List of Papers Browse by Subject Areas Author List
Abstract
Intracortical Brain–Computer Interfaces (iBCIs) for speech decoding show immense promise in translating neural activity into text. However, clinical viability remains severely constrained by labeled data scarcity and the continuous physical drift of neural representations over time. Contemporary approaches increasingly rely on massive cross-species pretraining and large ensembles, incurring prohibitive costs while still requiring daily recalibration to counteract electrode drift. We propose a data-efficient pipeline that requires only 40 hrs of patient data and achieves domain generalization to unseen recording sessions. To overcome data limitations, we propose Neural Cut-and-Stitch (NCS), a signal-space augmentation strategy. By isolating word-level neural patterns from past recordings, we apply crossfade stitching to synthesize brain signals for entirely new sentences. We also introduce an Adversarial Domain Adaptation (ADAN) framework to improve generalization to unseen sessions. This explicitly enforces a session-invariant neural manifold, retaining acoustic decoding performance with far lesser calibration. Experiments on the Brain-to-Text ’25 benchmark demonstrate that NCS alone reduces the Phoneme Error Rate (PER) from 10.1% to 8.8% on the held-out test set. Combining it with the baseline, we reduce the Word Error Rate (WER) from 7.0% to 4.8%. Concurrently, ADAN confines PER degradation on unseen sessions ∼ 15 days into the future to <2% change. These gains show that signal augmentation and domain adaptation can be complementary to large-scale pretraining and reduce reliance on daily recalibration for deployment.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5367_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Miccai-26/brain-to-text-ncs-adan
Link to the Dataset(s)
Brain To Text’25 Dataset: https://datadryad.org/dataset/doi:10.5061/dryad.dncjsxm85
BibTex
@InProceedings{NasAni_Cut_MICCAI2026,
author = { Naskar, Animan AND Naskar, Ankush AND Bathula, Deepti},
title = { { Cut, Stitch, Adapt: Data-Efficient and Low-Calibration Speech BCIs } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16894},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The study proposes two approaches to improve speech decoding for intracortical BCIs. The Neural Cut-and-Stitch data augmentation segments neural data by word and realigns the segments to generate new, augmented sentences. The study also applies adversarial domain adaptation framework, which is designed to create a day-agnostic encoder.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The study tackles two critical problems for intracortical BCIs, which include requirement of considerable effort for data collection, and managing drifts in neural signals over time. The Neural Cut-and-Stitch data augmentation approach includes detailed approaches considering the characteristics of neural signals, such as adding black and space regions to provide natural boundary and conducting normalization statistics to support natural sequence distribution. The methods are presented clearly with details and the manuscript is well written overall.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1) The proposed approaches are implemented using a Bidirectional GRU as the neural decoder. To better isolate and highlight the specific contributions of ADAN and NCS, it would be beneficial to test their integration with other backbone architectures besides the GRU. 2) The manuscript’s clarity and reproducibility would be enhanced by providing more details on the training process. Specifically, more details on hyperparameter tuning such as sensitivity analysis, and with more clarity on the methods for optimal epoch selection based on the validation set. 3) The purpose of the Generative Synthesis approach is unclear, as it does not appear to improve decoding performance based on the current results. While the motivation is reasonable, the manuscript should more clearly emphasize its effect or provide advantages for using or including this method.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
As I understand it, the generative synthesis is also trained on neural recordings to produce synthetic signals. If this is correct, shouldn’t the ‘Total Training Data’ for the ‘Ours (Generative)’ model in Table 1 account for the original recordings used for its training? Clarification on this point would be appreciated.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
While data augmentation via segmentation and realignment, and adversarial domain adaptation for cross-session (across time) decoding are not entirely new concepts in neural signal processing, their application to intracortical BCIs for speech decoding is a valuable contribution. The study’s novelty could be further highlighted by demonstrating that the benefits of NCS and ADAN generalize across multiple decoding architectures, as this appears to be the core value of this study.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The study has some limitations, including the limited participant scope arising from the use if microelectrode recordings. The motivation for using the generative approach is clear, but the approach itself to me does not appear to contribute meaningfully to the study’s novelty. Despite these concerns, other concerns have been addressed and the study has value in the field of microelectrode BCIs as a demonstration of how data augmentation and domain adaptation techniques can be applied in this setting.
Review #2
- Please describe the contribution of the paper
This paper proposes two key methods for limited labeled data and longitudinal neural drift and the authors applies them within Speech BCI system. The work focuses on improving data efficiency and robustness through a signal-space data augmentation method and an ADAN framework to improve generalization to unseen future recordings.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.Clear identification of two key bottlenecks in speech BCI systems, namely limited labeled data and longitudinal neural drift. 2.Intuitive and practically motivated methodological design, including a signal-space data augmentation strategy and an adversarial domain adaptation framework. 3.A well-articulated system-level perspective with meaningful empirical improvements in decoding performance and robustness.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The overall methodological novelty is moderate. The contribution is primarily an effective system-level integration rather than a fundamentally new method. 2.The evaluation is limited to a single participant 3.Proposed Neural Cut-and-Stitch approach dose not explicitly preserve preserve phase continuity or fine-grained temporal dynamics. This may limit its ability to accurately capture neural signal structure.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper addresses important and practical challenges in intracortical speech BCIs, namely data scarcity and longitudinal signal drift. The proposed methods are intuitive and demonstrate consistent improvements in both decoding performance and robustness to unseen sessions. While the methodological novelty is moderate and largely builds upon existing techniques, the work provides a meaningful system-level contribution under realistic constraints. Some limitations remain, including the lack of analysis on temporal continuity (e.g., phase structure), as well as limited evaluation across subjects.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
All my concerns have been addressed in the manuscript and the response letter.
Review #3
- Please describe the contribution of the paper
The paper presents a Novel Stitch and Cut Algorithm for data augmentation of sentence level intracortical EEG data by extracting word/phenome level representations from the existing data and joining these extracted representations to augment new data in order to cater to the problem of limited recording data. The authors also propose Adversarial Domain Adaptation (ADAN) framework to achieve generalization across different session recordings, in order to reduce need for session/day level calibration.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The paper has the following strengths:
-
The paper tackles an important and pressing problem(lack of enough data) and session specific calibration which are key challenges in development of accurate and clinically deployable BCIs for neuroprosthesis
-
The proposed method seems model, framework and hardware agnostic, thus making it easily integrable in any existing ECoG/EEG based neuroprosthesis pipelines
-
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The paper has following major weaknesses:
- The paper lies on a core assumption that the individual words when rearranged to form another sentence won’t change the underlying distribution of the signal(while it maybe true for audio signals) but the EEG/ECoG signals are complex in the regard because 2 different framings from same set of words could trigger very different emotions/artifacts in brain which might be missing in the synthetic data for example
“The company hired a new CEO” “The CEO hired a new company”
the brain might generate slightly different signals for each, so a good way will be to see if the underlying distribution of a synthetically generated sentence signal aligns with its true counterpart(which exists in dataset) to support this key assumption of the paper
- The evaluation is done on a single dataset with a single participant which makes it highly limited to establish the credibility and the generalizability of the proposed techniques. The authors shall clearly justify the intuition behind the same, why other publicly available datasets for the same task were not used:
[1] Willett, Francis; Kunz, Erin; Fan, Chaofei et al. (2023). Data for: A high-performance speech neuroprosthesis [Dataset]. Dryad. https://doi.org/10.5061/dryad.x69p8czpq
[2] Berezutskaya, J., Vansteensel, M.J., Aarnoutse, E.J. et al. Open multimodal iEEG-fMRI dataset from naturalistic stimulation with a short audiovisual film. Sci Data 9, 91 (2022). https://doi.org/10.1038/s41597-022-01173-0
[3] Zaid Zada, Samuel A. Nastase, Bobbi Aubrey, Itamar Jalon, Ariel Goldstein, Sebastian Michelmann, Haocheng Wang, Liat Hasenfratz, Werner Doyle, Daniel Friedman, Patricia Dugan, Lucia Melloni, Sasha Devore, Orrin Devinsky, Adeen Flinker, and Uri Hasson (2025). The “Podcast” ECoG dataset. OpenNeuro. [Dataset] doi: doi:10.18112/openneuro.ds005574.v1.0.2
-
The authors have not clearly mentioned whether there are any other existing techniques for data augmentation or if this paper presents a first attempt for the same authors shall mention it explicitly
-
The authors have compared limited models in their evaluation, which limits the credibility of the results
“Although optimizing G via Soft-DTW and CTC consistency losses successfully produced phonetically valid macroscopic signals, the generator introduced subtle, high-frequency spectral artifacts. The primary acoustic decoder subsequently overfit to these artifacts during training. Consequently, this generative branch was omitted from the final evaluation pipeline in favor of the more stable NCS method.” The authors shall trying backing the same with data
-
In Figure 3, what is role of the GRL block in forward pass, if it is more like a skip connection to encoder it could be represented separately for more clarity. In Figure 4, the legends overlap with the the graph, authors shall consider plotting the legends outside for better clarity
-
The authors have used OPT 6.7B, is there a specific reason for choosing the same it is a bit older model(2022) and in recent days models like Qwen 3.5, Gemma etc give way better performance even under 4B param budget, so the authors shall justify the intuition behind the choice
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
There are some arbitrary choices, assumptions and claims that have not been backed by enough data to establish the reliability of the framework. Kindly see the weaknesses section and try providing justification for these choices in rebuttal.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors have made a sincere effort in the rebuttal to address the concerns raised earlier. Most of my questions have been adequately answered, and for the points that could not be fully addressed, the authors have appropriately acknowledged them as limitations and outside the current scope of the work.
However, I would still like to highlight one concern regarding the Dyrad dataset. The dataset corresponds to a micro-electrode recording setting, particularly involving recordings from the speech-motor cortex (Broca’s area), and the rebuttal did not provide sufficient clarification on how this setting is comparable to the other ECoG-based datasets discussed in the paper. Instead, it was grouped together with the other datasets without adequately addressing the differences in recording modality and neural acquisition setup. While the justification provided for the ECoG datasets appears reasonable overall, the distinction for the Dyrad dataset remains somewhat unclear.
Overall, I believe the paper would foster meaningful discussion at the MICCAI 2026 conference and could motivate future research in data augmentation for intracortical BCIs, where data collection is inherently scarce, expensive, and difficult to obtain, while also opening possibilities for extending such approaches to other neural and brain-signal modalities as well.
Author Feedback
We sincerely thank all reviewers for their constructive feedback. We address the major concerns below.
(New experiments not included as per MICCAI guidelines)
[R1, R2, R4] Single-Participant Scope: We acknowledge this limitation; however, microelectrode recordings of the speech motor cortex are rare and small-scale. The alternative datasets suggested use ECoG/fMRI/EEG, which operate on fundamentally different spatial and temporal scales than microelectrode arrays, making direct comparison out of scope. Both methods are biologically motivated to generalize across patients:
- NCS: Electrodes are implanted specifically in speech motor cortex, so recorded signals primarily reflect articulatory movements rather than higher-level semantic content — making word-level signatures structurally stable and reusable across rearrangements for any patient.
- ADAN: Drift from micro-motions and glial scarring is a universal constraint of intracortical BCIs. ADAN models this as a statistical domain shift, adversarially enforcing session-invariant representations regardless of individual neural patterns.
[R2, R4] NCS Signal Dynamics & Continuity: Cuts occur strictly within long BLANK/SPACE regions where co-articulatory coupling is minimal. Signal reconstruction applies mean/variance alignment and spectral crossfading at boundaries to suppress discontinuities. Additionally, 2,000 randomized word sequences regularize against the model exploiting linguistic structure in place of neural dynamics. The 10.1%→8.8% PER reduction empirically confirms the net benefit.
[R4] Prior Augmentation Literature: Existing microelectrode array augmentation methods primarily use time shifting/warping, noise injection and channel/time masking; to our knowledge, none improved PER/WER over baseline metrics. We acknowledge the absence of a formal comparative study as a limitation.
[R1, R4] Generative Branch: Motivation & Interpretation: The generator was trained on the original 40-hour dataset using Soft-DTW and CTC-consistency losses (Table 1 will be updated). It was designed to model co-articulatory context by inverting sentence-level phoneme sequences back into continuous neural signals — an explicit attempt to go beyond NCS’s static word-segment assumption. We retained it to transparently document this exploration. Its inferior performance stems from compounding failure modes: the inversion starts from soft-logits carrying 10.1% inherent decoder error; the CTC training signal depends on this same imperfect decoder; and Soft-DTW optimizes a soft minimum over alignments rather than a single ground-truth path, introducing path-averaging artifacts. Together, these produce spectral artifacts that explains the large performance gap.
[R4] GRL Clarification: The GRL is not a skip connection — it sits sequentially between the statistical pooling layer and the day classifier. During backpropagation, gradients pass through the pooling layer to the encoder with their sign negated by the GRL, suppressing day-specific information while preserving phonetic structure. We will clarify this in Figure 3. [R1, R4] Architecture & OPT-6.7B During development we evaluated biGRU, biLSTM, and standard RNN architectures; the GRU consistently outperformed alternatives. It also enables a direct 1-to-1 comparison with the official benchmark baseline, ensuring reported WER gains reflect solely our neural decoding contributions. Both NCS and ADAN are architecture-agnostic; multi-backbone evaluation is flagged as future work.
[R1] Training Protocol Models were trained for a fixed number of epochs on merged train+validation data; the optimal epoch count was determined on per-session validation splits in a preliminary run. The held-out test session was never used for tuning. We will make this explicit in the final manuscript.
[R4] Figure Clarity: Legend overlap in Figure 4 will be corrected.
Code will be made public upon acceptance.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
Reviewers gave many positive comments on this paper, including (1) it attempts to tackle two critical problems for intracortical BCIs, (2) the methodological design is reasonable and practical, and (3) it is well written.
However, they also raise some major concerns. (1) A core assumption of this study may greatly affect its performance. (2) The evaluation is limited to a single participant. (3) The proposed approach does not explicitly preserve phase continuity or fine-grained temporal dynamics.
Besides, reviewers also consider the comparison experiments are not sufficient and some implementation and experimental details can be further clarified.
Please carefully study the comments raised by the reviewers and provide responses to their concerns in the rebuttal.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
After rebuttal, all reviewers suggest accepting this paper.
It is suggested that the authors sufficiently consider the comments raised by the reviewers and further improve the paper in the final version.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The reviewers unanimously agree this work deserves acceptance and presentation at the conference.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
All reviewers agree that the authors have address the major concerns about the manuscript. The small sample size is confirmed by the authors to be a limitation of the current publication I encourage the authors to ensure this point is included in the final manuscript.