Plug-and-Play Stability for Intracortical Brain-Computer Interfaces: A One-Year Demonstration of Seamless Brain-to-Text Communication

Intracortical brain-computer interfaces (iBCIs) have shown promise for restoring rapid communication to people with neurological disorders such as amyotrophic lateral sclerosis (ALS). However, to maintain high performance over time, iBCIs typically need frequent recalibration to combat changes in the neural recordings that accrue over days. This requires iBCI users to stop using the iBCI and engage in supervised data collection, making the iBCI system hard to use. In this paper, we propose a method that enables self-recalibration of communication iBCIs without interrupting the user. Our method leverages large language models (LMs) to automatically correct errors in iBCI outputs. The self-recalibration process uses these corrected outputs ("pseudo-labels") to continually update the iBCI decoder online. Over a period of more than one year (403 days), we evaluated our Continual Online Recalibration with Pseudo-labels (CORP) framework with one clinical trial participant. CORP achieved a stable decoding accuracy of 93.84% in an online handwriting iBCI task, significantly outperforming other baseline methods. Notably, this is the longest-running iBCI stability demonstration involving a human participant. Our results provide the first evidence for long-term stabilization of a plug-and-play, high-performance communication iBCI, addressing a major barrier for the clinical translation of iBCIs.

Paper

Similar papers

Peer review

Reviewer QkLb7/10 · confidence 3/52023-07-06

Summary

This paper introduces a method called Continual Online Recalibration with Pseudo-labels (CORP) that allows for self-recalibration of intracortical brain-computer interfaces (iBCIs) without the need for user interruption. iBCIs are used to restore communication abilities in individuals with neurological disorders like ALS but require frequent recalibration due to changes in neural recordings over time. The proposed method utilizes large language models (LMs) to automatically correct errors in iBCI outputs and uses these corrected outputs as "pseudo-labels" to update the iBCI decoder in real time. The CORP framework was evaluated over an 8-month period with a single participant in a clinical trial. The results showed a stable decoding accuracy of 93.71% in an online handwriting task, outperforming other baseline methods. This study demonstrates the longest-running iBCI stability with a human participant and presents a plug-and-play, high-performance communication iBCI that addresses a significant challenge in the clinical application of iBCIs.

Strengths

1. originality: This work presents an online recalibration method for iBCI which uses LLM to perform re-calibration for communication iBCI. 2. quality and clarity: This paper provides detailed experimental validation of the effectiveness of the proposed method. The figures and text are clear, easy to understand, and flow smoothly, ensuring a clear expression of ideas. 3. significance: The CORP method proposed in this paper achieves higher accuracy and stability in brain-to-text communication through intracortical brain-computer interfaces (iBCIs) than other recalibration methods. It has been validated on the longest-running iBCI involving human participants, proving its effectiveness and potential.

Weaknesses

1. The method presented in this paper has only been validated on a single subject, which may raise concerns about the limited sample size and lack of generalizability. 2. The stopping criteria mentioned in the article are based on whether the loss falls below a certain threshold to determine the termination of training. This criterion may not be sufficiently objective. It could be considered to explore alternative quantitative metrics for evaluating the training performance of the model.

Questions

If a participant does not use the iBCI for an extended period without recalibrating the RNN model, the initial Word Error Rate (WER) and Character Error Rate (CER) can be very high. In such cases, using pseudo-labels generated with the help of LLMs may potentially mislead the training of the RNN model. Has this scenario been tested or are there any alternative approaches?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

As mentioned by the authors in the paper, the main limitation of CORP lies in the pseudo-labels generated by the language model. If the error rate of the pseudo-labels is high, it directly affects the Character Error Rate (CER) of the recalibrated RNN model. Therefore, it is crucial to select more powerful language models to minimize errors in the pseudo-labels.

Reviewer hfNe8/10 · confidence 4/52023-07-07

Summary

The paper proposes a self-recalibrating brain-computer interface system where inputs come from implanted electrodes in patient's motor cortex and the outputs characters that are patient images to hand-write. The output is then corrected by a language model to match to the most probable desired sentence. The recalibration in needed because in the real biological system there are many factors that change over time and the input data distribution drifts because of that. While it is possible to recalibrate the system by running a dedicated session, this is not optimal is it requires downtime, participation of a technician and ~daily mental effort from the patient. This work proposes a smart trick to used LLM-corrected word predictions as new labels for continuous retraining: the input signal distribution drift is gradual on the next day, while the drift has already occurred, the language model can still correctly recover what the intended word were. And these corrected words can be used as new labels to retrain the system a bit and this cancel the drift. If performed often the system will never drift too far and will always be able to self-correct. The experimental results from one test subject with implanted microelectrode array confirm that the error rate of the system stays stable thanks to continuous recalibration and it is shows that it would degrade without the proposed mechanism.

Strengths

The paper is very clearly written and well-structured, to the extent that I find myself in trouble performing my role as a reviewer's and ask questions, because I had to delete most most of them as I progressed through the pages :) The application is, of course, amazing and the fact that this work was tested on a human subject with implanted electrodes makes it a unique contribution to the field of ML applications to BCIs.

Weaknesses

I could not come up with any. I think paper is a very clear account of what was done and presents sufficient support for the claims. Perhaps more test subjects would be beneficial, but even with one subject the results stand.

Questions

Fig 1: What does the arrow from "Language model" to "Current session's data" mean? Is it somehow part of LM's training to predict session data? Or does it provide labels and the data in "Tx Fetures" is the same as in "Current session's data"? 126: How does the system know that the subject has finished thinking a sentence? Is there a special "stop"-though and if so - how reliably it can be detected? Q1: What was X's subjective feedback on the performance of the system after a few months? While the numbers clearly show that the system was more accurate, what it also noticeable from the user's perspective? Q2: How are spaces handles? I would image there is no clear motor imagery to think about to imagine making a space, how does the patient do that? 361: Perhaps this should go to Limitations?

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

4 excellent

Limitations

I would be curious to know about the natural limits of speed of writing with this approach. After all having a direct brain connection would seem like an opportunity to forgo "clunky" written language and make a more direct transmission possible. Just curious what are the authors' thought on feasibility of this and what could be the way to do it.

Reviewer 8J1g5/10 · confidence 3/52023-07-07

Summary

There is a vastly growing literature reconstructing continuous language from non-invasive brain recordings using popular deep-learning models. These papers typically use invasive recordings from surgically implanted electrodes, while decoders that use non-invasive recordings can only identify stimuli from among a small set of letters, words, or phrases. This paper contributes to that literature by introducing a new approach, CORP, a continual online recalibration method for intracranial Brain-Computer Interface (iBCI) that maps neural activity to text. Specifically, the authors use a language model (RNN) to automatically correct errors in iBCI outputs by continually updating the iBCI decoder with an online learning method. The experimental results revealed that the proposed framework achieved a stable decoding accuracy of 93.71% in an online handwriting iBCI task, significantly outperforming other baseline methods.

Strengths

The paper contains the following key contributions: * The novelty of this work: Different from previous continuous language reconstruction methods from non-invasive brain recordings, the authors build a framework to leverage the structure in language to enable self-recalibration of communication iBCIs without interrupting the user. * The proposed framework is agnostic to the communication task and can recalibrate any iBCI decoder that maps input neural signals to text output. Originality: * The prototype of a self-recalibrating handwriting iBCI system and assessed its performance over a period of time is interesting.

Weaknesses

Weaknesses: * Although the paper's main idea is quite interesting, however, the experiment was conducted on one participant in a pilot clinical trial. Hence, the proposed method needs to be tested on more subjects. * Authors have not discussed recent brain decoding works: [1] Semantic reconstruction of continuous language from non-invasive brain recordings, Jerry Tang, Amanda LeBel, Shailee Jain, Alexander G. Huth [2] Decoding speech from non-invasive brain recordings, Alexandre Défossez, Charlotte Caucheteux, Jérémy Rapin, Ori Kabeli, and Jean-Rémi King * In the above works, the authors reconstruct the continuous language from non-invasive brain recordings (fMRI and MEG) with better perplexity. * Did the authors compare their methods on any existing datasets? * Since the RNN language model is limited in handling long-term memory information and vanishing gradient problems, did the authors try with recent pretrained Transformer language models? We may expect better accuracy since the Transformer model was pretrained on larger corpora. Moreover, we may expect that a replay buffer is not required in the case of Transformer models. Quality: The paper supports its claims with few details. Specifically, the methodology and experimental details are limited. Clarity: The paper needs a lot of improvement with clear methodology details and motivation behind using the RNN model. The information provided in the submission needs to be more comprehensive to reproduce the results.

Questions

* Did the authors interpret the states of RNN in a continual online learning setup? Is there any trend across days? Which information is overwriting more in the language model?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

3 good

Contribution

3 good

Limitations

The authors have discussed several limitations and made future directions for the research community.

Reviewer 3txU7/10 · confidence 4/52023-07-12

Summary

This work focuses on

Strengths

Originality: Related work in Chen et al., IEEE SMC 2022 uses a language model for pseudo label corrections during BCI self-recalibration, though the study is focused on simulations with EEG data from a longitudinal study with participants with ALS using the P300 speller. This work is a creative combination of existing ideas to develop a new method to recalibrate BCIs for communication by using language models to improve pseudolabel quality and enhancing continuous learning during recalibration via the use of a replay buffer and data augmentation. Quality: This paper presents results from a longitudinal online BCI study to demonstrate utility of the proposed approach, which is the gold standard in evaluating BCI algorithms. The inclusion of results from offline analysis also enhances the paper. Clarity: The paper is very well-written and organised. Areas needing clarity and suggestions to improve readability are noted below. Significance: This work is highly relevant to developing automated approaches to periodically recalibrate BCIs for communication for long-term BCI use with minimal user disruptions. The approach is applicable to general BCIs for communication. Results from a longitudinal online study with a BCI user from a target end user population increase the impact of the paper.

Weaknesses

- Low sample size. The paper presents results from one participant with generally high performance level, so difficult to assess the utility across a broad range of user performance levels. The low sample size is understandable given the challenge with conducting studies in target BCI end-user populations; in particular, this is an iBCI study, in contrast to a non-invasive BCI study. The authors recognise the limitation of the lack of generalisability of results given the low sample size. The authors include results from simulations using data from the current participant to investigate the impact of a broad range of character error rates on the recalibration performance (Figure 5). - Potential order effects due to lack of randomisation of the no recalibration block (block 2) and the recalibration blocks (blocks 3 and 4). If understood correctly, the RNN decoder is updated with the data from the current calibration block and does not rely on data from the seed model block, so the testing order could be randomised daily to mitigate order effects. - There is the confound of the recalibration blocks (blocks 3 and 4) displaying the LM-decoded outputs (“the top-scored result was displayed on the screen as the final decoded sentence.”) vs. the no recalibration blocks (block 2) displaying the RNN-decoded outputs. This difference in feedback may potentially impact the BCI user experience (mental state, motivation, etc.) and further compound order effects as the user is aware given the fixed block order.

Questions

- “The second block employed a frozen seed model, trained on a combination of data from [46] and data collected prior to this evaluation (21 sessions in total).” Are the “data collected prior to this evaluation” from the current participant? Also referred to as “newly collected data” earlier in the paper. Are these 21 sessions prior to day 0? - “updating the decoder after every sentence.” How is the end of a sentence detected? Automatically? - Equation 1: \theta_k, where k denotes a day implies that the recalibration uses all the data from that day. Is Equation 1 supposed to be \theta_{x, k}, where x refers to sentence? - What are the character error rates of the LM-decoded outputs (Figure 2b)? This is to assess whether the use of LM-based correction at word level introduces errors at the character level (vs. Figure 2a with RNN-decoder outputs). (Can be inferred based on simulations in figure 5). - Figure 2: It would be useful to include results from recalibration with the ground truth labels. Why are the amounts of data collection different on day 0 and day 105 different? If understood correctly, day 0 does not have four blocks. This needs to be specified/clarified in the text/caption. - Inconsistency: Over an approximately 8-month period, our participant used the iBCI system monthly and wrote on average 57.7 sentences per usage session.” vs. “X’s writing speed was 69.5 ± 8.6 characters per minute on average.” - What is “per-frame labeling”? - x_{i, t}, y_{i, t}: define subscript i. - Define all acronyms and variables in the captions and provide more context such that the captions standalone to understand the content of the presented information without necessarily referencing the text. Figure and table captions should be more informative to minimise confusion/misinterpreting the CER% or WER% results across figures/tables. For example, the mismatch between the average online WER % with CORP in table 1 vs. figure 2 is explained in the text and not the figure caption. Same with Figure 3. Captions should state if results are from offline vs. online analysis, specific blocks used during recalibration, etc., for clarity. - Check that the contrast between line styles is preserved when figures are in grayscale.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

4 excellent

Limitations

- More discussion is needed on the societal impact of the BCI technology. In particular, how effectively the BCI communicates a user’s intent when there is no alternative. There is the potential concern that the LM may be more dominant than the user’s intent, particularly in cases with low BCI prediction accuracy. - “we do not anticipate pseudolabel quality to be a major concern in practice. This is because future clinically viable iBCIs are expected to have a high decoding accuracy... Users are also likely to utilize the iBCI frequently, resulting in small nonstationarities most of the time. ... we believe that the pseudo-labels will have high accuracy, allowing CORP to sustain the iBCIs accuracy indefinitely.” Given the low sample sizes and no current data from “future clinically viable iBCIs”, these claims are questionable. There are issues related to recording quality with long term use of intracortical electrodes.

Reviewer miLw8/10 · confidence 3/52023-07-26

Summary

Edit: I have read the rebuttal and am satisfied with the authors' comments. I stand by my score and would like to see this paper accepted. ######################################################################## In this work, the authors tackle the issue of non-stationarity in decoding technologies specifically for handwriting recognition. They propose several methodological steps to address like the use of a replay buffer, data augmentation and most importantly, an n-gram language model. They use data from a longitudinal study spanning over 200 days where they collect data from a single participant. Their model comprises an affine transform for each day followed by a 2-layer RNN that outputs character probabilities. These probabilities are then fed into a pertained n-gram language model that then generates plausible words (with beam search). The top-n words are then fed into GPT2-XL to rerank and the best word is displayed back to the participant. The output of the LM is also fed into a calibrator that uses this data along with some percentage of past data to recalibrate the RNN. The authors show that this model achieves very good character error rate and reasonable word error rate over a model that doesn't recalibrate for stationarity and an alignment method. Through several experiments, they also show the usefulness of different model components, hyperparameter choices and the reliability of the pseudo-labels.

Strengths

Strengths: 1. Very interesting applications of ML techniques to handwriting decoding. 2. Well written and easy to follow. 3. Achieves very good performance and authors clearly show through many experiments how their method alleviates the problem of non-stationarity. 4. Good analyses on model choice with ablations and many different parameter sets.

Weaknesses

Weaknesses: No major weaknesses apart from some questions below and a few missing details.

Questions

How long does the assumption of some stationarity hold? The authors say hat prior work found only a 1.5% error rate but is this something we can evaluate with respect to the drop in calibration error to find the most optimal window? Re hyperparameter tuning in figure 4: It was not clear what the shaded region represented here- standard error across all decoded characters? Also, how stable is this across different days, ie., does the non-stationarity affect the plots substantially? There were limited details provided on the affine transform for each day. How is the affine transform for each day trained and what role does it play in the online decoding system? Clarification: Since the LM outputs word n-grams, I assume that this is segmented into characters to provide the pseudo `y's`? Minor: What `p`was used in the final model? (esp for Table 4) Similarly, what is the `n` used for the GPT2-XL ranking step? I would also encourage the authors to allude to the effect of replay buffer size and other modeling parameters on CER in the main text.

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

4 excellent

Contribution

4 excellent

Limitations

Yes, potential limitations were discussed.

Reviewer 3txU2023-08-14

Post-rebuttal

Reviewer has read and appreciates the author rebuttal. The main concerns are mostly addressed. Revising score upward. Other: Suggest including example trajectories of characters, RNN-decoded and LM-decoded outputs.

Reviewer 8J1g2023-08-15

Thanks for the rebuttal

Dear authors, Thanks for the rebuttal. By considering authors' feedback, it is clear that they have provided a comprehensive rebuttal and taken diligent care in addressing all the questions. Hence, I have decided to revise and increase my score accordingly.

Reviewer hfNe2023-08-20

Thank you for the clarifications!

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC