±0.00180.2193±0.00630.2227±0.0053
−3.2%−3.8%−0.4%−0.2%
FACED0.2934±0.00430.3098±0.00120.2502±0.00760.2497±0.0043
0.2362±0.01520.2118±0.01470.1961±0.01090.1907±0.0139
−19.5%−31.6%−21.7%−23.6%
M-AT0.6757±0.02160.7153±0.01410.6549±0.01950.6535±0.0040
0.6306±0.05350.6618±0.03350.5556±0.02820.5549±0.0309
−6.7%−7.5%−15.2%−15.1%
P-MI0.4306±0.00500.4137±0.00780.3543±0.00790.3442±0.0087
0.3722±0.01640.3177±0.01150.2611±0.01020.3036±0.0260
−13.6%−23.2%−26.3%−11.8%
ISRUC0.7201±0.00960.6966±0.00690.7085±0.00870.6821±0.0301
0.5449±0.02740.3716±0.04610.2855±0.03350.5202±0.1245
−24.3%−46.6%−59.7%−23.7%
HMC0.6091±0.00180.6271±0.00340.6190±0.00370.6221±0.0037
0.4332±0.05670.5016±0.01730.4719±0.02610.5014±0.0386
−28.9%−20.0%−23.8%−19.4%
Avg.0.62050.61980.61300.6026
0.55850.53170.51100.5442
Avg. change−10.8%−15.1%−17.1%−10.5%

Each dataset reports original B-Acc, B-Acc after 25% electrode dropout (mean±SD over five seeds), and relative change (%), in that order. Bold marks the smallest relative degradation per dataset.

SHPE has the smallest average relative degradation (−10.5%), nearly tied with Channel-ID (−10.8%) and below ACPE (−15.1%) and 4D-PE (−17.1%). It has the smallest degradation on 5/11 datasets (6 including the CHB-MIT tie) and the largest on TUSZ. Channel-ID has higher average absolute B-Acc (0.5585 vs. 0.5442) and naturally supports electrode removal through valid remaining IDs, so we treat this test as secondary to montage expansion. SHPE’s advantage over ACPE and 4D-PE partly reflects large ISRUC drops (−46.6% and −59.7%); excluding ISRUC, average degradation is −9.2% (SHPE), −9.4% (Channel-ID), −11.9% (ACPE), and −12.8% (4D-PE).

Summary. SHPE remains competitive in absolute B-Acc and has the smallest average relative decline under both expansion and electrode removal, though the best method varies by dataset. Given the limited evidence, we have replaced “montage-independent” with “independent of electrode names and a fixed channel set” and will add both experiments and dataset-level results to the paper.

Reviewer_SCy7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

DualBrain is an EEG foundation model combining two independently parameterized axial-attention pathways: temporal-then-spatial and spatial-then-temporal. At each of three stages, a learned feature-wise gate combines their outputs. A time-frequency patch tokenizer and spherical harmonic positional encoding (SHPE) represent the signal and electrode coordinates. The model is pretrained by masked waveform reconstruction on approximately 9,246 hours of TUEG and fully fine-tuned on 11 datasets spanning seven tasks. The reported average balanced accuracy is 0.7368, compared with 0.7134 for the strongest complete-model baseline, REVE. Additional experiments replace the encoder or positional encoding within a shared pretraining framework and evaluate single-path and fixed-fusion alternatives.

Critical strengths
  1. Substantial evaluation with unusually useful internal controls. Beyond the 11-dataset complete-model comparison, Tables 3–5 and Appendices D.1–D.3 retain the tokenizer, positional encoding or encoder as appropriate, reconstruction machinery, and pretraining recipe while changing the component of interest. The six-stage single-path variants in Table 24 address a substantial part of the capacity objection. The full model improves mean balanced accuracy over fixed fusion (0.7368 versus 0.7221) and both larger single-path controls (0.7201 and 0.7236). This is meaningful evidence that the combined architecture is effective under the tested recipe.
  2. A compact and reasonably motivated geometric representation. SHPE constructs continuous electrode features from 25 spherical harmonic components rather than a dataset-specific channel lookup. Its positional module is small, and Table 5 reports the highest average balanced accuracy among the evaluated encodings, with no dataset more than 1.18 percentage points below its best alternative. Together with a 6.33M-parameter model and broad task coverage, this makes the design practically interesting even without universal per-dataset superiority.
Critical weaknesses
  1. The central explanation in terms of opposite processing orders is not yet cleanly established. Under the common recipe, the global-attention LaBraM encoder reaches 0.7326 versus 0.7368 for MJDE (Table 3), and actually exceeds MJDE on CHB-MIT and Siena, the two tasks highlighted for large complete-model gains. Thus, those headline gains cannot be attributed specifically to MJDE. The single-path controls are valuable, but neither they nor fixed averaging distinguish a benefit of opposite orders from a benefit of two independently parameterized branches with learned fusion. A same-order, two-branch gated control would address that remaining alternative. Moreover, Appendix D.1 states that the five-seed deviations come from fine-tuning a single pretrained checkpoint per encoder; they do not quantify pretraining variability around the relatively small gap to global attention. This limits the strength of the paper's principal architectural conclusion.
  2. The transfer and superiority claims need a more auditable evaluation boundary. Section 3.1 describes using the processed TUEG corpus but does not document patient/recording exclusion against the TUAB, TUEV, and TUSZ evaluation sets. This is an unresolved overlap risk, not evidence that leakage occurred. Appendix C.1 also mixes reproduced baselines with published results, and C.3 explicitly identifies different TUAB input constructions for DualBrain and REVE. The CHB-MIT subsection does not specify the exact train/validation/test patient assignment. These details matter because the aggregate comparison combines small advantages with a few large gains. The CHB-MIT result also illustrates why balanced accuracy alone is insufficient: DualBrain improves it from 0.7450 to 0.8261 against REVE, while AUROC decreases from 0.9301 to 0.9180 and AUPRC from 0.5301 to 0.4213 (Table 13). The broad phrase “alongside AUROC gains” in Section 4.1 needs qualification for this comparison.
Questions for authors
  1. Clarification / evaluation validity: What patient- and recording-level exclusions separate TUEG pretraining from downstream TUAB, TUEV, and TUSZ evaluation? Please provide the audit procedure and counts, together with the exact CHB-MIT patient split, including treatment of repeated recordings from the same patient. A demonstrated disjoint boundary would remove an important uncertainty.
  2. Assessment-changing control: Does a two-branch model with the same attention order in both branches, independently parameterized and combined with the same gates, match MJDE under the same block count and training recipe? This would distinguish opposite-order complementarity from generic gated branch capacity. The existing six-stage single-path controls do not answer this particular question.
  3. Clarification / robustness of the comparison: How stable is the 0.0042 average advantage over the common-framework global-attention encoder across pretraining seeds? For the seizure comparisons, can you report validation-selected operating thresholds and sensitivity/specificity alongside AUPRC, and identify which published baselines used exactly matching inputs and splits? These results would affect the claimed magnitude and practical meaning of the gain.
5 comments
AuthorsOpenReview ↗

Response to Reviewer SCy7 (1/3)

Q1: Evaluation boundary. Thank you. We ran an explicit overlap audit and patient-level exclusion pretraining.

Audit. Our original pretraining used the processed TUEG corpus without excluding downstream subjects, as in prior EEG foundation models such as CBraMod and CodeBrain. We matched patient and recording/session identifiers between TUEG and each downstream set. Overlapping patients: TUAB 2,207/2,329 (95%), TUEV 237/290 (82%), TUSZ 592/675 (88%). Overlapping recordings: TUAB 2,615/2,993 (87%) and TUSZ 3,286/7,725 (43%). Overlapping sessions: TUAB 2,630/2,993 (88%) and TUSZ 1,189/1,642 (72%). For TUAB, 15 sessions match TUEG although their specific recordings do not, explaining the difference between the two counts. Because the available TUEV files lack TUEG-compatible recording and session IDs, we report overlap by patient ID only. The original results therefore do not have a disjoint boundary, and the same holds for other models pretrained on TUEG such as CBraMod, CSBrain, and CodeBrain also face this potential overlap.

Patient-level exclusion. We pretrained three separate models, each removing all TUEG recordings from patients in the corresponding downstream set, with one pretraining seed (42), and evaluated each over five fine-tuning seeds (42, 696, 1001, 1234, 3407), without dataset-specific seed selection.

DatasetMethodB-AccAUPRC / KappaAUROC / F1-w
TUABLaBraM-Base (excl.)0.8140 ± 0.00190.8965 ± 0.00160.9022 ± 0.0009
CBraMod (excl.)0.8249 ± 0.00250.9221 ± 0.00150.9156 ± 0.0017
CodeBrain (excl.)0.8288 ± 0.00640.9061 ± 0.00390.9012 ± 0.0020
DualBrain (original)0.8100 ± 0.00730.8977 ± 0.00840.8905 ± 0.0094
DualBrain (excl.)0.8086 ± 0.00400.8966 ± 0.00740.8882 ± 0.0091
TUEVLaBraM-Base (excl.)0.6409 ± 0.00650.6637 ± 0.00930.8312 ± 0.0052
CBraMod (excl.)0.6659 ± 0.01240.6744 ± 0.01210.8331 ± 0.0071
CodeBrain (excl.)0.6375 ± 0.01820.6838 ± 0.02910.8293 ± 0.0163
DualBrain (original)0.6978 ± 0.01040.6979 ± 0.01000.8430 ± 0.0045
DualBrain (excl.)0.6815 ± 0.00980.6863 ± 0.02860.8354 ± 0.0138
TUSZDualBrain (original)0.9279 ± 0.01870.9960 ± 0.00130.9852 ± 0.0047
DualBrain (excl.)0.9098 ± 0.01810.9921 ± 0.00380.9703 ± 0.0148

TUAB is essentially unchanged (−0.14 pp B-Acc, within seed std). TUEV and TUSZ decrease by 1.63 and 1.81 pp. On TUEV, DualBrain (0.6815) remains highest among the available exclusion-based results (CBraMod 0.6659, +1.56 pp). That margin is larger than the seed std, but the Kappa and weighted-F1 margins are within one std, and the baseline numbers come from other papers and may differ in protocol. On TUSZ, DualBrain reaches 0.9098 B-Acc, below REVE’s 0.9197, whose checkpoint we could not verify as patient-disjoint from TUSZ. We have no exclusion-based baseline results for TUSZ, so this comparison is not matched. After exclusion, DualBrain remains competitive with a compact model (6.03M encoder / 6.33M backbone parameters, vs. about 69M for REVE-Base and 15.17M for CodeBrain), while TUEV and TUSZ lose accuracy.

CHB-MIT split. We followed the CBraMod preprocessing pipeline, also used by CSBrain. Cases chb12, chb13, and chb17 were excluded in preprocessing. The remaining 21 cases were split into training (chb01–chb11, chb14–chb16, chb18–chb20), validation (chb21–chb22), and test (chb23–chb24). All recordings and segments from a case stay in one partition. The split is recording-disjoint but not strictly patient-disjoint: chb01 (training) and chb21 (validation) are the same person, recorded about 1.5 years apart. This affects validation only, and no test case shares a patient with training, so test results are unaffected, but checkpoint selection could be slightly influenced. We will add the audit, exclusion protocol, results, and split to Section 3 and the Appendix.

AuthorsOpenReview ↗

Response to Reviewer SCy7 (2/3)

Q2. Assessment-changing control

We thank the reviewer for suggesting this control. A related concern was raised by another reviewer, and we ran an additional ablation to separate opposite-order complementarity from generic gated two-branch capacity.

Control. We compared MJDE (S→T/T→S) against two independently parameterized same-order dual-branch controls (S→T/S→T and T→S/T→S), matching gating mechanism, block count, parameters (6.33M backbone; 6.03M encoder), FLOPs (6.974G), and pretraining budget (40 epochs; 86,680 steps). All models used one pretraining seed and five fine-tuning seeds.

Does a same-order model match MJDE? Nearly, on average. The two controls reach 0.7289 mean B-Acc across the 11 datasets, vs. 0.7368 for DualBrain (+0.79 pp). Most of the margin comes from Mental Arithmetic (+6.26 pp over the better control). Excluding M-AT, the margin is +0.24 pp (0.7311 vs. 0.7287), within seed variability, so we do not claim a statistically significant advantage for opposite orders.

DatasetMetricS→T/S→TT→S/T→SOurs
TUABB-Acc0.8141 ± 0.01100.8040 ± 0.00530.8100 ± 0.0073
AUPRC0.9020 ± 0.01180.8909 ± 0.00720.8977 ± 0.0084
AUROC0.8959 ± 0.01470.8804 ± 0.00800.8905 ± 0.0094
TUEVB-Acc0.6788 ± 0.01200.6846 ± 0.01470.6978 ± 0.0104
Kappa0.6851 ± 0.02390.6952 ± 0.01010.6979 ± 0.0100
F1-W0.8357 ± 0.01310.8399 ± 0.00720.8430 ± 0.0045
TUSZB-Acc0.9123 ± 0.01180.8992 ± 0.01360.9279 ± 0.0187
AUPRC0.9950 ± 0.00160.9925 ± 0.00280.9960 ± 0.0013
AUROC0.9815 ± 0.00620.9726 ± 0.01020.9852 ± 0.0047
CHB-MITB-Acc0.8103 ± 0.02650.8426 ± 0.02730.8261 ± 0.0348
AUPRC0.3632 ± 0.01610.4277 ± 0.02020.4213 ± 0.0172
AUROC0.9015 ± 0.01180.9116 ± 0.01890.9180 ± 0.0064
SienaB-Acc0.8451 ± 0.01770.8576 ± 0.01660.8410 ± 0.0114
AUPRC0.4002 ± 0.04410.4316 ± 0.03530.4557 ± 0.0637
AUROC0.9299 ± 0.00340.9371 ± 0.00370.9354 ± 0.0025
SEED-VB-Acc0.4204 ± 0.00170.4154 ± 0.00810.4199 ± 0.0034
Kappa0.2798 ± 0.00230.2725 ± 0.01250.2781 ± 0.0045
F1-W0.4271 ± 0.00260.4214 ± 0.01170.4257 ± 0.0040
FACEDB-Acc0.6330 ± 0.00530.6184 ± 0.01690.6082 ± 0.0052
Kappa0.5855 ± 0.00480.5696 ± 0.01850.5578 ± 0.0061
F1-W0.6360 ± 0.00470.6230 ± 0.01440.6124 ± 0.0058
Mental ArithmeticB-Acc0.7312 ± 0.01210.7306 ± 0.03940.7938 ± 0.0393
AUPRC0.8136 ± 0.06140.8224 ± 0.02150.8130 ± 0.0365
AUROC0.8921 ± 0.03980.9119 ± 0.01290.8769 ± 0.0152
PhysioNet-MIB-Acc0.6350 ± 0.00710.6382 ± 0.00240.6503 ± 0.0034
Kappa0.5133 ± 0.00950.5176 ± 0.00320.5337 ± 0.0046
F1-W0.6359 ± 0.00730.6391 ± 0.00300.6510 ± 0.0036
ISRUCB-Acc0.7950 ± 0.00700.7884 ± 0.00490.7883 ± 0.0033
Kappa0.7456 ± 0.01500.7417 ± 0.01150.7480 ± 0.0093
F1-W0.8035 ± 0.01320.8017 ± 0.00920.8068 ± 0.0067
HMCB-Acc0.7427 ± 0.00110.7388 ± 0.00430.7414 ± 0.0042
Kappa0.6962 ± 0.00200.6881 ± 0.00640.6920 ± 0.0052
F1-W0.7622 ± 0.00160.7565 ± 0.00450.7604 ± 0.0041
AverageB-Acc0.72890.72890.7368

Cross-task consistency. We measured each model’s B-Acc shortfall from the best of the three configurations on each dataset. DualBrain’s mean and maximum shortfalls are 0.64 and 2.48 pp, vs. 1.43 and 6.26 pp for S→T/S→T and 1.43 and 6.32 pp for T→S/T→S. Excluding M-AT, the mean shortfalls are 0.71, 0.95, and 0.94 pp, so the advantage persists but is smaller. This indicates a more balanced profile, not uniform superiority. FACED is the clearest exception, where DualBrain trails both controls in B-Acc, Kappa, and weighted F1. T→S/T→S also outperforms DualBrain on CHB-MIT and Siena.

Interpretation and revision. These results are consistent with gated integration of two differently ordered views giving more balanced performance across heterogeneous tasks than repeating either order. They do not establish representational complementarity or the necessity of opposite orders, and they do not capture pretraining-seed variability (see Q3). We will add the same-order controls to the ablation table and revise the abstract, introduction, and ablation text to state this scope.

AuthorsOpenReview ↗

Response to Reviewer SCy7 (3/3)

Q3. Clarification / robustness of the comparison

Thank you. We address the three parts in turn.

Pretraining-seed stability. We pretrained both DualBrain and the LaBraM-style encoder with seeds 42, and 43 under the same recipe and evaluated the same five fine-tuning seeds on all 11 datasets. DualBrain has the higher mean B-Acc on both seeds (0.7368 vs. 0.7326 on seed 42, +0.42 pp,; 0.7202 vs. 0.7130 on seed 43, +0.72 pp), and the mean over the two seeds is 0.7285 vs. 0.7228 (+0.57 pp). The two encoders move together across seeds, so the paired gap is the more informative comparison. Two pretraining seeds are not enough to establish seed robustness or statistical reliability, so we describe DualBrain as matching or modestly exceeding the global-attention encoder under the common recipe and present this as limited evidence. We will add the results and protocol to the appendix.

Pretraining seedDualBrainLaBraM
420.7368 ± 0.00410.7326 ± 0.0048
430.7202 ± 0.00650.7130 ± 0.0056
Avg. ± SD0.7285 ± 0.00830.7228 ± 0.0098

Operating point and seizure metrics. Binary predictions use a fixed probability threshold of 0.5. We select checkpoints by validation B-Acc but do not tune the threshold, which is also the rule in the released CBraMod and CSBrain evaluators; the released REVE evaluator uses a fixed class decision (argmax). Sensitivity and specificity were computed from saved test predictions at this operating point (mean ± SD over five fine-tuning seeds):

DatasetThresholdSensitivitySpecificityAUPRC
CHB-MIT0.50.6897 ± 0.09200.9625 ± 0.02780.4213 ± 0.0172
Siena0.50.6963 ± 0.02100.9859 ± 0.00360.4557 ± 0.0637
TUSZ0.50.8719 ± 0.03670.9838 ± 0.01120.9960 ± 0.0013

Because B-Acc depends on the operating point, we also report threshold-free metrics. On CHB-MIT, DualBrain’s balanced accuracy is higher than REVE’s, but its AUROC (0.9180 vs. 0.9301) and AUPRC (0.4213 vs. 0.5301) are lower. We therefore corrected the Section 4.1 text, which described AUROC gains, and no longer describe the seizure comparison as a consistent gain.

Inputs and splits for seizure comparisons. We used established preprocessing pipelines for CHB-MIT and Siena. For CHB-MIT, we followed CBraMod’s 16-channel bipolar preprocessing, 10-second windows, case exclusions, and train/validation/test split. CSBrain and CodeBrain report the same input dimensions and sample count, and CSBrain directs users to CBraMod’s preprocessing, so these are the closest documented comparisons. For Siena, we used the unmodified CSBrain preprocessing and split, making CSBrain the directly matched published comparison; CBraMod’s Siena score comes from the CSBrain paper, not the original CBraMod paper. For TUSZ, we defined one protocol and applied it to the baselines we reproduced; copied results are not presented as protocol-matched.

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

This paper introduces DualBrain, a foundation model for EEG data. Its key components include:

(1) a Multi-view Joint Dynamics Encoder (MJDE), which models data through both spatial-to-temporal and temporal-to-spatial sequences, integrating information at each stage via a gating mechanism;

(2) Spherical Harmonic Positional Encoding (SHPE), based on the spherical coordinates of electrodes on the scalp;

(3) a self-supervised pre-training framework using geometric block masking and reconstruction on the TUEG dataset.

The model was pre-trained on approximately 1.11 million 30-second EEG segments (totaling about 9,246 hours) and evaluated across 11 downstream datasets covering seven task categories, including anomaly detection, seizure detection, emotion recognition, motor imagery, and sleep stage classification.

Critical strengths
  1. Robust performance. The 11 datasets encompass a wide range of downstream EEG tasks, demonstrating practical benefits that extend beyond any single task. This serves as the primary empirical evidence supporting acceptance.
  2. Effective ablation studies. The authors evaluate channel IDs, ACPE, and 4D positional encodings, while also providing additional comparisons involving fixed fusion and a six-stage single-path architecture.
Critical weaknesses
  1. The causal claim that "a bidirectional processing sequence is essential" has not been rigorously isolated. Neither the three-stage single-path nor the dual-path architectures are strictly equivalent to the baseline in terms of parameter count or computational operations; while the six-stage single-path architecture is roughly parameter-equivalent, it introduces a deeper serial network. Consequently, the current data more directly support the conclusion that "a complete gated architecture performs better on average," rather than establishing the independent necessity of each individual design component.

  2. Performance gains vary across datasets, and mean values ​​may mask specific shortcomings; the relationship between pre-training and downstream tasks within the TUH family requires further clarification. On the TUAB dataset, the model achieves a balanced accuracy of 0.8100, which is lower than the 0.8315 and 0.8289 reported for REVE and CBraMod, respectively. On the Siena dataset, the SHPE model’s score of 0.8410 is outperformed by several alternative positional encoding methods (page 9, Table 5). Given that TUEG is used for pre-training, while TUAB, TUEV, and TUSZ all belong to the TUH family, the paper should explicitly state whether data overlap—at the level of individual recordings or subjects—has been strictly eliminated. For reported marginal gains, it is advisable to include details such as matching random seeds, confidence intervals, or results from statistical significance tests.

Questions for authors
  1. Please provide an audit of the overlap in subject IDs and recording/session IDs between the TUEG pre-training corpus and the downstream subsets (TUAB, TUEV, and TUSZ), along with the corresponding exclusion strategy. Is it possible to report the results after performing complete deduplication?
  2. Could you construct control experiments—comparing single-path, dual-path, and gated architectures—where the number of parameters, FLOPs, serial depth, and pre-training budget are kept as consistent as possible, in order to assess the incremental value contributed by each path?
  3. Regarding Table 2 and the ablation study on positional encoding, could you provide paired error metrics (using the same random seed), cross-task statistics, and total inference costs? Additionally, please explain the cases—such as TUAB and the Siena dataset—where the proposed method does not achieve the best performance.
5 comments
AuthorsOpenReview ↗

Response to Reviewer MueH (1/3)

Q1: Overlap between TUEG pretraining and TUAB/TUEV/TUSZ. Thank you; we agree this needs to be stated explicitly. Our original pretraining followed the protocols of prior EEG foundation models such as CBraMod and CodeBrain, using the processed TUEG corpus, which did not exclude the downstream sets. We therefore ran two analyses.

(1) ID-level audit. We matched patient and recording/session IDs between TUEG and each downstream set. Overlapping patients: TUAB 2,207/2,329 (95%), TUEV 237/290 (82%), TUSZ 592/675 (88%). Overlapping recordings: TUAB 2,615/2,993 (87%) and TUSZ 3,286/7,725 (43%). Overlapping sessions: TUAB 2,630/2,993 (88%) and TUSZ 1,189/1,642 (72%). For TUAB, 15 sessions match TUEG although their specific recordings do not, explaining the difference between the two counts. Because the available TUEV files lack TUEG-compatible recording and session IDs, we report TUEV overlap by patient ID only. The original pretraining therefore does not have a disjoint boundary with these three sets.

(2) Pretraining with overlapping patients excluded. We pretrained three separate models, each removing all TUEG recordings from patients in the corresponding downstream set, with one pretraining seed (42), and evaluated each over five fine-tuning seeds (42, 696, 1001, 1234, 3407), without dataset-specific seed selection. Results are mean ± std; baseline “excl.” rows are taken from LaBraM, CBraMod, and CodeBrain and may differ in protocol.

DatasetMethodB-AccAUPRC / KappaAUROC / F1-w
TUABLaBraM-Base (excl.)0.8140 ± 0.00190.8965 ± 0.00160.9022 ± 0.0009
CBraMod (excl.)0.8249 ± 0.00250.9221 ± 0.00150.9156 ± 0.0017
CodeBrain (excl.)0.8288 ± 0.00640.9061 ± 0.00390.9012 ± 0.0020
DualBrain (original)0.8100 ± 0.00730.8977 ± 0.00840.8905 ± 0.0094
DualBrain (excl.)0.8086 ± 0.00400.8966 ± 0.00740.8882 ± 0.0091
TUEVLaBraM-Base (excl.)0.6409 ± 0.00650.6637 ± 0.00930.8312 ± 0.0052
CBraMod (excl.)0.6659 ± 0.01240.6744 ± 0.01210.8331 ± 0.0071
CodeBrain (excl.)0.6375 ± 0.01820.6838 ± 0.02910.8293 ± 0.0163
DualBrain (original)0.6978 ± 0.01040.6979 ± 0.01000.8430 ± 0.0045
DualBrain (excl.)0.6815 ± 0.00980.6863 ± 0.02860.8354 ± 0.0138
TUSZDualBrain (original)0.9279 ± 0.01870.9960 ± 0.00130.9852 ± 0.0047
DualBrain (excl.)0.9098 ± 0.01810.9921 ± 0.00380.9703 ± 0.0148

TUAB is essentially unchanged (B-Acc −0.14 pp, within seed std), and its ranking is unchanged: DualBrain trails the three baselines with or without exclusion. TUEV drops by 1.63 pp B-Acc. DualBrain (excl.) still has the highest B-Acc, Kappa, and weighted F1 among the listed models, but only the B-Acc margin (+1.56 pp over CBraMod) is clearly larger than the seed std; the Kappa and F1-w differences are within one std, so we regard them as comparable. TUSZ drops by 1.81 pp B-Acc. After exclusion, DualBrain (0.9098) remains above CodeBrain (0.8934) and CBraMod (0.8808) in Table 2 but below REVE (0.9197). REVE reports removing recordings used in its own downstream tasks, which include TUAB and TUEV but not TUSZ, and does not report a TUSZ-specific or patient-level exclusion, so we cannot verify that its released checkpoint is disjoint from TUSZ. We therefore keep REVE’s score as a reference, not as a pretraining-matched comparison. After patient exclusion, DualBrain remains competitive with a compact model (6.03M encoder parameters, 6.33M backbone, vs. about 69M for REVE-Base’s encoder), but we cannot make a controlled accuracy comparison on TUSZ.

These results suggest that the TUAB and TUEV trends are not an artifact of dataset-level overlap, although TUEV and TUSZ lose accuracy and the TUSZ comparison with REVE is unmatched. The reported variance comes from fine-tuning a single pretrained checkpoint. We will add the audit, the exclusion protocol, and these results to Section 3 and the appendix, and revise the TUSZ comparison in Section 4.1.

AuthorsOpenReview ↗

Response to Reviewer MueH (2/3)

Q2. Thank you. We agree that stacking single-path blocks increases serial depth, so it cannot isolate each path’s contribution. We therefore report two sets of comparisons.

(1) Gated dual-path vs. single-path and fixed fusion (existing ablations). Mean B-Acc across the 11 datasets is 0.7368 for DualBrain, versus 0.7221 for fixed (non-gated) fusion and 0.7201 and 0.7236 for the two larger six-stage single-path controls (+1.3 to +1.7 pp). Learned gating over two pathways is associated with higher mean B-Acc than these alternatives, which are deeper or larger.

DatasetMetricS→TT→SOurs
TUABB-Acc0.8141±0.01100.8040±0.00530.8100±0.0073
AUPRC0.9020±0.01180.8909±0.00720.8977±0.0084
AUROC0.8959±0.01470.8804±0.00800.8905±0.0094
TUEVB-Acc0.6788±0.01200.6846±0.01470.6978±0.0104
Kappa0.6851±0.02390.6952±0.01010.6979±0.0100
F1-W0.8357±0.01310.8399±0.00720.8430±0.0045
TUSZB-Acc0.9123±0.01180.8992±0.01360.9279±0.0187
AUPRC0.9950±0.00160.9925±0.00280.9960±0.0013
AUROC0.9815±0.00620.9726±0.01020.9852±0.0047
CHB-MITB-Acc0.8103±0.02650.8426±0.02730.8261±0.0348
AUPRC0.3632±0.01610.4277±0.02020.4213±0.0172
AUROC0.9015±0.01180.9116±0.01890.9180±0.0064
SienaB-Acc0.8451±0.01770.8576±0.01660.8410±0.0114
AUPRC0.4002±0.04410.4316±0.03530.4557±0.0637
AUROC0.9299±0.00340.9371±0.00370.9354±0.0025
SEED-VB-Acc0.4204±0.00170.4154±0.00810.4199±0.0034
Kappa0.2798±0.00230.2725±0.01250.2781±0.0045
F1-W0.4271±0.00260.4214±0.01170.4257±0.0040
FACEDB-Acc0.6330±0.00530.6184±0.01690.6082±0.0052
Kappa0.5855±0.00480.5696±0.01850.5578±0.0061
F1-W0.6360±0.00470.6230±0.01440.6124±0.0058
MentalArithmeticB-Acc0.7312±0.01210.7306±0.03940.7938±0.0393
AUPRC0.8136±0.06140.8224±0.02150.8130±0.0365
AUROC0.8921±0.03980.9119±0.01290.8769±0.0152
PhysioNetMIB-Acc0.6350±0.00710.6382±0.00240.6503±0.0034
Kappa0.5133±0.00950.5176±0.00320.5337±0.0046
F1-W0.6359±0.00730.6391±0.00300.6510±0.0036
ISRUCB-Acc0.7950±0.00700.7884±0.00490.7883±0.0033
Kappa0.7456±0.01500.7417±0.01150.7480±0.0093
F1-W0.8035±0.01320.8017±0.00920.8068±0.0067
HMCB-Acc0.7427±0.00110.7388±0.00430.7414±0.0042
Kappa0.6962±0.00200.6881±0.00640.6920±0.0052
F1-W0.7622±0.00160.7565±0.00450.7604±0.0041
Avg.B-Acc0.7289 ± 0.13460.7289 ± 0.13730.7368 ± 0.1388

(2) Same-order gated controls (new). To separate opposite-order complementarity from generic gated two-branch capacity (also raised by Reviewer SCy7), we added two gated dual-path controls with independently parameterized branches using the same attention order (S→T/S→T and T→S/T→S). Both match DualBrain in gating mechanism, parameters (6.03M encoder / 6.33M backbone), FLOPs (6.974G), serial depth, and pretraining budget (40 epochs, 86,680 steps). All models use one pretraining seed, and the reported std is over five fine-tuning seeds.

DualBrain has the highest mean B-Acc (0.7368 vs. 0.7289 for each control), but much of the margin comes from M-AT (+6.26 pp over the better control). Excluding M-AT, the advantage narrows to 0.24 pp (0.7311 vs. 0.7287), so we do not claim that opposite orders are independently superior. As a descriptive, post-hoc check, we measured each model’s B-Acc shortfall from the best of the three on each dataset. DualBrain’s mean and maximum shortfalls are 0.64 and 2.48 pp, vs. 1.43 and 6.26 pp for S→T/S→T and 1.43 and 6.32 pp for T→S/T→S. Excluding M-AT, the mean shortfalls are 0.71, 0.95, and 0.94 pp, so the advantage persists but is smaller. This indicates a more balanced profile across these tasks, not uniform superiority. FACED is the clearest exception, where DualBrain trails both controls on B-Acc, Kappa, and weighted F1, and T→S/T→S also leads on CHB-MIT and Siena.

The results are consistent with gated integration of two differently ordered views giving more balanced performance across heterogeneous tasks than repeating either order. This is an observed cross-task pattern, not proof that the representations are complementary or that opposite orders are necessary. We will revise the abstract, introduction, and ablation text to reflect this scope and add the same-order controls to the ablation table.

AuthorsOpenReview ↗

Response to Reviewer MueH (3/3)

Q3. Thank you. We separate seed-paired comparisons (positional-encoding ablation) from dataset-level comparisons (Table 2).

1. Positional-encoding ablation

Paired comparisons. We compare SHPE with channel-ID, ACPE, and 4D PE using the same five fine-tuning seeds (42, 696, 1001, 1234, 3407). The table reports the mean within-seed difference ΔB-Acc = SHPE − alternative, in pp; positive values indicate lower balanced error (1 − B-Acc).

Datasetvs. channel-IDvs. ACPEvs. 4D PE
TUAB0.790.81−0.03
TUEV2.641.37−0.37
TUSZ2.261.991.76
CHB-MIT1.592.730.17
Siena−0.891.56−0.99
SEED-V1.051.391.20
FACED0.72−0.261.76
M-AT−1.182.508.13
P-MI2.080.81−0.16
ISRUC0.77−1.140.48
HMC0.990.050.46

Cross-task statistics. We use exploratory two-sided dataset-paired t-tests across 11 datasets, Holm-corrected over three comparisons.

ComparisonMean ΔB-Acc (pp) [95% CI]Wins / lossesHolm-Adj. p
SHPE vs. channel-ID+0.98 [0.18, 1.79]9 / 20.0425
SHPE vs. ACPE+1.07 [0.28, 1.87]9 / 20.0387
SHPE vs. 4D PE+1.13 [−0.54, 2.79]7 / 40.1618

Confidence intervals are unadjusted and reflect variation across datasets, not fine-tuning seeds. SHPE’s mean improvement over channel-ID and ACPE is significant under this test, and over 4D PE it is not. With only 11 datasets, we treat these tests as exploratory.

Consistency across datasets. SHPE trails 4D PE on four datasets by less than 1 pp (TUAB: 0.03, TUEV: 0.37, Siena: 0.99, and PhysioNet-MI: 0.16 pp), but leads by 8.13 pp on M-AT and 1.76 pp on TUSZ and FACED. Relative to the best PE on each dataset, SHPE achieves a lower mean shortfall (0.38 vs. 1.50 pp) and maximum shortfall (1.18 vs. 9.31 pp) than 4D-PE. These descriptive results support SHPE's cross-task consistency rather than superiority on every dataset, which we will clarify in the revised manuscript.

2. Table 2

Matched seed-level results are unavailable for Table 2. We therefore compare DualBrain with each baseline using dataset-level B-Acc differences and apply two-sided exact sign-flip tests across 11 datasets, with Holm correction over nine comparisons.

BaselineMean ΔB-Acc (pp)Wins/lossesAdj. p
EEGNet16.0911/00.0088
ContraWR13.6011/00.0088
ST-Transformer15.1411/00.0088
BIOT7.5211/00.0088
LaBraM5.3210/10.0146
CBraMod3.7810/10.0273
CSBrain2.949/20.0273
REVE2.3410/10.0273
CodeBrain4.8310/10.0273

DualBrain achieves higher mean B-Acc than all nine baselines (adjusted p ≤ 0.0273). Against REVE, it gains 2.34 pp across 11 datasets (10 wins). As Table 2 mixes reproduced (†) and published results, this reflects cross-benchmark consistency, not matched conditions. We also test reproduced entries separately.

BaselineReproduced datasetsDualBrain B-AccBaseline B-AccΔ(pp)
REVETUSZ,CHB-MIT,Siena,SEED-V0.75370.70994.38
CodeBrainTUSZ,Siena,P-MI,HMC0.79020.70798.22

On reproduced entries, DualBrain outperforms two recent EEG foundation models, REVE and CodeBrain, on all four reproduced datasets (+4.38 and +8.22 pp, respectively). Neither comparison reaches statistical significance (two-sided exact p = 0.125); these results highlight DualBrain's parameter efficiency: its 6.33M parameters are substantially fewer than REVE-Base (69M) and CodeBrain (15.17M), suggesting a favorable performance–parameter trade-off despite the limited datasets and protocol differences.

3. Inference cost. Complexity is reported in Appendix C.1, Table 8. DualBrain has 6.33M parameters (vs. CSBrain: 8.8M, CodeBrain: 15.2M, REVE-Base: 69.0M) and 8.729G FLOPs (vs. 8.450G, 8.740G, and 104.440G), measured on 19-ch, 30-s inputs at 200 Hz. Across 11 datasets, the four PE variants have similar median downstream model sizes (70.73–70.76M) and FLOPs (2.124–2.133G). On a single A100 (batch size 1), median inference times are 6.64 ms (SHPE), 5.04 ms (channel-ID), 4.95 ms (ACPE), and 5.22 ms (4D PE).

4. Cases where DualBrain is not best

TUAB. DualBrain (0.8100 B-Acc) trails REVE (0.8315) and CBraMod (0.8289) by ~2 pp, exceeding its seed std (±0.0073). PE differences in our TUAB ablation are approximately 0.8 pp or less, so PE alone cannot explain the gap. Protocols also differ in preprocessing (DualBrain vs. REVE; Appendix C.3) and classification heads: flattened tokens (DualBrain/CBraMod), normalized flatten-based linear readout (REVE), and mean pooling (LaBraM). Their effects remain unisolated, which we will acknowledge in Section 4.1.

Siena. In the PE ablation, SHPE trails channel-ID and 4D PE by 0.89 and 0.99 pp, respectively, within approximately one seed std (±0.0114). With only 14 subjects and few seizure windows, the PE ranking remains uncertain. DualBrain (0.8410 B-Acc) leads all baselines in Table 2 but trails T→S/T→S (0.8576) in the same-order ablation, as reported above.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Final Summary of Rebuttal

Dear Area Chairs and Program Chairs,

Thank you for the discussion. DualBrain is a compact EEG foundation model combining temporal-then-spatial and spatial-then-temporal attention (MJDE) with spherical harmonic positional encoding (SHPE). We summarize what the reviews and our new experiments establish, including where results are mixed.

1. Performance and cost. Across 11 datasets spanning 7 tasks, DualBrain has the highest mean B-Acc in Table 2 (0.7368 vs. 0.7134 for REVE-Base), in a broad comparison that mixes reproduced and published results. It uses 6.33M parameters vs. 69.0M and 8.73G vs. 104.44G FLOPs (19-channel, 30-second input).

2. One architecture across tasks. Two same-order controls, matched in gating, parameters, FLOPs, depth, and budget, have mean B-Acc 0.7289 vs. 0.7368 for MJDE; the margin is small (+0.24 pp without Mental Arithmetic) and uses one pretraining seed. The best order varies by dataset (repeated temporal-then-spatial on CHB-MIT and Siena, the reverse on FACED, where MJDE trails by up to 2.48 points). MJDE stays within 2.48 points of the best on every dataset, vs. up to 6.26 and 6.32 for the controls.

3. One positional encoding across settings. SHPE has the highest mean fine-tuning B-Acc and stays within 1.18 points of the best encoding on every dataset. With the encoder and linear probe fixed, it has the smallest mean relative loss under electrode expansion (−2.4% vs. −9.8% for Channel-ID) and is essentially tied with Channel-ID under removal (−10.5% vs. −10.8%, with Channel-ID higher in absolute B-Acc). 4D-PE retains more performance on three of four expansion datasets.

4. Patient-exclusion pretraining. After excluding all downstream patients, B-Acc changes by −0.14 (TUAB), −1.63 (TUEV), and −1.81 points (TUSZ). DualBrain remains highest among available exclusion-based results on TUEV (0.6815); on TUSZ (0.9098) it is below REVE’s reported 0.9197, whose TUSZ exclusion is undocumented.

5. Ablation scope. For context, prior EEG foundation models typically report component ablations on selected benchmarks: CodeBrain compares tokenizer and module variants on 3 of 10 datasets (Section 3.4, Table 3) [1]; REVE compares positional encodings on 2 of 10 (Appendix D, Table 19) [2]; CSBrain evaluates tokenization and attention designs on 4 of 16 (Section 3.3, Figs. 4-5) and regional descriptors on 4 of 16 (Appendix E.3, Fig. 11) [3]; and CBraMod compares attention mechanisms and positional encodings on 4 of 12 (Section 3.4, Figs. 4-5) [4]. Because a component’s advantage can differ across datasets, we report our encoder and positional-encoding ablations under full fine-tuning on all 11 downstream datasets, including every case in which an alternative performed better, so readers can see where each design holds, where it reverses, and by how much.

Taken together, the ablations and the additional experiments in points 2 and 3 show that DualBrain stayed close to the best design on each dataset across the tasks and adaptation settings we evaluated, even as different designs led in individual settings. This matters for an EEG foundation model, whose architecture and positional encoding must be chosen before the downstream task, electrode layout, and opportunity for fine-tuning are known. We understand the remaining concern is the size of the contribution, and we have scoped our claims accordingly. We hope the full results help future researchers make these choices, and we thank you for your time.

References

[1] Ma et al. “CodeBrain: Bridging Decoupled Tokenizer and Multi-Scale Architecture for EEG Foundation Model.” International Conference on Learning Representations (ICLR), 2026.

[2] El Ouahidi et al. “REVE: A Foundation Model for EEG – Adapting to Any Setup with Large-Scale Pretraining on 25,000 Subjects.” Advances in Neural Information Processing Systems (NeurIPS), 2025.

[3] Zhou et al. “CSBrain: A Cross-Scale Spatiotemporal Brain Foundation Model for EEG Decoding.” Advances in Neural Information Processing Systems (NeurIPS), 2025.

[4] Wang et al. “CBraMod: A Criss-Cross Brain Foundation Model for EEG Decoding.” International Conference on Learning Representations (ICLR), 2025.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.