acceptodds
Under review as a conference paper at ICLR 2027

UniPopT: Where Does Intracranial EEG Help a Brain-Signal Foundation Model?

Abstract

Intracranial EEG (iEEG) is often described as an untapped resource for brain-signal foundation models, yet whether it improves decoding has not been established: the one prior report of a gain from joint EEG–iEEG pre-training used a model that carries no electrode coordinates. We present UniPopT, a Population Transformer that ingests scalp EEG and iEEG through one interface: a frozen single-channel encoder turns every electrode into a token, a spherical-coordinate and modality positional encoding places it, and the aggregator is pre-trained on both modalities with next-segment and replaced-token objectives. On the official evaluation splits of two widely used clinical EEG corpora UniPopT improves on existing methods trained on the same recordings; on the external Brain4FMs benchmark it obtains the best mean rank on the scalp-EEG leaderboards and the top accuracy on the intracranial ones. Decomposing the gain on the same external benchmark, we find that the full model beats every ablated variant, that the gain enters through iEEG in the aggregator's pre-training corpus, and that the spherical encoding is what lets the model exploit electrode montages it never saw during pre-training. Beyond benchmark accuracy, we analyse how decoding scales with the amount of fine-tuning data per subject and show that most of the gain from additional electrodes arrives through the spherical coordinates.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG, intracranial EEG, foundation models, self-supervised learning, electrode coordinates, positional encoding, clinical neurophysiology

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_dmeqOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

UniPopT asks whether intracranial EEG (iEEG) actually improves scalp-EEG decoding when both modalities share one model, extending the Population Transformer recipe: a frozen single-channel TS2Vec encoder embeds each electrode, a spherical positional encoding (r, θ, φ about the anterior commissure, plus depth, modality, domain, and segment slots) places scalp electrodes and patient-specific contacts in one frame, and a PopT aggregator is pretrained on a bimodal corpus (TUH + five NEMAR iEEG datasets; 1,136 h scalp + 691 h iEEG) with next-segment and replaced-token objectives. On the official TUAB/TUEV splits UniPopT exceeds data-matched retrained baselines (89.52 AUROC vs 88.88 for the PopT recipe), and on the external Brain4FMs benchmark it obtains the best mean AUROC rank on the 8 scalp boards (2.75 vs 3.50 for REVE) and top accuracy on the 2 intracranial boards. One-change-at-a-time pretraining variants (P1–P4) localize the gain: iEEG helps through the aggregator's pre-training corpus (+2.62 AUROC, 9/9 boards, p = .001), not the encoder bank; the spherical encoding is what exploits unseen montages (advantage ~4× the Cartesian variant, growing with recording length per subject); and a Laplacian-reference-geometry analysis explains a −14.8-point dense-montage loss.

Critical strengths
  • Exceptional experimental discipline and an answer to a genuinely open question. Data-matched retrained controls (P4), one-component-per-variant pretraining runs at equal compute, inclusion criteria fixed before outcomes, paired significance tests and Friedman statistics, and an explicit separation of published from retrained numbers — the decomposition of where iEEG gains enter (corpus vs encoder bank) is evidence of a kind rarely produced in this literature, and it corrects the modality-vs-volume confound in the prior BrainWave report.
  • The analyses go well beyond benchmark accuracy and are practically actionable: montage × recording-length curves (dense montages pay off only with enough per-subject data), the forensics tracing a dense-montage failure to Laplacian reference spacing, and the token-space redundancy analysis (linear CKA 0.958 between bimodal and EEG-only encoders) that explains why iEEG helps at the corpus level but not the encoder level. The limitations section is candid and specific.
Critical weaknesses
  • The headline claim — iEEG improves scalp decoding — remains partly confounded with corpus size: the bimodal corpus contains ~30% more recordings (4,043 vs 3,076) at matched compute. The appendix's scalp-heavy control (UNIPOPT-EX) points the right way but also changes the stride and encoder, so the modality contribution is not cleanly isolated.
  • Methodological novelty is moderate (PopT machinery + spherical encoding + bimodal data), in-distribution gains are small (+0.42 TUAB AUROC; flat on TUEV), external coverage is selective (10 of 26 Brain4FMs boards, 7 of 21 datasets), and some results are very noisy (UCSD-OFF 61.44 ± 25.22), which limits the strength of per-board conclusions.
Questions for authors
  • A size-matched EEG-only control (same recording count as the bimodal corpus) would separate modality from volume — is this feasible within your compute budget?
  • Does the spherical encoding's advantage survive a stronger Cartesian baseline (e.g., per-axis Fourier features of xyz, or a learned MLP of xyz) rather than the integer-lookup variant?
  • Is there evidence for the reverse direction (scalp pretraining improving iEEG decoding), or a specific reason the asymmetry is expected?
4 comments
AuthorsOpenReview ↗

Reply to Reviewer dmeq (part 1)

We sincerely thank the reviewer for the thorough evaluation and constructive suggestions. We have conducted additional experiments and analyses to further investigate these issues and clarify our findings. Due to limited computational resources resulting from upcoming conference deadlines, we apologize for the delayed response. Some additional experiments are also still ongoing or in preparation, and we appreciate your understanding. We address each concern in detail below.

1. To address the concern raised in your Critical Weaknesses and Question 1 that the benefits of adding iEEG may be confounded by corpus size and diversity, we conducted an ablation: On top of the 3,076 TUH recordings, we added 967 scalp EEG recordings from 320 distinct subject IDs, selected from three datasets included in NeuralBench (Banville et al., 2026): Shirazi2024Hbn, Hinss2023Open, and Levy2026, contributing 710, 145, and 112 recordings, respectively. These additional recordings have a total raw duration of approximately 90.54 hours of age prediction, mental workload, and typing tasks.

The resulting EEG-only corpus, denoted EEG4043, matches the bimodal corpus in the total number of recordings (4,043), subject IDs (TUH + 320), source studies (TUH + three studies), and number of training samples for the encoder (200,400). Using EEG4043, we pretrained both the single-channel encoder and the aggregator, followed by downstream fine-tuning, with the same experimental settings as UniPopT.

The external evaluation results are presented below. Evaluation on the two intracranial HUP leaderboards is still ongoing and is therefore not included here; nevertheless, the completed scalp EEG evaluations already provide evidence relevant to this concern.

BenchmarkEEG4043C6P1C6 − EEG4043
ADFD80.56±3.2380.56 \pm 3.2384.61±4.7284.61 \pm 4.7281.72±5.9181.72 \pm 5.91+4.05
SD-2879.26±9.6879.26 \pm 9.6881.35±7.3481.35 \pm 7.3478.37±8.9578.37 \pm 8.95+2.09
MDD-6494.33±4.2494.33 \pm 4.2496.90±2.4796.90 \pm 2.4795.35±2.5295.35 \pm 2.52+2.57
UCSD-ON62.15±20.6062.15 \pm 20.6077.93±6.4077.93 \pm 6.4072.90±10.1572.90 \pm 10.15+15.78
UCSD-OFF51.63±15.0251.63 \pm 15.0261.44±25.2261.44 \pm 25.2256.72±18.8756.72 \pm 18.87+9.81
Dep-BDI65.33±5.5065.33 \pm 5.5064.70±6.7264.70 \pm 6.7261.71±9.7961.71 \pm 9.79-0.63
Dep-STAI63.28±3.3163.28 \pm 3.3165.18±3.4165.18 \pm 3.4162.84±1.8562.84 \pm 1.85+1.90
Unweighted Mean (7 Benchmarks)70.9376.0272.80+5.08

Across the seven evaluated leaderboards, EEG4043 outperforms UniPopT (C6) only on Dep-BDI, while UniPopT achieves a mean AUROC advantage of 5.08 percentage points over EEG4043 (UniPopT − EEG4043). Moreover, EEG4043 performs similarly overall to P1, the variant pretrained using only TUH data. Under these matched settings, adding scalp EEG recordings from multiple studies does not reproduce the gains obtained by adding iEEG. These findings provide evidence against an explanation based solely on the approximately 31% increase in recording count and support the contribution of complementary information from iEEG. We will add this ablation and its results later into the manuscript.

We also evaluated EEG4043 on the TUH-based downstream tasks, with the results presented below.

MetricEEG4043C6P1
TUAB AUROC89.03±0.0989.03 \pm 0.0989.52±0.1089.52 \pm 0.1089.10±0.1089.10 \pm 0.10
TUEV Weighted F171.09±1.2371.09 \pm 1.2371.83±0.2671.83 \pm 0.2671.82±0.6071.82 \pm 0.60
TUEV κ×100\kappa \times 10061.49±1.7961.49 \pm 1.7962.39±0.5062.39 \pm 0.5062.12±0.8262.12 \pm 0.82

One possible explanation for these findings is that iEEG contributes primarily through joint dependencies across electrodes and time segments. Even if single-channel tokens from the two modalities occupy highly overlapping representation spaces, their relationships across electrodes and time segments may still differ. The aggregator is trained to determine whether two segments are temporally consecutive and to identify channel tokens that have been replaced with activity from other time points. Thus, iEEG may provide complementary signals for these pretraining tasks through activity relationships captured under different spatial sampling configurations, helping the aggregator learn more transferable contextual dependencies. This explanation is consistent with our existing representation analyses, but the specific mechanism requires further validation.

AuthorsOpenReview ↗

Reply to Reviewer dmeq (part 2)

2. To address the concern of your second question of “Does the spherical encoding's advantage survive a stronger Cartesian baseline?”, we followed the suggestion from another reviewer who raised a similar question. That is to “test REVE’s (x,y,z,t)(x, y, z, t) encoding within pipeline.”. We retained UniPopT’s original training data, architecture, and training configuration, replacing only the spherical positional encoding with REVE’s 512-dimensional continuous 4D positional encoding with few adaptations (details of these adaptations will be provided in the appendix). We conducted evaluations on the in-domain TUEV dataset and Dep-STAI (as the performance difference between the 19-channel and 64-channel settings on the latter can help reveal the properties of the spherical positional encoding). The results are as follows:

Position encodingDep-STAI 19-channel AUROCDep-STAI 64-channel AUROCTUEV weighted F1TUEV κ\kappa
P3: integer Cartesian59.89±4.2959.89 \pm 4.2961.42±3.8361.42 \pm 3.8370.41±1.4170.41 \pm 1.4160.36±1.6560.36 \pm 1.65
Adapted REVE PE60.60±5.00\mathbf{60.60 \pm 5.00}64.70±2.2464.70 \pm 2.2470.79±1.3870.79 \pm 1.3861.17±1.5561.17 \pm 1.55
C6: UniPopT PE57.75±7.7457.75 \pm 7.7465.18±3.41\mathbf{65.18 \pm 3.41}71.14±0.50\mathbf{71.14 \pm 0.50}61.31±0.54\mathbf{61.31 \pm 0.54}

The continuous REVE encoding improves mean 64-channel AUROC over P3 by approximately 3.28 percentage points, confirming the importance of testing a stronger Cartesian baseline. C6 retains the highest mean in the 64-channel setting and on TUEV, although its margins over REVE PE are small. REVE PE performs better in the 19-channel setting. These findings support the competitiveness of UniPopT’s encoding under a stronger, data-matched control, rather than universal or statistically established superiority of spherical encoding.


3. In response of your third question: Our study focused on whether including iEEG in pretraining improves scalp EEG decoding. Our current experiments do not isolate the reverse contribution through a matched iEEG-only control. However, there is a physiological basis for expecting the benefits to be task-dependent and potentially asymmetric.

Scalp EEG measures spatially mixed and attenuated neural activity, while iEEG provides more local measurements near the sampled generators. Simultaneous scalp–intracranial recordings show that focal events detectable intracranially can have weak or absent scalp correlates, and that their scalp visibility depends on the spatial extent and synchrony of the underlying activity. These findings demonstrate differences in the information accessible to the two recording modalities. Tao et al., 2005 · Casale et al., 2022

On this basis, we hypothesize that iEEG pretraining can expose the model to more spatially resolved examples of neural dynamics relevant to scalp decoding, whereas scalp pretraining may contribute less additional information to intracranial tasks that depend on focal activity. This provides a specific reason to expect unequal transfer benefits. Scalp recordings could still improve iEEG decoding through shared rhythms, broader spatial coverage, and greater training diversity. The expected asymmetry therefore depends on the downstream task, and its magnitude remains untested in our model.

The available evaluation resources also limit a systematic comparison. TUAB and TUEV provide established in-domain scalp evaluations with reference baselines. For the intracranial datasets used in our pretraining, we currently lack a comparable established in-domain evaluation protocol and baseline suite. Our external iEEG evaluation covers two HUP seizure-detection tasks. This uneven task coverage restricts how broadly we can assess reverse transfer or compare the two directions.

I thank the authors for their response. Most of my concerns have been addressed, and I keep my positive assessment as 3.

We are glad that our clarifications and additional experiments have addressed most of your concerns. Thank you for your positive feedback and for taking the time to review our responses

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

UNIPOPT extends the Population Transformer (PopT) architecture to jointly handle scalp EEG and intracranial EEG (iEEG) through a shared spherical-coordinate positional encoding and aggregator. The pipeline uses a frozen TS2Vec single-channel encoder, projects each electrode into a unified spherical coordinate frame, and pre-trains a transformer aggregator on a bimodal corpus (1,136h scalp EEG and 691h iEEG) with next-segment and replaced-token objectives. The paper evaluates on TUAB/TUEV official splits and ten Brain4FMs leaderboards, and runs a five-variant ablation decomposing gains into encoder bank, aggregator corpus, and positional encoding contributions. Key findings: iEEG helps through the aggregator’s pre-training corpus (not the encoder bank) and the spherical encoding enables generalization to unseen electrode montages.

Critical strengths
  1. The five-variant ablation design is well-constructed and provides insight into how iEEG contributes to a brain-signal foundation model. Varying encoder bank, aggregator corpus, and positional encoding independently (Table 3) under fixed compute is a clean experimental design. The token-space analysis (CKA 0.958, 92.1% subspace overlap) gives a concrete explanation for limited effect of iEEG at the encoder bank level.

  2. UNIPOPT is the first model to place both scalp and intracranial electrodes in a shared coordinate frame, enabling coordinate-aware processing of both modalities from one checkpoint. BrainWave already handles both modalities without coordinates and achieves stronger iEEG results in terms of AUROC. The novelty is therefore specifically the coordinate interface.

Critical weaknesses
  1. The central claim that iEEG improves scalp-EEG decoding is confounded with corpus diversity. UNIPOPT’s aggregator corpus has 4,043 recordings versus P1’s 3,076, i.e., 31% more distinct recordings. Token-space analysis (CKA 0.958) shows that iEEG is embedded into the same subspace as scalp EEG. This makes it more plausible that the aggregator’s gain reflects seeing more diverse recordings rather than a different modality contributing complementary information.

  2. The scalp EEG pre-training corpus is limited to one source (TUH), whereas competing models pre-train on multiple diverse sources (REVE 92 datasets, BrainWave 13 scalp EEG datasets). Combined with the absence of a data-matched REVE baseline, a state-of-the-art foundation model with continuous positional encoding and a close competitor on Brain4FM, this makes it difficult to tell whether UNIPOPT’s results reflect the architecture or the corpus characteristics.

  3. The in-domain evaluation is constrained to the TUEV and TUAB datasets, and the Brain4FM leaderboards to clinical/psychiatrics classification tasks. Thus, the evaluation of UNIPOPT misses standard EEG FM tasks such as sleep staging, and motor imagery BCI.

Questions for authors
  1. The CKA of 0.958 between bimodal and EEG-only encoders shows that iEEG tokens are embedded into the same subspace as scalp EEG after preprocessing. If the encoder cannot distinguish the modalities, the aggregator’s gain from iEEG in its corpus (+2.62 external AUROC) could simply reflect seeing 31% more diverse recordings (4,043 vs. 3,076) rather than a different modality contributing complementary information. Could you run an ablation with an EEG-only corpus of approximately 4,043 recordings matching diversity?

  2. REVE, with continuous coordinate-aware positional encoding, is your closest competitor on Brain4FMs and a state-of-the-art foundation model. Could REVE be added as a baseline, either by testing REVE’s (x, y, z, t) encoding within your pipeline or retraining REVE on the TUH corpus?

  3. The ‘best mean rank’ metric does not represent the variance across the leaderboards. UNIPOPT is 13 AUROC points behind REVE on EEGMat and nearly 7 behind BrainWave on SD-28, while leading by 9 on UCSD-ON. Using mean rank, a 0.02-point positive gap (MDD-64) is weighted the same as a 13-point negative gap. Could you provide a weighted rank or mean AUROC difference?

  4. The 200 Hz sampling rate removes the high-gamma band (90–150 Hz), arguably iEEG’s most distinctive feature. While this is framed as a field convention, e.g., BrainWave has used an embedding layer that aligns the inherently different sampling frequencies of scalp and iEEG. Could you train an encoder at higher sampling frequency to account for the high-gamma band?

5 comments
AuthorsOpenReview ↗

Reply to Reviewer aLja (part 1)

We sincerely thank the reviewer for the thorough evaluation and constructive suggestions. We have conducted additional experiments and analyses to further investigate these issues and clarify our findings. Due to limited computational resources resulting from upcoming conference deadlines, we apologize for the delayed response. Some additional experiments are also still ongoing or in preparation, and we appreciate your understanding. We address each concern in detail below.

1. To address the concern raised in your Critical Weaknesses and Question 1, we conducted an ablation: On top of the 3,076 TUH recordings, we added 967 scalp EEG recordings from 320 distinct subject IDs, selected from three datasets included in NeuralBench (Banville et al., 2026): Shirazi2024Hbn, Hinss2023Open, and Levy2026, contributing 710, 145, and 112 recordings, respectively. These additional recordings have a total raw duration of approximately 90.54 hours of age prediction, mental workload, and typing tasks.

The resulting EEG-only corpus, denoted EEG4043, matches the bimodal corpus in the total number of recordings (4,043), subject IDs (TUH + 320), source studies (TUH + three studies), and number of training samples for the encoder (200,400). Using EEG4043, we pretrained both the single-channel encoder and the aggregator, followed by downstream fine-tuning, with the same experimental settings as UniPopT.

The external evaluation results are presented below. Evaluation on the two intracranial HUP leaderboards is still ongoing and is therefore not included here; nevertheless, the completed scalp EEG evaluations already provide evidence relevant to this concern.

BenchmarkEEG4043C6P1C6 − EEG4043
ADFD80.56±3.2380.56 \pm 3.2384.61±4.7284.61 \pm 4.7281.72±5.9181.72 \pm 5.91+4.05
SD-2879.26±9.6879.26 \pm 9.6881.35±7.3481.35 \pm 7.3478.37±8.9578.37 \pm 8.95+2.09
MDD-6494.33±4.2494.33 \pm 4.2496.90±2.4796.90 \pm 2.4795.35±2.5295.35 \pm 2.52+2.57
UCSD-ON62.15±20.6062.15 \pm 20.6077.93±6.4077.93 \pm 6.4072.90±10.1572.90 \pm 10.15+15.78
UCSD-OFF51.63±15.0251.63 \pm 15.0261.44±25.2261.44 \pm 25.2256.72±18.8756.72 \pm 18.87+9.81
Dep-BDI65.33±5.5065.33 \pm 5.5064.70±6.7264.70 \pm 6.7261.71±9.7961.71 \pm 9.79-0.63
Dep-STAI63.28±3.3163.28 \pm 3.3165.18±3.4165.18 \pm 3.4162.84±1.8562.84 \pm 1.85+1.90
Unweighted Mean (7 Benchmarks)70.9376.0272.80+5.08

Across the seven evaluated leaderboards, EEG4043 outperforms UniPopT (C6) only on Dep-BDI, while UniPopT achieves a mean AUROC advantage of 5.08 percentage points over EEG4043 (UniPopT − EEG4043). Moreover, EEG4043 performs similarly overall to P1, the variant pretrained using only TUH data. Under these matched settings, adding scalp EEG recordings from multiple studies does not reproduce the gains obtained by adding iEEG. These findings provide evidence against an explanation based solely on the approximately 31% increase in recording count and support the contribution of complementary information from iEEG. We will add this ablation and its results later into the manuscript.

We also evaluated EEG4043 on the TUH-based downstream tasks, with the results presented below.

MetricEEG4043C6P1
TUAB AUROC89.03±0.0989.03 \pm 0.0989.52±0.1089.52 \pm 0.1089.10±0.1089.10 \pm 0.10
TUEV Weighted F171.09±1.2371.09 \pm 1.2371.83±0.2671.83 \pm 0.2671.82±0.6071.82 \pm 0.60
TUEV κ×100\kappa \times 10061.49±1.7961.49 \pm 1.7962.39±0.5062.39 \pm 0.5062.12±0.8262.12 \pm 0.82
One possible explanation is that iEEG contributes primarily through joint dependencies across electrodes and time segments. Even if single-channel tokens from the two modalities occupy highly overlapping representation spaces, their relationships across electrodes and time segments may still differ. The aggregator is trained to determine whether two segments are temporally consecutive and to identify channel tokens that have been replaced with activity from other time points. Thus, iEEG may provide complementary signals for these pretraining tasks through activity relationships captured under different spatial sampling configurations, helping the aggregator learn more transferable contextual dependencies. This explanation is consistent with our representation analyses, but the specific mechanism requires further validation.
AuthorsOpenReview ↗

Reply to Reviewer aLja (part 2)

2. To address the concern of your second question and in the Weaknesses that “it difficult to tell whether UNIPOPT’s results reflect the architecture or the corpus characteristics,” we followed your suggestion of “testing REVE’s (x,y,z,t)(x, y, z, t) encoding within pipeline.” We retained UniPopT’s original training data, architecture, and training configuration, replacing only the spherical positional encoding with REVE’s 512-dimensional continuous 4D positional encoding with few adaptations (details of these adaptations will be provided in the appendix). We conducted evaluations on the in-domain TUEV dataset and Dep-STAI (as the performance difference between the 19-channel and 64-channel settings on the latter can help reveal the properties of the spherical positional encoding). The results are as follows:

Position encodingDep-STAI 19-channel AUROCDep-STAI 64-channel AUROCTUEV weighted F1TUEV κ\kappa
P3: integer Cartesian59.89±4.2959.89 \pm 4.2961.42±3.8361.42 \pm 3.8370.41±1.4170.41 \pm 1.4160.36±1.6560.36 \pm 1.65
Adapted REVE PE60.60±5.00\mathbf{60.60 \pm 5.00}64.70±2.2464.70 \pm 2.2470.79±1.3870.79 \pm 1.3861.17±1.5561.17 \pm 1.55
C6: UniPopT PE57.75±7.7457.75 \pm 7.7465.18±3.41\mathbf{65.18 \pm 3.41}71.14±0.50\mathbf{71.14 \pm 0.50}61.31±0.54\mathbf{61.31 \pm 0.54}

The continuous REVE encoding improves mean 64-channel AUROC over P3 by approximately 3.28 percentage points, confirming the importance of testing a stronger Cartesian baseline. C6 retains the highest mean in the 64-channel setting and on TUEV, although its margins over REVE PE are small. REVE PE performs better in the 19-channel setting. These findings support the competitiveness of UniPopT’s encoding under a stronger, data-matched control, rather than universal or statistically established superiority of spherical encoding.


3. About the third question, we agree that our emphasis on the best mean rank overstated the consistency of UniPopT’s advantage across the scalp EEG tasks. We used the mean AUROC method as you mentioned: On all leaderboards with equal weights, we compared UniPopT with REVE and BrainWave.

The mean AUROC difference between UniPopT and BrainWave (UniPopT − BrainWave) was +1.53 percentage points across the scalp EEG leaderboards and −1.03 percentage points across all ten when the two iEEG leaderboards were included. The difference between UniPopT and REVE was +1.27 percentage points across the eight scalp EEG leaderboards. iEEG leaderboards were not evaluated on REVE.

The reason of using equal weights is that, we believe the appropriate weight for a board depends on many complicated factors. How best to weight heterogeneous tasks for a particular application is a broader question of benchmark design that is not in this scope of research, so we want to keep it simple here.

Meanwhile, we found out that UCSD-ON and UCSD-OFF use recordings from the same Parkinson’s disease cohort (OpenNeuro ds002778), which BrainWave included in its self-supervised pretraining corpus as PD-31 (Yuan et al., 2024). The two leaderboards compare the same patients, recorded on or off medication, with the same healthy controls. This establishes dataset-level overlap, and may affect the overall evaluation here.

In conclusion, if you suggest so, we will replace the headline claim about mean rank with the per-board differences and their means, and add the following table into the manuscript.

Leaderboard /UniPopT − REVEUniPopT − BrainWave
ADFD0.3310.5
SD-280.34−6.81
MDD-640.023.38
UCSD-ON15.129.22
UCSD-OFF−1.053.46
Dep-BDI−2.37−7.70
Dep-STAI10.851.63
EEGMat−13.06−1.44
HUP-ECoG—−9.55
HUP-SEEG—−13.03
Mean1.271.53/-1.03
               *AUROC differences*
AuthorsOpenReview ↗

Reply to Reviewer aLja (part 3)

4. For the fourth question: To investigate whether high gamma band provides additional predictive information, we conducted a complementary HUP sensitivity analysis. We froze the pretrained TS2Vec encoder and aggregator, extracted 90–150 Hz log-power summaries from the original EDF recordings, and compared three logistic-regression probes: frozen representations alone, representations plus high-gamma features, and representations plus within-patient shuffled high-gamma features. Bandwidth screening retained 20/20 ECoG and 34/37 SEEG patients.

The observed AUROC differences, in percentage points, were:

DatasetHG vs. baseline: pooled-window meanHG vs. shuffled: pooled-window meanHG vs. baseline: patient-averaged
HUP-ECoG+2.184+2.179−0.233
HUP-SEEG−0.296−0.265−0.124

The patient-averaged metric first averages repeated test appearances within each patient, then averages across patients.

These results show a pooled-window improvement on ECoG, but no positive patient-averaged gain on either dataset. They support a limited conclusion: adding these high-gamma power summaries does not yield uniformly beneficial results in the frozen-representation setting. This analysis suggests, to some extent, that directly training an encoder at a higher sampling rate may not yield substantial improvements. We believe that an embedding layer similar to that used in BrainWave to align the different native sampling rates of scalp EEG and iEEG may be necessary. However, this would require changes to the overall model architecture and may not be feasible within the available rebuttal time.

Reviewer_7kNWOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents UniPopT, a foundational model jointly pretrained on scalp EEG and iEEG. It uses a frozen single-channel TS2Vec encoder to embed each channel as a token, then uses an aggregator to produce a contextual token for the brain stage. The model demonstrates competitive performance on the scalp EEG tasks with fine-tuning. Through analysis, the authors claimed that iEEG improves performance when added to the aggregator's pretrain cohort, but not when added to the encoder's pretraining data. Later, the authors argue that the spherical encoding allows the model to exploit electrode montages unseen during pretraining.

Critical strengths
  1. The UniPopT model demonstrates competitive performance on the scalp EEG tasks with fine-tuning.
  2. The paper provides analyses that adding iEEG to the aggregator's pretraining improves AUROC on all nine external boards, but adding the iEEG to the encoder's training data does not improve AUROC.
Critical weaknesses
  1. The current evaluation does not demonstrate that the model is useful for iEEG beyond seizure detection: The paper sells UniPopT as a brain signal foundational model that serves both scalp and iEEG signals (L85). However, its iEEG evaluation only covers two seizure detection tasks from HUP (L854) and UniPopT is behind BrainWave on both tasks. Other clinically relevant iEEG tasks, such as Omni-iEEG, are not evaluated; moreover, UniPopT also did not evaluate the Neuroprobe, where the original PopT is evaluated. Furthermore, resampling iEEG to 200 Hz removes most of the frequency range that defines the high-frequency components in clinically relevant tasks.

  2. The results show that adding iEEG recordings helps, but do not demonstrate that the improvement actually came from the intracranial information. The paper claimed that iEEG improves scalp recordings because its high signal-to-noise recordings directly from the cortex (L31) The authors’ own analysis shows that training the encoder with iEEG does not change how it represents scalp EEG. The gain from adding iEEG may come from increasing the training data from 3076 to 4043 recordings (i.e. more diverse cohort) and no size-matched EEG-only control was run. Thus, the main claim of the paper "A bimodal pre-training recipe in which iEEG improves scalp-EEG decoding" may just because of the diversity of the training data.

Questions for authors
  1. Will it help if we add more diverse iEEG data? Most of the iEEG data comes from memory tasks. Will more hetergenious iEEG dataset help, such as sleep and ictal data?
  2. What is the performance of the linear probe? Since UniPopT is a foundational model, it should have sufficient zero-shot capability. However, all reported metrics are based on task-specific fine-tuning.
  3. Does the joint pretraining also help iEEG decoding and classification? The paper only measures whether iEEG helps scalp EEG, but does not measure the reverse direction.
8 comments
AuthorsOpenReview ↗

Reply to Reviewer 7kNW (part 1)

We sincerely thank the reviewer for the thorough evaluation and constructive suggestions. We have conducted additional experiments and analyses to further investigate these issues and clarify our findings. Due to limited computational resources resulting from upcoming conference deadlines, we apologize for the delayed response. Some additional experiments are also still ongoing or in preparation, and we appreciate your understanding. We address each concern in detail below.

1. Regarding your concern in the Weaknesses section about the limited evaluation of iEEG performance: Our original intention was to align with the brain4fm benchmark. However, after excluding some brain4fm datasets that could not be downloaded or were incompatible, the range of tasks was indeed limited. We acknowledge this and, following your suggestion, evaluated two iEEG leaderboards: 1) the Detroit/UCLA subset of Omni-iEEG, for pathological channel identification; 2) Neuroprobe Lite, previously evaluated in the Population Transformer study, for sentence onset detection. The results are as follows:

Model and submissionFold 0 AUROCFold 1 AUROCMean
PopT, submitted by the original author Geeling Chau, Laplacian STFT + global standardization0.98440.91370.9491
Frozen C6 + logistic regression (this evaluation)0.88650.81590.8512
ModelChannel-level AUROCMacro-F1
TimeConv-CNN0.80610.6469
SEEG-NET0.78500.5259
CLAP0.76840.6009
EEGNet, single-channel adaptation0.74680.5640
PyHFO-Omni0.73510.5635
Handcrafted features + linear classifier0.50450.4899
Frozen C6, Detroit/UCLA subset (this evaluation)0.78930.6061

On Neuroprobe, UniPopT is 9.79 AUROC percentage points below PopT, with lower performance in both folds. However, Neuroprobe is an in-domain training dataset for PopT, placing UniPopT at an inherent disadvantage in this comparison. On Omni-iEEG, performance is close to that of SEEG-NET and 1.68 AUROC percentage points below that of TimeConv-CNN. These results extend the iEEG evaluation beyond seizure detection and show that the model also performs well on other iEEG evaluation leaderboards.

AuthorsOpenReview ↗

Reply to Reviewer 7kNW (part 2)

2. Regarding the concern about 200 Hz raised in your Weaknesses section, “Resampling iEEG to 200 Hz removes much of the high-frequency range relevant to clinical tasks”: To investigate whether high gamma band provides additional predictive information, we conducted a complementary HUP sensitivity analysis. We froze the pretrained TS2Vec encoder and aggregator, extracted 90–150 Hz log-power summaries from the original EDF recordings, and compared three logistic-regression probes: frozen representations alone, representations plus high-gamma features, and representations plus within-patient shuffled high-gamma features. Bandwidth screening retained 20/20 ECoG and 34/37 SEEG patients.

The observed AUROC differences, in percentage points, were:

DatasetHG vs. baseline: pooled-window meanHG vs. shuffled: pooled-window meanHG vs. baseline: patient-averaged
HUP-ECoG+2.184+2.179−0.233
HUP-SEEG−0.296−0.265−0.124

The patient-averaged metric first averages repeated test appearances within each patient, then averages across patients.

These results show a pooled-window improvement on ECoG, but no positive patient-averaged gain on either dataset. They support a limited conclusion: adding these high-gamma power summaries does not yield uniformly beneficial results in the frozen-representation setting. This analysis suggests, to some extent, that directly training an encoder at a higher sampling rate may not yield substantial improvements. We believe that an embedding layer similar to that used in BrainWave to align the different native sampling rates of scalp EEG and iEEG may be necessary. However, this would require changes to the overall model architecture and may not be feasible within the available rebuttal time.

AuthorsOpenReview ↗

Reply to Reviewer 7kNW (part 3)

3. Regarding your observation in the Weaknesses section that “no size-matched EEG-only control was run”: We conducted an ablation: On top of the 3,076 TUH recordings, we added 967 scalp EEG recordings from 320 distinct subject IDs, selected from three datasets included in NeuralBench (Banville et al., 2026): Shirazi2024Hbn, Hinss2023Open, and Levy2026, contributing 710, 145, and 112 recordings, respectively. These additional recordings have a total raw duration of approximately 90.54 hours of age prediction, mental workload, and typing tasks.

The resulting EEG-only corpus, denoted EEG4043, matches the bimodal corpus in the total number of recordings (4,043), subject IDs (TUH + 320), source studies (TUH + three studies), and number of training samples for the encoder (200,400). Using EEG4043, we pretrained both the single-channel encoder and the aggregator, followed by downstream fine-tuning, with the same experimental settings as UniPopT.

The external evaluation results are presented below. Evaluation on the two intracranial HUP leaderboards is still ongoing and is therefore not included here; nevertheless, the completed scalp EEG evaluations already provide evidence relevant to this concern.

BenchmarkEEG4043C6P1C6 − EEG4043
ADFD80.56±3.2380.56 \pm 3.2384.61±4.7284.61 \pm 4.7281.72±5.9181.72 \pm 5.91+4.05
SD-2879.26±9.6879.26 \pm 9.6881.35±7.3481.35 \pm 7.3478.37±8.9578.37 \pm 8.95+2.09
MDD-6494.33±4.2494.33 \pm 4.2496.90±2.4796.90 \pm 2.4795.35±2.5295.35 \pm 2.52+2.57
UCSD-ON62.15±20.6062.15 \pm 20.6077.93±6.4077.93 \pm 6.4072.90±10.1572.90 \pm 10.15+15.78
UCSD-OFF51.63±15.0251.63 \pm 15.0261.44±25.2261.44 \pm 25.2256.72±18.8756.72 \pm 18.87+9.81
Dep-BDI65.33±5.5065.33 \pm 5.5064.70±6.7264.70 \pm 6.7261.71±9.7961.71 \pm 9.79-0.63
Dep-STAI63.28±3.3163.28 \pm 3.3165.18±3.4165.18 \pm 3.4162.84±1.8562.84 \pm 1.85+1.90
Unweighted Mean (7 Benchmarks)70.9376.0272.80+5.08

Across the seven evaluated leaderboards, EEG4043 outperforms UniPopT (C6) only on Dep-BDI, while UniPopT achieves a mean AUROC advantage of 5.08 percentage points over EEG4043 (UniPopT − EEG4043). Moreover, EEG4043 performs similarly overall to P1, the variant pretrained using only TUH data. Under these matched settings, adding scalp EEG recordings from multiple studies does not reproduce the gains obtained by adding iEEG. These findings provide evidence against an explanation based solely on the approximately 31% increase in recording count and support the contribution of complementary information from iEEG. We will add this ablation and its results later into the manuscript.

We also evaluated EEG4043 on the TUH-based downstream tasks, with the results presented below.

MetricEEG4043C6P1
TUAB AUROC89.03±0.0989.03 \pm 0.0989.52±0.1089.52 \pm 0.1089.10±0.1089.10 \pm 0.10
TUEV Weighted F171.09±1.2371.09 \pm 1.2371.83±0.2671.83 \pm 0.2671.82±0.6071.82 \pm 0.60
TUEV κ×100\kappa \times 10061.49±1.7961.49 \pm 1.7962.39±0.5062.39 \pm 0.5062.12±0.8262.12 \pm 0.82

One possible explanation for these findings is that iEEG contributes primarily through joint dependencies across electrodes and time segments. Even if single-channel tokens from the two modalities occupy highly overlapping representation spaces, their relationships across electrodes and time segments may still differ. The aggregator is trained to determine whether two segments are temporally consecutive and to identify channel tokens that have been replaced with activity from other time points. Thus, iEEG may provide complementary signals for these pretraining tasks through activity relationships captured under different spatial sampling configurations, helping the aggregator learn more transferable contextual dependencies. This explanation is consistent with our existing representation analyses, but the specific mechanism requires further validation.


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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.