acceptodds
Under review as a conference paper at ICLR 2027

MCD-EEG: Can a Non-Stationary Spatio-Temporal EEG Benchmark Evaluate Brainprint Recognition Across Tasks, Re-caps, and Populations?

Abstract

Electroencephalography (EEG), owing to its high temporal resolution, non-invasive acquisition, convenient deployment, and compatibility with wearable devices, has become a fundamental technology for brain–computer interfaces (BCIs) and task-based brainprint recognition. However, existing public EEG datasets are generally limited to a single cognitive task and a single acquisition session, lacking realistic distribution shifts across tasks, populations, and acquisition stages (e.g., EEG cap removal and reattachment, referred to as re-cap). Consequently, they are insufficient for systematically evaluating the robustness, generalization, and fairness of EEG representation learning and brainprint recognition models in real-world scenarios. To address these challenges, we introduce MCD-EEG, a large-scale EEG benchmark designed for cross-temporal robust and neurodiversity-aware representation learning. MCD-EEG contains recordings from 45 participants, including 22 prelingually deaf individuals and 23 hearing controls, collected using 60-channel high-density EEG together with behavioral measurements. A multi-stage cognitive paradigm consisting of reading, interference, and delayed recall tasks was employed, with repeated recordings acquired across sessions. Notably, the second session involved complete EEG cap removal and reattachment, explicitly introducing realistic cross-temporal session drift. Based on this dataset, we further establish standardized benchmark tasks for within-task, cross-task, and cross-re-cap brainprint recognition, providing a new testbed for developing robust and fair EEG representation learning algorithms.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Electroencephalography, brainprint recognition, cross-session drift, neurodiversity, EEG benchmark

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_ijX7OpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces MCD-EEG, a benchmark for EEG-based brainprint recognition across cognitive tasks, acquisition conditions, and populations. The dataset includes EEG recordings from 45 hearing and prelingually deaf participants, covering multiple cognitive tasks and EEG cap removal and reattachment (Re-cap). The authors establish within-task, cross-task, and cross-Re-cap evaluation protocols and assess several existing EEG models through performance comparisons and feature analyses.

Critical strengths

The paper considers multiple sources of distribution shifts in EEG-based identity recognition, including different cognitive tasks, populations, and acquisition changes caused by EEG cap removal and reattachment. Compared with datasets focusing on a single task or acquisition session, this benchmark provides more diverse evaluation scenarios and facilitates research on the robustness and generalization of EEG representations.

Critical weaknesses

1、The paper treats Re-cap as a core evaluation scenario. However, the Re-cap stage involves not only removing and reattaching the EEG cap but also changing the cognitive task from memory retrieval to an IQ assessment. Consequently, the observed performance degradation may result from both electrode repositioning and cognitive task changes, making it difficult to attribute the degradation specifically to Re-cap. I suggest adding a control experiment in which participants perform the same cognitive task before and after EEG cap reattachment to more accurately assess the distribution shift introduced by Re-cap.

2、Some experimental details need further clarification and consistency. For example, the main text represents EEG samples as 60 × 1600, whereas the Re-cap stage is described as extracting 2000-ms EEG segments. These descriptions are inconsistent in terms of temporal length. In addition, some important details, such as model training hyperparameters and specific data splitting procedures, are insufficiently documented. I suggest clarifying these details, ensuring consistency throughout the paper, and providing complete experimental configurations to improve the reproducibility of the benchmark.

3、1. The main contributions of this paper lie in data collection and benchmark construction, without proposing a new algorithm or representation learning method. The cognitive experimental paradigms and baseline models are also largely based on existing research. Although the work involves substantial effort, it primarily represents an integration of existing experimental designs and evaluation methods, with limited additional methodological innovation.

Questions for authors

In the within-task and cross-task identity recognition experiments, models such as EEGNet and FBCNet achieve 100% accuracy under multiple settings. Could these results potentially be affected by data leakage? I would appreciate further clarification on the specific data preprocessing and train-test splitting procedures, particularly whether there are overlapping or highly correlated EEG segments between the training and test sets, and whether any preprocessing steps utilize information from the test set.

7 comments
AuthorsOpenReview ↗

Re-cap Control Experiment with 10 Scheduled Participants (Expected Completion: November 11) and Preliminary Responses to Other Concerns

We sincerely thank the reviewer for the constructive comments and suggestions. We address the concerns as follows.

  1. Re-cap Control Experiment. We agree that the original Re-cap protocol may introduce confounding effects from cognitive task differences. To address this concern, we have scheduled 10 participants for an additional controlled experiment, in which each participant will perform the same cognitive task before and after EEG cap removal and reattachment. This paired design will allow us to more directly evaluate the acquisition-related distribution shift while controlling for cognitive task differences. We expect to complete the experiment by November 11 and will report the results during the discussion period.
  2. Experimental Details and Consistency. We apologize for the unclear description. The Re-cap EEG samples used for identity recognition have a dimension of 60 × 1600, corresponding to 60 channels and a 4-second recording at 400 Hz. We clarify that the previously mentioned 2,000 ms refers to the inter-trial interval (ITI) in the experimental paradigm, rather than the ERP analysis window. Specifically, ERP analysis uses a 1,500-ms poststimulus window with a 200-ms prestimulus baseline. These timing parameters serve different purposes and should not be confused with the 4-second EEG inputs used for identity recognition. We have clarified these distinctions and corrected the inconsistent descriptions in the revised manuscript, along with detailed preprocessing procedures, training hyperparameters, and data partitioning protocols.
  3. Methodological Novelty. We acknowledge the reviewer's concern regarding methodological novelty. The primary contribution of this work is not a new recognition architecture, but a benchmark designed to systematically evaluate EEG identity recognition under cognitive, population-related, and acquisition-related distribution shifts. We will further clarify the distinctive contributions and research value of these evaluation protocols. The additional Re-cap control experiment will also strengthen the validity of the proposed benchmark.
  4. High Recognition Accuracy and Data Leakage. We thank the reviewer for raising this concern. We confirm that there is no data leakage between the training and test sets. The high recognition accuracy can be attributed to the fact that the Reading and Memory Retrieval stages were recorded from the same participants within a single EEG acquisition session, where individual-specific EEG characteristics remain relatively stable. This observation is consistent with previous research. For example, Jin et al. (IEEE TCDS, 2024), in Disentangled Adversarial Generalization Network for Cross-Session Task-Independent Brainprint Recognition, reported an average cross-task identification accuracy of 99.50% on the single-session MTED dataset. These findings support the feasibility of near-perfect identity recognition across cognitive tasks within the same acquisition session. We will further clarify the data partitioning and preprocessing procedures in the revised manuscript to address the reviewer's concerns regarding reproducibility and data leak
AuthorsOpenReview ↗

Response to Reviewer ijX7(Part 1/3: Comments 1)

Thank you for recognizing the value of evaluating multiple sources of EEG distribution change and for recommending matched-condition controls. Following our November 8 interim response, we have completed the additional ten-participant experiments. The new results address the task confound directly, and we clarify the segment definitions, training procedures, and evidence for train–test separation below.

1. Same-condition controls for the Re-cap evaluation

The completed controls show that substantial cross-recording degradation persists when the cognitive condition is held constant. The original Memory-to-IQ/Re-cap comparison changes both task and recording visit: IQ/Re-cap was recorded three days after Reading, Interference, and Memory Retrieval. We therefore added same-condition measurements to address this confound.

Between November 6 and 11, 2026, ten additional participants completed two recordings separated by 12 hours, with cap removal and reapplication between recordings. Section 4.4 and Table 6 evaluate resting state and the same 0-back, 1-back, or 2-back condition across recordings. FBCNet achieves resting-state accuracy of 79.66% ± 5.83%, while EEGNet achieves 68.23% ± 1.67%, 70.77% ± 3.05%, and 71.55% ± 2.41% for the three N-back conditions, respectively.

Table 9 further separates load and recording changes using the same participants' 0-back and 2-back data. For all eight models, task-specific training uses only first-recording 0-back, followed by four tests:

  • First-recording 0-back: the reference condition.
  • First-recording 2-back: load changes.
  • Second-recording 0-back: recording changes while load is retained.
  • Second-recording 2-back: both change.

Across all eight models, changing load within the first recording reduces accuracy by 0.00–6.42 percentage points, whereas changing the recording at the same load reduces it by 29.42–63.33 percentage points. Within each fold, the four tests share the same ten identities, selected checkpoint, and test-set size. These results directly support recording-associated changes as the larger source of generalization loss under this protocol.

The controls use resting state and N-back in an additional cohort. They do not separately estimate the contributions of cap reapplication, elapsed time, and physiological variation. We have aligned the manuscript's interpretation with the recording-associated effect established by these comparisons.

AuthorsOpenReview ↗

Response to Reviewer ijX7(Part 2/3: Comments 2)

2. Consistent segment lengths and complete experimental configurations

The identity-recognition tensor dimensions should be interpreted using the analysis-specific sampling rate and window:

AnalysisInput or analysis windowSampling rateDuration or role
Original Reading, Memory Retrieval, and IQ/Re-cap identity recognition60 channels × 1,600 samples400 Hz4-second identity-recognition segments
Supplementary resting-state and N-back identity recognition60 channels × 800 samples400 Hz2-second identity-recognition segments
ERP analysis1,500 ms after stimulus onset, with a baseline from −200 to 0 ms400 Hz after downsamplingSeparate event-related analysis

We clarify that the previously mentioned 2,000-ms duration is a timing parameter of the experimental paradigm, distinct from the ERP analysis window. Specifically, the paradigm uses a 2,000-ms inter-trial interval, while ERP analysis is conducted over a 1,500-ms post-stimulus window with a 200-ms prestimulus baseline. We have clarified this distinction in the revised manuscript and appendix.

Preprocessing and training. The documented acquisition preprocessing includes downsampling to 400 Hz, 0.1–45 Hz filtering, bad-channel processing, ICA-based artifact removal, and re-referencing. For the original Reading/Memory/IQ evaluation, data-dependent preprocessing decisions and ICA parameters are fitted using training data within each fold and then held fixed for validation and testing. Classifier-input normalization is also fitted only on the corresponding source training partition.

The local baselines use AdamW and cross-entropy with label smoothing of 0.1. Learning rate is 0.001, except for EEGViT (0.0003) and the local EEGPT-named implementation (0.0005). Appendix E.3 and Tables 13–14 report architecture- and task-specific weight decay, batch size, and epoch budgets. Supplementary experiments use weight decay 0.01, batch size 32 except for BrainprintNet (8), gradient-norm clipping at 5, and at most 80 epochs, with a minimum of 20 epochs and early-stopping patience of 15. Checkpoint selection uses the lowest source-validation cross-entropy.

The original EEGPT-named baseline is a local implementation trained from scratch. We distinguish it from the additional evaluation using publicly released EEGPT weights, with a frozen encoder and a source-trained linear readout, reported in Table 11 and used for the EEGPT entry in Table 9.

Data partitioning. Within-task identity evaluation holds out complete epochs for testing, with validation confined to the remaining source data. Reading-to-Memory and Memory-to-IQ transfer use five source-side training/validation folds; each selected model is evaluated on the complete target recording. Their mean and standard deviation summarize five source-trained models on a fixed target set. Table 9 uses five blocked outer test folds, each with 720 source training segments, 240 separate source validation segments, and 240 test segments per condition. Validation and test blocks are disjoint. We have made these evaluation roles explicit in the captions and methods.

Reviewer_e4hXOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes MCD-EEG, a brainprint benchmark consisting of 60-channel EEG recorded from 22 prelingually deaf and 23 hearing participants. The claimed contributions are (1) the first benchmark addressing re-cap (cap removal and reattachment) shift, (2) a paradigm that induces cross-task shift, and (3) a neurodiversity-aware fairness evaluation. Eight models are evaluated on within-task, cross-task, and cross-re-cap identity recognition. Within-task and cross-task identification is near-perfect (FBCNet/EEGNet reach 1.0), but accuracy drops sharply after re-cap, from which the authors conclude that sensor-level shift is the main barrier.

Critical strengths
  • The deaf cohort is characterized in detail, including degree of hearing loss, assistive-device use, and sign-language exposure, and is matched to hearing controls on age, reading ability, and IQ. This is the one component that is hard to obtain from existing public resources, which is why it matters. However, the paper does not turn this cohort into new analyses or findings, so its weight in my assessment is limited.
Critical weaknesses
  • Cross-session shift involving cap reattachment is already inherent in multi-day datasets such as SEED and BCI IV-2a, and cross-task identification can be studied with PhysioNet EEGMMIDB, among others. Combining these resources would allow the same evaluations at larger scale and with real between-day intervals. In addition, no task or condition labels are defined in the benchmark (the only labels are identity and group), so the carefully designed cognitive paradigm is never used as a decoding target. Since a dataset paper’s contribution lies in what it enables beyond existing resources, this weakness largely determines my recommendation.

  • That within-session identification is near ceiling and drops sharply across sessions, and that deaf readers show different ERPs, are both well-known facts. The conclusion that sensor shift is the main barrier has already been studied extensively using datasets that provide multiple sessions. Furthermore, the within-task and cross-task tracks are saturated at 1.0 and therefore cannot discriminate between methods. As a result, the paper establishes neither new scientific insight nor benchmark utility.

Questions for authors
  • What research question can be answered only with MCD-EEG and not with a combination of existing datasets?

  • How does re-cap differ from the cap reattachment that occurs in existing multi-session recordings?

  • Are task-related labels, such as the CN/IN/CT/IT condition labels and behavioral responses, provided with the dataset?

10 comments
AuthorsOpenReview ↗

Response to Reviewer e4hX (Part 1/3: Comments 1)

Thank you for highlighting the need to distinguish the contribution of MCD-EEG from established EEG resources. We address this through the characterized deaf/hearing cohort, the additional matched load-and-recording evaluation, and explicit cognitive-stage prediction targets. We also clarify which conclusions these data support.

1. What research question does MCD-EEG support beyond combining existing datasets?

MCD-EEG supports two related research questions: Can Reading and Memory Retrieval be distinguished from EEG in previously unseen hearing and prelingually deaf participants, and are model robustness patterns consistent across these cohorts? These questions are investigated within a shared language-cognition protocol, using cohorts characterized for hearing history and matched on age, reading ability, and nonverbal reasoning.

First, the cognitive paradigm is directly evaluated as a prediction target. Appendix E.2 and Table 8 report Reading-versus-Memory Retrieval decoding using participant-disjoint five-fold evaluation. All trials from a participant remain in the same outer fold, so testing evaluates participants not encountered during training. The cohorts are pooled for training, and performance is reported separately for hearing and deaf test participants.

MS-EEGNet-SF achieves 66.47% ± 2.82% accuracy in the hearing cohort, while EEGNet-lite-Ens2 achieves 66.54% ± 4.72% in the deaf cohort, compared with a 50% chance level for this balanced binary task. These results provide evidence of predictable stage-associated information in both cohorts. They also establish a non-saturated evaluation using cognitive-stage labels beyond participant identity and population membership. We have made this auxiliary prediction task explicit in the main benchmark description.

Second, the cohort enables subgroup-specific evaluation of identity robustness. Table 5 shows that EEGNet outperforms FBCNet in the hearing cohort (51.90% versus 48.23%), whereas FBCNet outperforms EEGNet in the deaf cohort (82.39% versus 73.58%). This descriptive reversal demonstrates the relevance of examining model rankings separately across cohorts. It should not be interpreted as a statistically established interaction or a causal effect of deafness.

The added value therefore lies in the combination of characterized deaf/hearing cohorts, linked Reading and Memory Retrieval observations, behavioral annotations, and repeated acquisition under a shared protocol. Combining datasets that do not contain these cohorts and linked observations cannot reconstruct this comparison merely by pooling samples.

We acknowledge that cognitive decoding and cross-session EEG recognition can be studied using existing resources, including BCI IV-2a and M3CV. Our claim is that MCD-EEG provides a complementary population and experimental setting for investigating these questions, supported by the cognitive-stage results in Table 8 and subgroup robustness results in Table 5. This complementary design supports subgroup comparisons that cannot be directly reconstructed by pooling existing datasets.

AuthorsOpenReview ↗

Response to Reviewer e4hX (Part 2/3: Comments 2-3)

2. How does re-cap differ from cap reattachment in existing multi-session recordings?

We acknowledge that the physical re-cap procedure is not fundamentally different from cap reattachment in existing multi-session EEG studies. However, MCD-EEG enables identity-recognition evaluation across cognitive tasks and recording sessions under a unified protocol, together with subgroup-specific robustness comparisons between prelingually deaf and hearing participants.

We clarify that the original acquisition involves a genuine cross-day interval: Reading, Interference, and Memory Retrieval were acquired during the initial visit, while IQ/Re-cap was acquired three days later, as stated in Appendix D. We have made this timing consistent throughout the manuscript.

However, the original Memory-to-IQ/Re-cap comparison changes both task and recording visit. Tables 2–4 provide within-session and cross-task references, but do not isolate electrode repositioning in Table 5.

To further strengthen this comparison, we added Table 9 to the revised manuscript. Between November 6 and 11, 2026, we collected new EEG data from ten additional participants, each undergoing two recording sessions separated by 12 hours. All eight models are trained only on first-recording 0-back and evaluated under four conditions:

  • First-recording 0-back: neither load nor recording changes.
  • First-recording 2-back: load changes.
  • Second-recording 0-back: recording changes.
  • Second-recording 2-back: both change.

Five blocked outer folds separate training, validation, and testing. Each fold uses 720 source training segments, 240 independent source validation segments, and 240 test segments per condition. All four tests use the same selected checkpoint and corresponding held-out index blocks. Normalization and checkpoint selection use only source training/validation data.

Across all eight models, the same-recording load change reduces accuracy by 0.00–6.42 percentage points, while the matched-load recording change reduces accuracy by 29.42–63.33 percentage points. These contrasts are obtained within the same cohort and evaluation design.

Nevertheless, the recording change includes elapsed time, cap reapplication, and other physiological/acquisition variation. We have replaced the claim that sensor drift is the established dominant cause with the supported observation that the recording-associated contrast is larger than the load-associated contrast under this protocol.

3. Does the dataset contain task-related labels and behavioral responses, and is the cognitive paradigm used as a decoding target?

Yes. The trial-level metadata contain participant and trial identifiers, semantic correctness, sentence type, and the four conditions:

  • CN: Correct Normal;
  • IN: Incorrect Normal;
  • CT: Correct Tongue Twister;
  • IT: Incorrect Tongue Twister.

The original experiment contains 2,880 trials, comprising 45 participants × 64 trials, with 16 trials per condition per participant. These trial identifiers link the Reading and Memory Retrieval epochs. Behavioral observations include accuracy and response time for Reading Judgment (RA1/RT1) and Memory Retrieval/cloze testing (RA2/RT2). Section 2.2 defines the conditions, and Appendices E.6 and E.8 describe the behavioral methods and results. We have made the annotation dictionary and trial-index linkage explicit in the revised data documentation.

Furthermore, Appendix E.2 and Table 8 directly evaluate Reading-versus-Memory Retrieval decoding, using cognitive stage as the prediction target. This balanced binary task uses participant-disjoint five-fold evaluation, with all trials from a participant assigned to one outer fold. Training pools the cohorts, and test performance is reported separately for hearing and deaf participants.

MS-EEGNet-SF achieves 66.47% ± 2.82% accuracy in the hearing cohort, while EEGNet-lite-Ens2 achieves 66.54% ± 4.72% in the deaf cohort, against a 50% chance level. Thus, the revised evaluation uses a cognitive-stage target beyond identity and population labels.

AuthorsOpenReview ↗

Response to Reviewer e4hX(Part 3/3: Comments 4)

4. How is the benchmark useful when some identity-recognition tracks saturate?

The benchmark is designed to evaluate EEG representations under progressively more challenging conditions. Tables 2–4 establish within-session and cross-task reference performance, while Tables 5 and 9 examine cross-recording generalization. Table 8 further extends the evaluation to participant-independent cognitive-stage decoding. Together, these complementary protocols assess different aspects of EEG representation robustness beyond within-session recognition accuracy.

For example, Table 9 reports:

ModelRecording 1: 0-backRecording 1: 2-backRecording 2: 0-backRecording 2: 2-back
FBCNet100.00%100.00%70.58%71.17%
EEGNet100.00%100.00%68.08%68.75%
EEGPT, publicly pretrained frozen encoder90.33%86.83%57.00%51.67%

EEGConformer additionally achieves 99.75% for first-recording 0-back but 36.42% for second-recording 0-back. Therefore, near-perfect within-recording performance does not resolve the comparison of cross-recording performance. Table 8 provides another unsaturated evaluation with a different prediction target and participant-disjoint testing. Publicly pretrained EEGPT is explicitly distinguished from the earlier local implementation; its additional evaluations are reported in Table 11.

Although cross-session performance variations and group-related ERP differences have been investigated in prior studies, the distinctive value of MCD-EEG lies in providing a well-annotated dataset with reproducible evaluation protocols that integrate cognitive-stage decoding, cross-recording identity recognition, and robustness assessment across prelingually deaf and hearing cohorts. This resource enables systematic comparisons of EEG model performance across participant subgroups and experimental conditions, while providing complementary empirical evidence for investigating the relationships among cognitive tasks, recording conditions, population characteristics, and EEG representations. We have revised the abstract, contribution statement, Table 1, and Section 4.2 so that the claims match this evidence and its limitations.

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

This paper introduces MCD-EEG, an EEG benchmark of 45 participants, namely 22 prelingually deaf and 23 hearing participants, recorded with 60 channel EEG during a multi-stage cognitive paradigm of reading judgment, arithmetic interference, and delayed recall, plus EEG cap removal and reattachment, which the authors call Re-cap. Four benchmark tasks are defined: hearing versus deaf population classification, within-task brainprint identification, cross-task identification from reading to memory, and cross-Re-cap identification from memory to the Re-cap stage. The strongest baselines reach near perfect within-task and cross-task identification, while all models degrade sharply under Re-cap, with architecture and subgroup dependent robustness. The dataset includes behavioral and ERP validation of the paradigm.

Critical strengths
  1. The benchmark includes an explicit Re-cap condition in which the EEG cap is removed and reattached between recordings, which corresponds to a situation that occurs in real deployment. The results show that a change of cognitive task has little effect on identification, while Re-cap degrades performance for all models tested. The dataset is therefore a useful testbed for methods that address sensor related distribution shift.
  2. The deaf cohort is a distinctive part of the resource. The paper reports audiological inclusion criteria for the deaf group and shows that the two groups are matched in age, reading ability, and nonverbal reasoning. Identification is evaluated separately for the deaf and hearing groups, which makes the dataset suitable for fairness analyses.
Critical weaknesses
  1. The central cross-temporal framing is not clearly supported. Section 2.2 presents Re-cap as an additional stage that follows the first three stages, which reads as one recording visit, while the abstract calls it a second session, Section 2.3 calls it a cross-session re-setup protocol, Appendix D splits the data into Session 1 and Session 2, and Table 1 marks the dataset as longitudinal. No time interval between the recordings is reported anywhere. If Re-cap occurred within the same sitting, the benchmark measures electrode repositioning noise rather than cross-temporal drift, and the longitudinal claim must be reworded. By the definition implied by its own Table 1, where SEED and BCI IV-2a are marked as longitudinal because their sessions were recorded on different days, a same-day Re-cap would not qualify as a longitudinal session.
  2. The core Re-cap result is confounded and under-specified. Task 4 changes the cognitive task and the sensor setup at the same time, since models are trained on the memory stage and tested on recordings from a standardized IQ assessment that is described in a single sentence, so attributing the performance cliff to sensor drift alone requires a same-task control across Re-cap that is absent. In addition the five-fold cross-validation in the Table 4 and Table 5 captions contradicts the zero-shot definitions of both cross-condition tasks, and the number of Re-cap segments per subject is unreported, which makes the key result hard to interpret or reproduce.
Questions for authors
  1. The paper does not state whether all stages were recorded within a single visit or across different days. Section 2.2 presents Re-cap as an additional stage that follows the first three stages, while the abstract, Section 2.3, Table 1, and Appendix D describe a second session and mark the dataset as longitudinal. Were the Re-cap recording and the first three stages collected on the same day, and what was the time interval between them? This determines whether the cross-temporal and longitudinal framing is justified.
  2. How do you decouple the sensor shift from the task shift in Task 4, given that training uses memory stage data and the Re-cap data come from an IQ assessment? Would a same-task control across Re-cap, for example repeating the memory task after re-wearing the cap, preserve the performance cliff? How many segments per subject are available in the Re-cap condition?
  3. Please clarify the protocol behind Table 5. What exactly is cross-validated when models are trained on all memory data and tested on Re-cap data? Please also report baseline training details, including hyperparameters, the tuning protocol, and whether EEGPT uses public pretrained weights, so that the claimed strict separation between training and test data can be verified.
  4. For the population classification task in Table 9, were the five cross-validation folds subject independent? With 5632 trial level connectivity matrices, trial level folding would leak subject information and inflate results. Also, the paper states that the 5632 matrices come from 45 participants across 8 stimulus conditions and 16 trials, which would yield 5760 matrices, while 5632 corresponds to 44 participants at 128 matrices each. Please explain how the total arises, for example whether a participant or a subset of trials was excluded from this analysis.
  5. Two clarifications. First, the condition labels CN, IN, CT, and IT in Section 2.2 are never expanded in the main text, and the definitions appear only in Appendix E.2. Please define the labels at first use or add a pointer to the appendix. Second, the Re-cap task that provides the test data for Task 4 is described only as a standardized IQ assessment. Was this a re-administration of the Raven test used for baseline matching, or a different assessment?
4 comments
AuthorsOpenReview ↗

Response to Reviewer MYrA (Part 1/3: Comments 1–2)

Thank you for recognizing the value of the Re-cap setting and the characterized deaf/hearing cohorts. Your questions helped us clarify the acquisition timeline and evaluation protocols. We have also completed matched-condition control experiments with ten additional participants. We present the evidence and experimental details below.

1. Were the original and Re-cap recordings acquired on the same day, and what was the interval between them?

The original Re-cap recording was acquired three days after the initial visit. Reading, Interference, and Memory Retrieval were recorded during the initial visit; IQ/Re-cap was recorded after complete cap removal and reapplication at the second visit. Appendix D states this interval. We have specified this three-day repeated-acquisition protocol consistently in Sections 2.2–2.3, the abstract, and Table 1.

The supplementary resting-state and N-back data were collected between November 6 and 11, 2026, from ten additional participants, each completing two recordings separated by 12 hours. This supplementary cohort and timeline are reported separately from the original 45-participant acquisition.

2. How can task changes and recording changes be distinguished in Task 4, and how many Re-cap segments are available per participant?

The new controls show that substantial degradation persists when the cognitive condition is held constant across recordings. Tables 2 and 3 establish within-session Reading and Memory Retrieval performance, and Table 4 evaluates Reading-to-Memory transfer within the initial visit. Original Table 5 changes both task and recording visit. We therefore added matched-condition experiments to address the task confound directly.

Section 4.4 and Table 6 report resting-state and matched 0-back, 1-back, and 2-back evaluations in the supplementary cohort. Models are trained on the first recording and tested on the same condition in the second recording. FBCNet achieves resting-state accuracy of 0.7966 ± 0.0583; EEGNet achieves the highest N-back accuracies of 0.6823 ± 0.0167, 0.7077 ± 0.0305, and 0.7155 ± 0.0241, respectively.

Table 9 further separates load and recording changes using the same participants' 0-back and 2-back data. All eight models are trained only on first-recording 0-back and tested on four conditions:

  • First-recording 0-back: neither load nor recording changes.
  • First-recording 2-back: load changes.
  • Second-recording 0-back: recording changes while load is retained.
  • Second-recording 2-back: both change.

Within each of five blocked outer folds, the four conditions share the same selected checkpoint, identities, and corresponding held-out segment-index blocks. Across all eight models, changing load within the first recording reduces accuracy by 0.00–6.42 percentage points, whereas changing the recording at the same load reduces it by 29.42–63.33 percentage points. This directly addresses the task-confounding concern and supports recording-associated changes as the larger source of generalization loss under this protocol. These controls use resting state and N-back in an additional cohort; the comparison does not separately estimate the contributions of cap reapplication, elapsed time, and physiological variation. We have aligned the manuscript's interpretation with this evidence.

Segment counts in the original IQ/Re-cap target recording are as follows:

CohortParticipantsSegments per participantTotal target segments
Prelingually deaf22641,408
Hearing23641,472

Segment counts in the supplementary recordings are as follows:

ConditionParticipantsSegments per participant per recordingSegments per recordingTotal across both recordings
Resting state1063, 61, 63, 65, 65, 67, 66, 63, 65, 676451,290
0-back10120 each1,2002,400
1-back10120 each1,2002,400
2-back10120 each1,2002,400

All supplementary segments are 2 seconds long at 400 Hz, giving 1,290 resting-state segments and 7,200 N-back segments: 8,490 unique segments in total. Table 9 uses a subset of these same data.

AuthorsOpenReview ↗

Response to Reviewer MYrA (Part 2/3: Comment 3-4)

3. What is cross-validated in Table 5, and what are the training, tuning, and EEGPT pretraining procedures?

Table 5 cross-validates source-side training and validation. Five-fold stratified splitting is performed within Memory Retrieval. Normalization is fitted on each source training partition, and checkpoint selection uses its source validation partition. IQ/Re-cap is held out from training, normalization fitting, checkpoint selection, and hyperparameter tuning.

The five selected models are evaluated on the same complete IQ/Re-cap target recording. The mean and sample standard deviation therefore describe variation across five source-trained models on a fixed target set, rather than five independent target test folds. Table 4 follows the same procedure with Reading as source and Memory Retrieval as target. We have clarified this distinction in both captions. Here, zero-shot transfer denotes closed-set identity recognition without target-data adaptation: source and target contain the same enrolled identities.

The supplementary protocols are:

  • Table 6: five training/validation splits within the first recording; each selected model is evaluated on the same complete second recording.
  • Table 9: five blocked outer test folds, each with 720 source training segments, 240 separate source validation segments, and 240 test segments per condition. Each segment receives one outer-fold test prediction; test blocks and checkpoint-selection validation blocks are disjoint.

Training details. The local baseline implementations use AdamW and cross-entropy with label smoothing of 0.1. The learning rate is 0.001, except for EEGViT (0.0003) and the local EEGPT-named implementation (0.0005). In the original identity experiments, weight decay is 0.001 for LRMD, FBCNet, EEGNet, EEGConformer, and ATCNet, and 0.01 for EEGViT, the local EEGPT variant, and BrainprintNet. Batch size is 32, except for BrainprintNet (64). Original epoch budgets vary by architecture and task and are reported in Appendix E.3, Table 14.

For the supplementary resting-state and N-back experiments, weight decay is 0.01, the gradient-norm limit is 5, and batch size is 32, except for BrainprintNet (8). Training runs for at most 80 epochs, with a minimum of 20 epochs and early-stopping patience of 15 epochs. We retain the checkpoint with the lowest source-validation cross-entropy and use one fixed configuration per architecture, with no target-based hyperparameter tuning.

Table 6 uses split seed 42 and training seeds 4201–4205; Table 9 uses blocked partitions and the same training seeds. Partitions and selection rules are fixed before target evaluation.

EEGPT pretrained weights. The original EEGPT-named baseline is a local variant trained from scratch. To evaluate the public pretrained model, we added Table 11, using the publicly released EEGPT encoder with frozen parameters and a linear readout trained on source data. This evaluation uses AdamW, learning rate 0.0005, weight decay 0.01, batch size 32, and source-validation checkpoint selection. Target recordings remain held out under the source-only protocol described above.

The training paragraph preceding Section 4.1 and Appendix E.3, including Tables 13–14, document the architecture-specific hyperparameters, preprocessing, model selection, and pretrained-weight usage to support reproducibility.

4. Is population classification participant-independent, and why are there 5,632 rather than 5,760 connectivity matrices?

Yes. The five folds are participant-disjoint. For population classification in the original Table 9, now Table 17, all connectivity matrices from one participant belong to the same fold. Fitted preprocessing and hyperparameter selection use training participants only. We have stated this grouping rule and reported each fold's participant composition and matrix counts.

The full dataset contains 45 participants: 22 prelingually deaf and 23 hearing participants. One participant was excluded only from this population-classification analysis because of poor task-response quality, leaving 44 participants. All other benchmark tasks retain the original cohort.

The correct counting unit is eight condition-stage combinations: four sentence conditions (CN, IN, CT, and IT) across Reading and Memory Retrieval. With 16 trials per condition per stage, each participant contributes 128 matrices:

44 × 4 conditions × 2 stages × 16 trials = 5,632 matrices, comprising 2,816 Reading matrices and 2,816 Memory Retrieval matrices.

Thus, 5,632 matrices correspond to the 44 analyzed participants; 5,760 would correspond to the full 45-participant cohort. We have updated the methods and caption with the analysis-specific exclusion and the condition-stage counting scheme.

AuthorsOpenReview ↗

Response to Reviewer MYrA (Part 3/3: Comments 5)

5. What do CN, IN, CT, and IT mean, and which IQ assessment was used during Re-cap?

The four labels represent the 2 × 2 factorial combination of semantic correctness and sentence type:

  • CN — Correct Normal: a semantically correct normal sentence.
  • IN — Incorrect Normal: a semantically incorrect normal sentence.
  • CT — Correct Tongue Twister: a semantically correct tongue-twister sentence.
  • IT — Incorrect Tongue Twister: a semantically incorrect tongue-twister sentence.

The revised Section 2.2 defines these abbreviations at their first occurrence.

The original IQ/Re-cap task was a repeat administration of the same Raven assessment used for baseline characterization, conducted three days after the initial visit. We have identified the assessment explicitly in the acquisition protocol.

AuthorsOpenReview ↗

Follow-up on Additional Benchmark Evidence (Tables 18–20)

Dear Reviewer MYrA,

Thank you again for your constructive evaluation and recognition of MCD-EEG's contributions.

Following our previous responses, we have completed three additional evaluations (Tables 18–20), examining language-conditioned cross-visit identity transfer, sentence-condition generalization within the same recording, and matched-randomization controls. These results further strengthen the benchmark's scientific informativeness and demonstrate its distinctive value for evaluating EEG representation robustness across hearing and prelingually deaf cohorts.

As the discussion period approaches its conclusion, we would sincerely appreciate any further feedback on these new findings. We would also be grateful if you could take this additional evidence into account when finalizing your assessment.

Thank you very much for your time and thoughtful consideration.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of the Rebuttal

Summary of the Rebuttal

We sincerely thank Reviewers MYrA, e4hX, and ijX7 for their careful reviews and constructive suggestions. We appreciate their recognition of the characterized prelingually deaf/hearing cohorts and the value\color{green}{\text{the characterized prelingually deaf/hearing cohorts and the value}} of evaluating EEG robustness across cognitive tasks and recording\color{green}{\text{of evaluating EEG robustness across cognitive tasks and recording}} conditions\color{green}{\text{conditions}}. Their feedback has helped us strengthen the experimental evidence and clarify the benchmark's contribution.

We have responded point by point and completed additional experiments. The main responses are summarized below:

  • Reviewer MYrA: (1) Acquisition timeline:\color{blue}{\text{(1) Acquisition timeline:}} We clarify that the original IQ/Re-cap recording occurred three days after the initial visit; the supplementary ten-participant recordings were 12 hours apart. (2) Task confounding:\color{blue}{\text{(2) Task confounding:}} We added same-condition resting-state/N-back tests (Table 6) and an eight-model load × recording comparison (Table 9). Same-recording load changes reduce accuracy by 0.00–6.42\color{green}{\text{Same-recording load changes reduce accuracy by 0.00–6.42}} percentage points, compared with 29.42–63.33 points for\color{green}{\text{percentage points, compared with 29.42–63.33 points for}} matched-load recording changes.\color{green}{\text{matched-load recording changes.}} (3) Evaluation and training:\color{blue}{\text{(3) Evaluation and training:}} We clarify source-only model selection, the interpretation of five-fold results, hyperparameters, and the public pretrained EEGPT evaluation. (4) Sample accounting and annotations:\color{blue}{\text{(4) Sample accounting and annotations:}} We explain participant-disjoint population-classification folds, the analysis-specific exclusion yielding 5,632 matrices, condition labels, and the repeated Raven assessment.

  • Reviewer e4hX: (1) Contribution beyond existing resources:\color{blue}{\text{(1) Contribution beyond existing resources:}} We identify the shared protocol linking characterized deaf/hearing cohorts, cognitive stages, behavioral annotations, and repeated recordings as the basis for subgroup robustness comparisons. (2) Cognitive decoding and task labels:\color{blue}{\text{(2) Cognitive decoding and task labels:}} Table 8 evaluates Reading-versus-Memory Retrieval decoding with participant-disjoint five-fold testing. The best accuracies are 66.47% ± 2.82% for hearing participants\color{green}{\text{The best accuracies are 66.47\% ± 2.82\% for hearing participants}} and 66.54% ± 4.72% for deaf participants, against 50% chance.\color{green}{\text{and 66.54\% ± 4.72\% for deaf participants, against 50\% chance.}} We also detail CN/IN/CT/IT labels and response accuracy/times. (3) Benchmark utility despite saturation:\color{blue}{\text{(3) Benchmark utility despite saturation:}} Table 9 distinguishes models with near-perfect within-recording results: matched-load second-recording accuracy is 70.58% for FBCNet, 68.08% for EEGNet, and 36.42% for EEGConformer.

  • Reviewer ijX7: (1) Same-condition controls:\color{blue}{\text{(1) Same-condition controls:}} The completed ten-participant experiments show substantial degradation with cognitive condition held constant, supporting a recording-associated effect without isolating cap reapplication from elapsed time and physiological variation. (2) Experimental consistency:\color{blue}{\text{(2) Experimental consistency:}} We distinguish 4-second original identity inputs, 2-second supplementary inputs, and 1,500-ms poststimulus ERP analysis with a 200-ms prestimulus baseline, correcting our interim response. (3) Leakage and reproducibility:\color{blue}{\text{(3) Leakage and reproducibility:}} We specify complete-epoch partitioning, training-only preprocessing fitting, and source-validation checkpoint selection. Hash checks found no exact duplicate epochs among 5,760\color{green}{\text{Hash checks found no exact duplicate epochs among 5,760}} Reading/Memory epochs or 4,800 supplementary N-back segments; all\color{green}{\text{Reading/Memory epochs or 4,800 supplementary N-back segments; all}} 160 Table 9 prediction files were independently rescored.\color{green}{\text{160 Table 9 prediction files were independently rescored.}} (4) Contribution:\color{blue}{\text{(4) Contribution:}} We substantiate the resource's value through subgroup robustness, cognitive-stage decoding, and controlled recording-transfer evaluation.

We thank the reviewers again for helping improve the paper. The additional experiments provide direct evidence for\color{green}{\text{The additional experiments provide direct evidence for}} matched-condition recording transfer and a non-saturated\color{green}{\text{matched-condition recording transfer and a non-saturated}} cognitive-stage decoding task.\color{green}{\text{cognitive-stage decoding task.}} We respectfully ask the reviewers to consider this additional evidence in reassessing our submission.

The revised manuscript has been uploaded, and the complete\color{green}{\text{The revised manuscript has been uploaded, and the complete}} anonymous dataset, including the newly collected data, is\color{green}{\text{anonymous dataset, including the newly collected data, is}} available at:\color{green}{\text{available at:}} https://huggingface.co/datasets/anonymous-588/MCD-EEG/tree/main.

Best,
The Authors

AuthorsOpenReview ↗

Request for Assistance with Reviewer Follow-up

Dear ICLR 2027 AC, SAC, and PC,

We sincerely thank all three reviewers for their valuable feedback and constructive suggestions. Their comments have greatly helped us improve our manuscript and strengthen the scientific contribution of our benchmark.

Throughout the discussion period, we have carefully addressed the reviewers' concerns, conducted substantial additional experiments, and provided further evidence in Tables 18–20 to clarify the benchmark's distinctive value and scientific significance. In particular, we promptly recruited ten additional participants and conducted new controlled EEG experiments to directly address the methodological concerns raised during the review.

As the discussion period approaches its conclusion, we would be very grateful for your kind assistance in facilitating any further reviewer feedback on our latest responses and experimental findings, if appropriate. We sincerely welcome any additional suggestions and fully respect the reviewers' academic judgments.

Thank you very much for your time, dedication, and continued support throughout the review process.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.