acceptodds
Under review as a conference paper at ICLR 2027

PHASE: A Physiology-Guided Hierarchical Foundation Model for Intracranial EEG

11 views

Abstract

Clinicians and neuroscientists have long analyzed intracranial electroencephalography (iEEG) through directly measurable physiological characteristics, which carry much of the information that downstream tasks depend on. Recent iEEG foundation models learn by reconstructing or predicting their inputs, which leaves the retention of these characteristics implicit. They are also evaluated mainly on cognitive decoding and a narrow clinical task, i.e., seizure detection. On a broad, clinically relevant benchmark such as Omni-iEEG, they remain below task-specific models when used frozen. We introduce PHASE, a physiology-guided foundation model that makes these characteristics explicit learning targets, pairing them with masked latent prediction in a temporal stage (PHASE-T) within each channel and a spatiotemporal stage (PHASE-ST) across synchronized channels. PHASE is pretrained on heterogeneous recordings from 222 participants at nine clinical sites. On all five Omni-iEEG clinical tasks, frozen PHASE-T outperforms every evaluated foundation model by up to 31%, and fine-tuned PHASE-T surpasses the task-specific models, setting a new state of the art. PHASE-T benefits from physiological supervision, outperforming variants trained with latent prediction alone or auxiliary waveform reconstruction on every task in matched ablations. PHASE-T generalizes to unseen institutions, outperforming the compared models with few or no local labels. PHASE-ST further improves seizure-onset-zone identification over PHASE-T and, when frozen, decodes sound volume and pitch on BrainTreebank better than published models. Beyond task performance, PHASE learns to encapsulate the physiological characteristics clinicians recognize, from seizure onset and its propagation to anatomical region identity, even though its pretraining contains no ictal recordings or anatomical labels.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Computational neuroscience, iEEG, Computer Aided Diagnosis, Neurophysiology

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_r2L9OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PHASE: A Physiology-Guided Hierarchical Foundation Model for Intracranial EEG proposes a two stage approach for iEEG analysis. The first stage is dedicated to multi-scale temporal analysis. The second stage focuses on spatial features and is channel-agnostic. Both backbones are JEPA-based, but this decision is not ablated. Overall, the evaluation is thorough and the results are strong.

Critical strengths
  1. The paper is well written and clearly structured
  2. The model is well described in all its components, which are meaningfully justified as well
  3. The model is lightweight and reaches SoTA performance already
Critical weaknesses
  1. The JEPA objective is one of the more catchy contributions, but it’s not ablated; indeed, Table 2 seems to suggest that it’s not particularly effective
  2. The training split of the datasets used for downstream tasks comprise the vast majority of the pretraining data, so this could partly explain the improved performance on them. For example, NeuroCLUS ablates this choice, though in their case one dataset, SWEC, made up for 95% of their training data, so this was even much less important than for this work. Nonetheless, they found a statistically significant improvement at TreeBank Speech.
Questions for authors
  1. An ablation of the training dataset over the results of the downstream tasks is necessary to distinguish the effects of the architecture choices, following or at least borrowing from NeuroCLUS would be a good idea
  2. The model scaling is not explored; while this is not quite necessary due to the already SoTA results, the model lends itself well to being scaled upwards both in parameters and in training dataset. A positive scaling result would strengthen the contribution significantly
5 comments
AuthorsOpenReview ↗

Response to Reviewer r2L9 1/2

We thank Reviewer r2L9 for the constructive review and for finding the paper clearly written, the model components well justified, the evaluation thorough, and the results strong for a lightweight model. We added three new analyses: an ablation that removes the JEPA objective (Physio-only), a PHASE-T model pretrained on BrainTreebank alone (no Omni-iEEG training participants), and a model-size scaling study. The detailed responses follow.

[W1] The JEPA objective is one of the more catchy contributions, but it's not ablated; indeed, Table 2 seems to suggest that it's not particularly effective

Thank you for pointing this out. We added an ablation that removes the JEPA objective. Physio-only drops the latent-prediction term from Equation 2 and keeps the physiological loss; the corpus, 60K updates, optimization settings, and frozen-probe protocol are the same as in Table 2.

Pretraining variantHFO F1Ictal F1Sleep F1Lobe-5 F1Region-12 F1Channel AUCOutcome AUC
Full PHASE-T (latent + physiological)0.81380.89580.74130.48220.33980.83940.6369
Physio-only (new)0.80450.89240.65340.46100.32270.81650.6215
Latent-only (Table 2)0.62760.85970.53980.32710.19960.75700.6285

Removing the JEPA objective lowers all seven scores. Since Full PHASE-T also exceeds Latent-only on every score, the JEPA objective and the physiological targets both contribute to the effectiveness of the PHASE.

[Q1, W2] The training splits of the downstream datasets make up most of the pretraining data, which could partly explain the downstream gains; NeuroCLUS, whose training data were 95% SWEC, still found a significant improvement on TreeBank Speech. An ablation of the pretraining data, following or borrowing from NeuroCLUS, is needed to separate data effects from architecture choices.

Thank you for this helpful suggestion. We retrained PHASE-T for 60K updates on the BrainTreebank pretraining split alone: 10 participants from a single site, recorded during naturalistic listening. This removes the other eight sites from pretraining, including Detroit and UCLA, the two largest sites by recording hours, and all clinical-monitoring recordings. The resulting model remains comparable to or better than the released foundation models in Table 1 on most scores, so PHASE-T does not rely on the largest sites or on clinical-monitoring data; adding them improves it further, which is also consistent with the in-domain gain NeuroCLUS reports.

ModelHFO F1Ictal F1Sleep F1Lobe-5 F1Region-12 F1Channel AUCOutcome AUC
PHASE-T, full corpus0.81380.89580.74130.48220.33980.83940.6369
PHASE-T, BrainTreebank only0.79770.85050.68570.46980.30990.80720.6429
Best released foundation model (Table 1)0.79370.83230.72130.40540.25920.80550.6017
AuthorsOpenReview ↗

Response to Reviewer r2L9 2/2

[Q2] The model scaling is not explored; while this is not quite necessary due to the already SoTA results, the model lends itself well to being scaled upwards both in parameters and in training dataset. A positive scaling result would strengthen the contribution significantly

Thank you for the encouraging suggestion. We added a model-size scaling analysis. We pretrained PHASE-T encoders at two additional sizes, 18.7M and 40.3M parameters (coarse/fine widths 384/192 and 512/256, versus 768/384 in the 106.7M model of the paper), scaling width and depth together. The pretraining corpus, 60K updates, and optimization settings are the same as for the 106.7M model, with no per-size retuning, and all three encoders are evaluated with the same frozen probes.

EncoderHFO F1Ictal F1Sleep F1Lobe-5 F1Region-12 F1Channel AUCOutcome AUC
18.7M0.79570.84900.72130.44570.29750.78750.6473
40.3M0.80480.91620.73110.47000.32010.81020.6364
106.7M (Table 1)0.81380.89580.74130.48220.33980.83940.6369

Performance increases with model size from 18.7M to 106.7M on most scores. Since the data and number of updates are fixed, these gains reflect model capacity. Even the 18.7M encoder is comparable to or better than the released foundation models in Table 1 on most scores. We scaled model size only; scaling the pretraining corpus together with the update budget is left to future work.

The Authors have addressed my concerns regarding the training dataset, but the results are being presented more positively than what is actually shown. Specifically, removing the in-domain data narrows the gap with the previous models substantially, and on one task (sleep) the results are actually below the baselines. This points to substantial contributions from the data mixture, with some improvements coming from the architecture itself (NeuroCLUS, in contrast, saw minor improvements). These should be the new headline results for a fair comparison, but since the manuscript has not been updated I cannot check.

Regarding the JEPA objective, once again the new ablation weakly corroborates the contribution, but the results mostly show that the physiological targets are sufficient, with JEPA only significantly improving on sleep. Moreover, the proper ablation would also replace JEPA with a generic self-supervised term to keep EMA, predictor, and loss weights.

Interestingly, sleep is the only task on which both ablations are significant, suggesting data-dependent behavior that is not explored by the Authors (perhaps a training run of physio-only with BrainTreeBank and no sleep?).

Reviewer_qyZ2OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

PHASE is a two-stage foundation model for intracranial EEG. PHASE-T learns single-channel temporal representations using a fine–coarse–fine architecture and combines masked latent prediction with physiological targets computed from the signal, including wavelet power, amplitude descriptors, and local high-frequency power. PHASE-ST subsequently composes frozen temporal representations across synchronized channels, using additional supervision from measured pairwise relationships. Evaluation covers clinical prediction on Omni-iEEG, cross-institution transfer, auditory decoding on BrainTreebank, and retrospective analyses of seizure dynamics and anatomical correspondence.

Critical strengths
  1. Broad evaluation demonstrates useful transferable representations. The same frozen PHASE-T encoder performs consistently well across clinical tasks spanning transient events, brain states, anatomical localization, and pathological-channel identification. The cross-institution experiments further strengthen the practical relevance.

  2. The ablation studies help clarify the contributions of the temporal and cross-channel learning components. For PHASE-T, the authors evaluate the benefits of introducing handcrafted physiological targets into JEPA-style self-supervised learning, comparing against latent-only prediction and auxiliary waveform reconstruction, and examine the role of fine-scale feature fusion. For PHASE-ST, ablations of cross-channel attention and relation supervision assess the contributions of multichannel context and explicit relationship targets. Together, these experiments provide useful evidence supporting the main design choices across both training stages.

Critical weaknesses
  • Different pretraining data complicate attribution of the gains. PHASE and the baseline checkpoints use different pretraining corpora, making it difficult to distinguish data-related advantages from the benefits of PHASE’s two-stage training, JEPA-style prediction, and explicit physiological supervision.
  • The motivation for two-stage pretraining needs further discussion. The experiments support incorporating explicit physiological supervision into JEPA-style learning, but do not clearly establish why temporal and cross-channel learning should be separated into two stages rather than jointly pretrained with the same objectives.
Questions for authors
  1. Pretraining-data differences: How much of PHASE’s advantage could be explained by differences in pretraining data coverage and domain alignment? Do the authors have a comparison with a representative baseline pretrained on the same corpus under a comparable training budget?
  2. Two-stage training: What motivates freezing the temporal encoder before cross-channel pretraining rather than jointly training both components with the same objectives? Is there empirical evidence supporting this choice, or a practical constraint that makes joint pretraining unsuitable?
4 comments
AuthorsOpenReview ↗

Response to Reviewer qyZ2 1/2

We thank Reviewer qyZ2 for the constructive review and for recognizing the breadth of the evaluation, the cross-institution transfer, and the ablations of both stages. We added two analyses: a Physio-only ablation that removes the JEPA objective, and a PHASE-T model pretrained on BrainTreebank alone, which shares its pretraining participants with BrainBERT. We also explain why the temporal encoder is frozen. The detailed responses follow.

[Q1, W1] Pretraining-data differences: Because PHASE and the baselines are pretrained on different corpora, how much of PHASE's advantage could be explained by data coverage and domain alignment rather than two-stage training, JEPA-style prediction, and explicit physiological supervision? Is there a comparison with a representative baseline pretrained on the same corpus under a comparable training budget?

Thank you for this helpful question. With the corpus fixed, Latent-only (JEPA-style) and Latent+Recon in Table 2 are such baselines: they share PHASE-T's architecture, corpus, and 60K updates and differ only in the objective. Full PHASE-T exceeds both on all seven scores. We added the reverse ablation, Physio-only, which drops the latent-prediction term and keeps everything else the same.

Pretraining variantHFO F1Ictal F1Sleep F1Lobe-5 F1Region-12 F1Channel AUCOutcome AUC
Full PHASE-T0.81380.89580.74130.48220.33980.83940.6369
Physio-only (new)0.80450.89240.65340.46100.32270.81650.6215
Latent-only (Table 2)0.62760.85970.53980.32710.19960.75700.6285

Removing the JEPA objective lowers performance, demonstrating that both JEPA objective and physiological supervision contribute to the effectiveness of PHASE. These scores come from PHASE-T alone, so two-stage training does not enter them (see Q2).

We retrained PHASE-T for 60K updates on the BrainTreebank pretraining split alone, drawn from the same participants as BrainBERT's pretraining data, with a comparable amount of pretraining input (about 6.9 × 10^8 vs. 6.4 × 10^8 channel-seconds). With the pretraining data matched, PHASE-T is comparable to or better than BrainBERT, so data coverage explains part of PHASE-T’s advantage but not all of it.

ModelHFO F1Ictal F1Sleep F1Lobe-5 F1Region-12 F1Channel AUCOutcome AUC
PHASE-T, full corpus0.81380.89580.74130.48220.33980.83940.6369
PHASE-T, BrainTreebank only0.79770.85050.68570.46980.30990.80720.6429
BrainBERT (Table 1)0.79370.83230.69410.40540.25920.80550.6017
AuthorsOpenReview ↗

Response to Reviewer qyZ2 2/2

[Q2, W2] Two-stage training: The experiments support adding physiological supervision to JEPA-style learning but do not show why temporal and cross-channel learning should be separated. What motivates freezing the temporal encoder before cross-channel pretraining rather than jointly training both components with the same objectives? Is there empirical evidence for this choice, or a practical constraint that makes joint pretraining unsuitable?

Thank you for this helpful question. We did not pretrain the two stages jointly, and Section 6 lists joint pretraining as future work. The separation follows from a design goal, and it also keeps the spatial stage inexpensive.

First, many clinical tasks, such as HFO classification and anatomical localization, are labeled per channel, whereas tasks such as auditory decoding draw on many synchronized channels. Freezing PHASE-T keeps its single-channel representation from being reshaped by the cross-channel objective, so the encoder evaluated in Table 1 is exactly the one inside PHASE-ST. Because the spatial module's output weights start at zero (Appendix C.1), PHASE-ST begins from this representation and adds synchronized context to it.

Second, each spatial update covers 128 sets of 64 synchronized crops, or 8,192 crops, versus 192 per PHASE-T update. With PHASE-T frozen, these crops only pass forward through it, and the spatial stage takes about 1.9 GPU-days. Joint training would backpropagate through the 106.7M-parameter encoder for all of them: over 5,000 updates, 3.6 times as many crops as the entire PHASE-T pretraining.

The staged model works at both levels. Frozen PHASE-T outperforms all three foundation models on Omni-iEEG scores (Table 1), and PHASE-ST's context raises per-participant SOZ ROC-AUC by 0.048 (95% CI: 0.026–0.071) and improves auditory decoding for all seven BrainTreebank participants (Figure 3).

Dear Reviewer qyZ2 Thank you again for your constructive review. With less than a day remaining in the discussion period, we would be grateful to hear whether our responses have addressed your concerns. We would be happy to clarify any points that are still unclear.

Best regards,
 The Authors

Thank the authors for the additional experiments. They address some of my concerns regarding the pretraining data and training recipe. The results suggest that the data mixture contributes substantially to the reported gains, while the incremental benefits of JEPA-style learning are more task-dependent. Further controlled ablations under matched large-scale pretraining conditions would help distinguish the contributions of the architecture, learning objectives, and data mixture. I will therefore maintain my original score.

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

The paper presents PHASE, a physiology-guided foundation model for intracranial EEG. PHASE-T combines masked latent prediction with explicit physiological targets within each channel, while PHASE-ST adds cross-channel targets over synchronized recordings. Pretraining uses heterogeneous recordings from 222 participants and nine sites. Frozen PHASE-T outperforms the compared foundation models on five Omni-iEEG clinical tasks, fine-tuning improves all reported task-specific comparisons, and the ablations suggest that physiological supervision is more useful than waveform reconstruction. The topic is important and the experimental coverage is strong, but the paper needs tighter control of cohort overlap, probe capacity, and clinical generalization claims.

Critical strengths
  1. The paper has a clear and plausible pretraining principle: physiological measurements are made explicit targets instead of being left as an implicit consequence of reconstruction. The temporal hierarchy is well aligned with the different timescales of iEEG phenomena.
  2. The evaluation is broad. One frozen PHASE-T representation is tested on HFO, ictal, sleep, anatomy, and pathological-region tasks, while PHASE-ST is evaluated on multichannel auditory decoding. This is substantially more informative than a single downstream task.
  3. The reported gains are meaningful and the matched ablations are useful. For example, physiological supervision increases frozen HFO macro-F1 from 0.63 to 0.81 over the latent-only variant, and the paper reports improvements in both cross-institution transfer and spatial decoding.
  4. The manuscript is relatively transparent about pretraining exclusions, model size, update counts, and compute, and it acknowledges the one-minute context limitation and the need for broader cohorts.
Critical weaknesses
  1. The paper relies heavily on comparisons across datasets and published numbers. Please clarify whether all competing foundation models use the same preprocessing, channel selection, input duration, linear-probe regularization, and participant-level split. A small difference in readout capacity or aggregation could materially affect the comparison.
  2. The nine-site pretraining corpus is a strength, but the analysis of site and participant composition is not sufficiently central. The authors should report per-site sample counts, recording hardware or sampling differences, and results separately by held-out institution. It would also help to quantify whether performance is driven by a subset of high-resource sites.
  3. The physiology targets may be computed from the same signal windows used as inputs. The manuscript should explain how target estimation avoids leakage from preprocessing, filtering, or future context, especially for event-related targets and spectral quantities. A temporal-shuffle or target-corruption control would test whether the gains come from meaningful physiology rather than additional signal statistics.
  4. The clinical claims should be qualified. The model improves proxy tasks and a resection-ratio outcome metric, but the paper does not demonstrate prospective seizure localization, calibration at the patient level, or decision-curve utility. The Outcome AUC should be accompanied by patient counts, confidence intervals, and comparison to simple clinical or score-based baselines.
  5. PHASE-ST is described as learning cross-channel relationships without fixed identities or coordinates, yet its pretraining uses synchronized 64-channel sets. Please show robustness to missing channels, different channel counts, channel permutations, and clinically realistic montage changes.
Questions for authors
  1. Are preprocessing, channel selection, input duration, probe regularization, and patient splits identical for all foundation-model comparisons?
  2. Please provide per-site participant counts and held-out-site results with confidence intervals.
  3. How are physiological targets computed, and do they use any future context or information unavailable at deployment time?
  4. What are the patient counts and uncertainty estimates for the Outcome AUC, and how does it compare with simple clinical baselines?
  5. How does PHASE-ST behave with missing channels, channel permutations, and montages other than synchronized 64-channel sets?
5 comments
AuthorsOpenReview ↗

Response to Reviewer wycu 1/4

We thank Reviewer wycu for the constructive review and for recognizing the physiological learning principle, the breadth of the evaluation, and the value of the matched ablations. We added new analyses: uncertainty estimates for the four held-out institutions; a PHASE-T model pretrained on BrainTreebank alone, a single site, to test whether performance is driven by a few high-resource sites; confidence intervals for the Outcome AUC and a comparison with two simple baselines; and a missing-channel analysis of PHASE-ST. We also clarify the evaluation protocol, the per-institution composition of the pretraining corpus, the target construction, and the scope of the clinical evaluation. The detailed responses follow.

[Q1, W1] The comparisons rely on cross-dataset and published numbers. Are preprocessing, channel selection, input duration, linear-probe regularization, and participant-level splits identical for all foundation-model comparisons, given that small differences in readout capacity or aggregation could materially affect them?

Thank you for this important question. The comparison is matched in everything except each model’s own input front-end. In each comparison, all models receive the same channels and the same 60-s windows or HFO events from the same participant-level splits, and are read out by the same linear probe with the same regularization (Appendix D.2.1). Each model, including PHASE-T, applies its own front-end to these inputs, as its released checkpoint requires (Appendix D.2.2). For MAYO/FNUSA, we follow MVPFormer's evaluation protocol, with the first four participants used for training (Appendix D.4), and for BrainTreebank, BaRISTA's recording split and PopT's evaluation protocol (Appendix D.5).

AuthorsOpenReview ↗

Response to Reviewer wycu 2/4

[Q2, W2] Please report per-site participant counts and recording hardware or sampling differences, provide held-out-institution results with confidence intervals, and quantify whether performance is driven by a subset of high-resource sites.

Thank you for the helpful suggestion. Appendix A.1 reports participants, recordings, and sampling shares for each pretraining source, and sampling is balanced across participants within each source. Native sampling rates differ across releases. All recordings are resampled to 1000 Hz, and physiological targets use only frequencies within each recording's usable bandwidth (Appendix A.2). The participants of each release come from the following institutions. Hours are total recording durations.

ReleaseInstitutionParticipantsRecordingsHours
Omni-iEEGDetroit (Wayne State University)81818.0
Omni-iEEGUCLA303046.7
Omni-iEEGZurich1225020.6
Omni-iEEGHospital of the University of Pennsylvania9181.5
Omni-iEEGJohns Hopkins881.6
Omni-iEEGNational Institutes of Health781.2
Omni-iEEGUniversity of Miami440.1
HatanoDetroit (Wayne State University)12914485.0
NYU PodcastNew York University884.0
BrainTreebankBoston Children's Hospital101737.3

Omni-iEEG's Detroit participants and the Hatano release share 76 participants (Table 5), so the corpus includes 222 distinct participants from nine sites. Omni-iEEG test participants are excluded from both sources (Appendix A.1).

Tohoku, NCNP, MAYO, and FNUSA contribute no pretraining data (Section 4), and Table 3 reports each of them separately. The Omni-iEEG-trained pathological-channel probe is applied unchanged to held-out Tohoku and NCNP (Table 3; Appendix D.3), and we now add participant-bootstrap intervals: ROC-AUC 0.7775 (95% CI: 0.6588–0.8815; 26 participants) and 0.6292 (0.4826–0.7713; 30 participants). At MAYO and FNUSA, where four labeled participants per institution are used for adaptation (Appendix D.4), we repeated the fine-tuning in Table 3 with two additional random seeds, keeping all other settings fixed. Across the three runs, F1 is 0.7418 ± 0.0092 at MAYO and 0.6651 ± 0.0052 at FNUSA (mean ± standard deviation across seeds, the form in which NeuroCLUS reports 0.40 ± 0.02 and 0.51 ± 0.02).

To quantify whether performance is driven by a subset of sites, we retrained PHASE-T for 60K updates on the BrainTreebank pretraining split alone: 10 participants from a single site, recorded during naturalistic listening. This removes the other eight sites from pretraining, including Detroit and UCLA, the two largest sites by recording hours, and all clinical-monitoring recordings. The resulting model remains comparable to or better than the released foundation models in Table 1 on most scores, so PHASE-T does not rely on the largest sites or on clinical-monitoring data; adding them improves it further.

ModelHFO F1Ictal F1Sleep F1Lobe-5 F1Region-12 F1Channel AUCOutcome AUC
PHASE-T, full corpus0.81380.89580.74130.48220.33980.83940.6369
PHASE-T, BrainTreebank only0.79770.85050.68570.46980.30990.80720.6429
Best released foundation model (Table 1)0.79370.83230.72130.40540.25920.80550.6017
AuthorsOpenReview ↗

Response to Reviewer wycu 3/4

[Q3, W3] How are the physiological targets computed from the input windows, and do they avoid leakage from preprocessing, filtering, or future context, especially for event-related and spectral targets? Would a temporal-shuffle or target-corruption control show that the gains come from meaningful physiology rather than additional signal statistics?

Thank you for this detailed question. The physiological targets serve only as pretraining supervision. Together with their fixed standardization statistics, they enter only the pretraining loss and are never computed at deployment, where the model receives only the waveform. Each target is measured from the same 60-s window that the encoder receives, with no annotations and no data from outside the window. The targets are wavelet power from 2 to 450 Hz, amplitude and local variation, and high-frequency power in the ripple and fast-ripple ranges (Appendix B.2). Pretraining and evaluation use the same high-pass and notch filters (Appendix A.2).

Two controls in the manuscript test whether the gains come from additional signal statistics rather than from the physiological targets. First, a linear probe on the targets themselves, without the encoder, is worse than PHASE-T on all seven scores (Appendix D.2.3), so the targets alone do not explain the gains. Second, replacing the targets with reconstruction of the masked waveform, a target that contains all of the masked signal, is also worse on all seven scores (Latent+Recon, Table 2), so supervision with more of the signal does not reproduce them.

[Q4, W4] The clinical claims should be qualified, since prospective seizure localization, patient-level calibration, and decision-curve utility are not demonstrated. What are the patient counts and confidence intervals for the Outcome AUC, and how does it compare with simple clinical or score-based baselines?

Thank you for the valuable suggestion. Our clinical evaluation follows Omni-iEEG, a published clinical iEEG benchmark, and uses its tasks and evaluators unchanged. Its pathological brain region identification task scores channels from preoperative interictal recordings and evaluates the scores with two criteria (Omni-iEEG, Section 4.2.2): separating SOZ channels from preserved channels of seizure-free patients (Channel), and predicting postoperative seizure freedom from the share of predicted pathological activity within the actual resection (Outcome). Outcome is therefore not a separate prediction but is computed from the same pathological-channel scores (Section 4.2). It tests whether patients whose resection removed more of the predicted pathological activity were more likely to become seizure-free. Although the task mirrors presurgical localization, it is built from evidence available only after surgery, the resection and postoperative outcome, because the pathological region cannot be determined prospectively (Omni-iEEG, Section 4.2.2). This evaluation therefore does not establish prospective seizure localization, nor does it assess patient-level calibration or decision-curve utility. As our Ethics Statement notes, PHASE is intended for research, and clinical use would require further validation.

We added an analysis of Outcome uncertainty and simple baselines on the same cohort and evaluator. The cohort has 95 participants (63 seizure-free and 32 non-seizure-free). Besides fine-tuned PHASE-T, we computed the resection ratio from two channel scores that need no model. A uniform score gives every window the same value, so the ratio reduces to the extent of the resection. The clinical SOZ label measures how much of the marked SOZ was resected. Intervals come from a participant bootstrap stratified by outcome, with 2,000 resamples.

Channel scoreOutcome AUROC (95% CI)
Fine-tuned PHASE-T0.7713 (0.6731–0.8581)
Clinical SOZ label0.7247 (0.6185–0.8242)
Uniform score (extent of resection)0.5890 (0.4583–0.7173)

PHASE-T is well above chance and exceeds the uniform score by 0.182 (paired 95% CI: 0.088–0.281). The clinical SOZ is marked from seizures recorded during monitoring, while PHASE-T scores only interictal recordings. Even so, PHASE-T reaches outcome discrimination comparable to the clinical SOZ (paired difference 0.047; 95% CI: −0.046 to 0.144).

Comments on the paper

1 comment
AuthorsOpenReview ↗

Final Remarks by Authors

Dear Reviewers and Area Chairs,

We sincerely thank all reviewers and ACs for their insightful and constructive feedback throughout the review process. The reviewers consistently highlighted the core strengths of PHASE as a foundation model for iEEG:

  • A clear and plausible pretraining principle. PHASE makes physiological measurements explicit learning targets, and its components are well justified.

  • Broad evaluation demonstrating transferable representations. One frozen encoder performs consistently well across clinical tasks and institutions.

  • Meaningful gains supported by matched ablations. The model is lightweight and already reaches state-of-the-art performance.


The remaining questions raised during the initial review fell into two categories: clarifications that were already presented in the main text or appendix, and suggestions for additional analyses.

During the discussion, we addressed all points by referencing existing sections and by adding new analyses. Two of the new analyses respond directly to the reviewers’ main questions: an ablation of the pretraining objectives and an ablation of the pretraining data. Both are now in the main text of the revised manuscript (Tables 2 and 3). We also extend our analysis and report uncertainty estimates for the results at unseen institutions and on surgical outcomes. We further test how the model scales with size and how robust it is to missing channels. All changes are marked in blue.

We are grateful for the constructive discussion with reviewers, which further highlighted the two main contributions of PHASE stated in the Introduction. The original results establish both, and the new analyses reinforce them:

  • Contribution 1: Physiological supervision leads to more informative representations across temporal and spatial scales. In matched ablations, physiological supervision improves every Omni-iEEG score. In the data-matched comparison requested by the reviewers, PHASE-T remains comparable to or better than BrainBERT and BaRISTA.

  • Contribution 2: Diverse pretraining and superior generalization to unseen institutions. Pretrained on a heterogeneous iEEG dataset, PHASE reaches state-of-the-art performance and generalizes to unseen institutions. In the new ablation, heterogeneous pretraining improves PHASE-T over pretraining on a single site.


After reading our responses, Reviewer r2L9 stated that we “addressed my concerns regarding the training dataset,” and Reviewer qyZ2 stated that the additional experiments “address some of my concerns regarding the pretraining data and training recipe.” Since our rebuttal directly addressed each reviewer’s comments, we believe the discussion confirms the completeness and relevance of the contribution.

PHASE offers iEEG research a foundation model grounded in the physiological characteristics that clinicians and neuroscientists already read. Its two main contributions are complementary. Physiological supervision gives a principle for learning iEEG representations. Diverse pretraining gives a lightweight encoder that outperforms every evaluated foundation model on the Omni-iEEG clinical tasks and generalizes to unseen institutions.

We sincerely appreciate the reviewers', ACs', and PCs' time, effort, and thoughtful engagement throughout this process.

Best,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.