The Therapeutic Loop: Closed-Loop Epilepsy Systems Mirroring the Read–Write Architecture of Brain–Computer Interfaces

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The review by Alessandro Principe and coworkers summarizes the “therapeutic loop” for Drug-resistant epilepsy (DRE) and potential path forward to bridge the translation gap between research and clinical trials. Overall, the paper is clearly written and scientifically sound; however, the following points are suggested to improve clarity, completeness, and interpretive depth.

1. Line 29; The authors are suggested to define ILAE.
2. Line 77; Although the sentence groups closed-loop and open-loop CLS together, it is not demonstrated that both modalities share the same mechanistic underlying process. If there are differences in stimulation timing or pattern, their mechanistic contributions may not be equivalent.
3. Line 78; The term “progressive adaptations within the seizure network” is not clearly defined. It is unclear whether the authors refer to synaptic plasticity, network-level reorganization, neuromodulatory changes, or another mechanism. The authors are encouraged to specify in detail to substantiate the proposed interpretation.
4. Line 91-94; How should true long-term seizure freedom be defined? Is a clinical benchmark sufficient, or should a more stringent metric, such as multi-year remission or statistically demonstrated long-term stability, be required? Additionally, if relapse occurs after one year, does this commonly used criterion genuinely capture long-term seizure control, or does it highlight the need for a more robust and standardized definition?
5. Line 108-113; The authors need to support their statement.
6. Line 157-158; The authors are advised to provide citation(s) for the released data.
7. Line 159; Could the authors provide details on how these sensitivity gains were balanced against changes in other performance metrics, and how “the most challenging patient cases” were defined?
8. Line 357-364; The authors need to clarify how reliance on subject-specific workflows affects the scalability and real-world implementation of BCI systems, and whether any generalizable or cross-subject approaches were considered or evaluated in comparison.
9. Line 395; The authors are encouraged to support their statement.
10. Line 460-468; How did the authors assess whether the augmented or synthetic EEG data preserved physiologically meaningful features? What criteria were used to ensure that these augmentation techniques did not introduce misleading or non-biological patterns into the training set?

Author Response

Comment 1: Line 29; The authors are suggested to define ILAE.
Response 1: Agree. The definition is now provided in line 29.

Comment 2: Line 77; Although the sentence groups closed-loop and open-loop CLS together, it is not demonstrated that both modalities share the same mechanistic underlying process. If there are differences in stimulation timing or pattern, their mechanistic contributions may not be equivalent.
Response 2: The term CLS refers specifically to closed-loop systems, so no open-loop CLSs exist by definition. We realise that our wording may have caused confusion, and we now clarify this distinction more explicitly. Lines 59–63 and Figure 1 described the differences between open-loop stimulation (OLS) and closed-loop stimulation (CLS). In the original line 77, the expression “both modalities” referred not to OLS and CLS, but to the two mechanisms of action within CLS discussed in the preceding two sentences: suppression of epileptiform activity and direct frequency modulation. The cited study suggests that the therapeutic benefit of these two mechanisms arises from similar progressive network adaptation. To avoid ambiguity, we have revised the text and added an explicit clarification in lines 76 and 81 indicating that the subsequent discussion concerns CLS only.

Comment 3: Line 78; The term “progressive adaptations within the seizure network” is not clearly defined. It is unclear whether the authors refer to synaptic plasticity, network-level reorganization, neuromodulatory changes, or another mechanism. The authors are encouraged to specify in detail to substantiate the proposed interpretation.
Response 3: Agree, we thank the reviewer for this comment. Details are now provided in lines 81-89.

Comment 4: Line 91-94; How should true long-term seizure freedom be defined? Is a clinical benchmark sufficient, or should a more stringent metric, such as multi-year remission or statistically demonstrated long-term stability, be required? Additionally, if relapse occurs after one year, does this commonly used criterion genuinely capture long-term seizure control, or does it highlight the need for a more robust and standardized definition?
Response 4: We thank the reviewer for raising this point. In the Introduction, we aimed to outline the existing challenges regarding the definition of long-term seizure freedom. Following the reviewer’s suggestion, we expanded this discussion in more detail in the Discussion section (lines 538–547). There, we clarify that current evidence indicates that a one-year seizure-free period is insufficient to define true long-term seizure control, as relapse can still occur beyond this timeframe. We now explicitly state that more stringent, statistically supported criteria are required to ensure a robust and clinically meaningful definition of long-term seizure freedom.

Comment 5: Line 108-113; The authors need to support their statement.
Response 5: Agree. The references have been added in line 119

Comment 6: Line 157-158; The authors are advised to provide citation(s) for the released data.
Response 6: Agree. The data was removed from the Kaggle webpage and then migrated to the Epilepsyecosystem.org page. Nevertheless, we added a footnote in line 169 directing to the location of the original dataset in the Kaggle contest, and to the page within epilepsyecosystem.org with the instructions to access the data.

Comment 7: Line 159; Could the authors provide details on how these sensitivity gains were balanced against changes in other performance metrics, and how “the most challenging patient cases” were defined?
Response 7: We thank the reviewer for requesting this clarification. We have updated the manuscript to explicitly define “the most challenging patient cases” as referring to the fact that the publicly released dataset includes only the three patients with the poorest seizure prediction performance in the original Cook et al. (2013) trial (now stated in line 167-168).
With respect to how sensitivity gains were balanced against other performance metrics, the available information does not provide sufficient detail to assess how improvements in sensitivity translated into overall device performance. As discussed in lines 175–184, the NeuroVista project was ultimately discontinued, and subsequent efforts shifted from predictive systems toward forecasting approaches. We have clarified this context in the revised text.

Comment 8: Line 357-364; The authors need to clarify how reliance on subject-specific workflows affects the scalability and real-world implementation of BCI systems, and whether any generalizable or cross-subject approaches were considered or evaluated in comparison.
Response 8: We thank the reviewer for this observation. Our intention was to convey this distinction in the text: lines 376–379 emphasise that general models offer high scalability and thus greater potential for real-world deployment, whereas lines 391–394 explain that subject-specific models, while often more accurate, do not scale well beyond the individual patient. Furthermore, we have added a table between lines 402 and 403 explicitly comparing both paradigms. 

Comment 9: Line 395; The authors are encouraged to support their statement.
Response 9: We thank the reviewer for this suggestion. Although the sentence in original line 395 referred to the preceding paragraph, which already included four supporting references, we have now strengthened the statement by adding two additional citations directly within the sentence (now line 433) to make the supporting evidence more explicit.

Comment 10: Line 460-468; How did the authors assess whether the augmented or synthetic EEG data preserved physiologically meaningful features? What criteria were used to ensure that these augmentation techniques did not introduce misleading or non-biological patterns into the training set?
Response 10: We thank the reviewer for raising this important point. In the studies discussed, the quality of synthetic EEG data was primarily evaluated indirectly, through the improvement it produced in classifier performance when models were trained on augmented data but tested exclusively on real EEG. However, we note that no systematic comparison of physiological features between real and synthetic data has been performed. We have expanded the paragraph on data augmentation (lines 504–518) to emphasise this limitation in the current literature. Nevertheless, from a practical standpoint, most augmentation techniques aim to enhance model robustness rather than to generate physiologically faithful EEG signals, and some methods intentionally introduce non-biological perturbations (for example, additive noise). This lack of physiological validity becomes a concern only if synthetic data were to be included in the test set, which is not the case in the studies reviewed. We now clarify these considerations in the revised manuscript.

Additional unprompted revisions:

In addition to the changes made in direct response to the reviewers’ comments, we have implemented unprompted revisions to further improve the clarity and completeness of the manuscript. Although not the primary focus of this review, which centres on closed-loop neuromodulation, we now also briefly mention the most recently approved open-loop neuromodulation therapy (epicranial focal cortex stimulation) when introducing neuromodulation (line 58), for completeness.

Furthermore, Figure 6 (lines 432-433) has been updated to match the visual style of the newly introduced figures. The content of the figure remains unchanged. 

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

In this review, the author describes the main limitations currently restricting the progress of neural regulation in epilepsy and conceptually presents these systems as examples of the read-write structural features of brain-computer interfaces. In addition, the author has synthesized the latest advancements in this field and depicted the path of fully automatic clinical effective closed-loop neural regulation as a feasible treatment paradigm for DRE, providing effective assistance in improving the quality of life of patients. The following issues still need to be resolved before this manuscript is accepted.

1. Judging from the full text, this is a review on epilepsy, but it is hard to tell from the title that it is a review article. Therefore, the term "review" needs to be added to the title.
2. Although the author described the treatment or solutions for epilepsy in the introduction, it is difficult to effectively judge their effects merely from the language. The author should appropriately increase the data related to epilepsy treatment, so as to visually distinguish its effect.
3. Judging from the literature cited by the author, there have been reports on research related to epilepsy. Compared with the literature, the highlights of such reviews need to be highlighted in the introduction.
4. There are only two limited pictures in this manuscript. In order to enable readers to understand the current research status. The author needs to appropriately add some pictures to the manuscript to enrich its content.
5. Although the author cited a large number of references, in order to highlight the cutting-edge nature of such reviews, the author needs to appropriately delete references that are more than five years old.
6. The textual descriptions in the manuscript are a bit cumbersome, and the language in the manuscript needs to be improved.
7. The article lacks in-depth analysis of the core issue of "Why the efficacy of the closed-loop system is still far lower than that of surgery".

Author Response

Comment 1: Judging from the full text, this is a review on epilepsy, but it is hard to tell from the title that it is a review article. Therefore, the term "review" needs to be added to the title.

Response 1: We thank the reviewer for this suggestion. The current journal template already indicates the article type before the title, so we initially did not consider it necessary to include the word “review” in the title. However, we are happy to revise the title to explicitly include “review” if the reviewer feels that this would improve clarity for readers.

 

Comment 2: Although the author described the treatment or solutions for epilepsy in the introduction, it is difficult to effectively judge their effects merely from the language. The author should appropriately increase the data related to epilepsy treatment, so as to visually distinguish its effect.

Response 2: We appreciate the reviewer’s comment and have clarified this point in the revised manuscript. The Introduction summarises treatment outcomes for drug-resistant epilepsy using seizure-freedom rates (defined as no seizures three years after the intervention), which is a standard and widely accepted metric for evaluating therapeutic effectiveness. For resective surgery, we report average seizure-freedom rates of 50–66 percent (lines 45-47), supported by eight references. We then present the seizure-freedom rates of neuromodulation therapies (RNS, VNS, and DBS) in lines 97–100 and provide a direct comparison with resective surgery in lines 104–106.

To further address the reviewer’s suggestion and allow a more visual appreciation of the relative therapeutic effects, we have added a new figure (Fig. 2) between lines 106 and 107 that summarises the reported seizure-freedom rates across the different neuromodulation treatment modalities.

 

Comment 3: Judging from the literature cited by the author, there have been reports on research related to epilepsy. Compared with the literature, the highlights of such reviews need to be highlighted in the introduction.

Response 3: We agree with the reviewer that clearly positioning the present work within the existing epilepsy-related review literature is important. The reviews cited in the manuscript address a wide range of topics, and we reference them only in relation to aspects that are directly relevant to the aims of our article.

To better highlight the specific contribution of this review, we have revised the Introduction to more explicitly emphasise its novel focus and scope. In particular, we now clarify in lines 33–34 and 49–53 the key highlights of other reviews, and we explicitly identify the existing gap in lines 110–115. Unlike previous reviews, which primarily describe methods or clinical outcomes, the present work focuses on analysing why the performance of closed-loop systems has not substantially improved in the clinical field.

We believe that these revisions distinguish our review from prior literature and strengthen the presentation of its unique contribution.

 

Comment 4: There are only two limited pictures in this manuscript. In order to enable readers to understand the current research status. The author needs to appropriately add some pictures to the manuscript to enrich its content.

Response 4: We thank the reviewer for this suggestion. In response, we have expanded the visual content of the manuscript to improve clarity and provide a clearer overview of the current research landscape.

Specifically, we have added four new figures: Figure 2 presents a visual comparison of seizure-freedom rates across different neuromodulation modalities and over time, facilitating comparison with surgical outcomes; Figure 3 provides a timeline of key milestones in epilepsy neuromodulation, highlighting that no new CLS neuromodulation therapies have been approved in the past decade; Figure 4 summarises the currently available datasets in a visual and easily interpretable format; and Figure 5 offers a schematic representation of the challenges associated with defining brain states.

In addition, we included a graphical abstract to complement the main text. 

We have also added Table 3 to compare modelling paradigms, as we believe this format offers a clearer and more effective visual summary for this particular section.

 

Comment 5: Although the author cited a large number of references, in order to highlight the cutting-edge nature of such reviews, the author needs to appropriately delete references that are more than five years old.

Response 5: We thank the reviewer for this suggestion. We carefully reviewed all citations and removed or replaced older references wherever they were not strictly necessary. 

In several cases, however, retaining references older than five years is essential, either because they represent foundational work or because the relevant clinical or regulatory milestones occurred more than five years ago. 

For clarity, we summarise these cases below:

1. Resective surgery outcomes: We intentionally cite studies that are approximately 10 years old to illustrate that seizure-freedom rates have remained stable over the past decade. These are complemented by more recent papers (all <2 years old) showing comparable outcomes.
2. Neuromodulation devices (RNS, VNS, DBS): The original clinical trials establishing the effectiveness of these devices were conducted more than five years ago, so the primary evidence necessarily relies on older literature. We also include more recent follow-up or replication studies where available. 
3. tDCS and direct frequency modulation: The foundational studies for these techniques are older than five years. We retain these for historical accuracy while also citing more recent work that reflects current developments.
4. Table 1: Some references exceed five years because they document the approval status and regulatory history of neuromodulation devices, most of which were approved around a decade ago.
5. DBS trials: The Salanova et al. and Fisher et al. studies are 10–15 years old because they are the original investigations establishing DBS efficacy. We also provide more recent supporting literature.
6. NeuroVista: All references are necessarily older, as the project ended more than a decade ago and no newer studies exist. These papers are crucial for understanding the evolution of seizure prediction research.
7. BIDs: We cite the foundational papers that introduced it, which predate the five-year window.
8. Regulatory frameworks: Documents such as the EU GDPR predate five years but remain the governing regulations and must therefore be cited.
9. Definitions of ictal states: The consensus definitions were established more than five years ago and have not been superseded.
10. Surveys on patient and caregiver acceptance: Very little research has been conducted recently in this area. We cite all relevant studies we could identify, even if older, because they represent the only available evidence.

Overall, we have made every effort to update references where possible, while retaining older ones only when they are indispensable for accuracy, completeness, or historical context.

 

Comment 6: The textual descriptions in the manuscript are a bit cumbersome, and the language in the manuscript needs to be improved.

Response 6: We thank the reviewer for this comment. The manuscript has been carefully revised throughout to improve clarity, conciseness, and linguistic precision, with particular attention to reducing redundancy and enhancing readability. In addition, the manuscript was reviewed by a native English-speaking colleague, now acknowledged in the Acknowledgements section, to further ensure linguistic accuracy and fluency.

We believe that these revisions have substantially improved the overall presentation of the work.

 

Comment 7: The article lacks in-depth analysis of the core issue of "Why the efficacy of the closed-loop system is still far lower than that of surgery".

Response 7: We thank the reviewer for highlighting this central question. Addressing why the efficacy of closed-loop systems remains lower than that of resective surgery is the primary motivation of this review. In the Introduction, we note that despite substantial methodological advances reported in the academic literature, no new closed-loop neuromodulation devices have received regulatory approval in the past decade. We therefore frame the core issue as a gap between academic progress and successful clinical translation, rather than a simple limitation of neuromodulation efficacy per se.

Throughout the manuscript, we analyse several factors that may contribute to the current limitations of CLS performance, including: (1) the limited availability of large, publicly accessible datasets for algorithm development, (2) insufficient characterisation of brain states, (3) the lack of explicit assessment of brain drift, (4) the trade-off between generalised and subject-specific modelling strategies, and (5) common methodological limitations in classifier performance evaluation.

As this is a comprehensive review, we do not claim that any single factor can be definitively identified as the cause of the limited clinical efficacy or the observed translational gap. Establishing such causal relationships would require dedicated systematic and longitudinal studies. Accordingly, the original manuscript presented the available evidence and limitations to allow readers to assess the relative importance of these issues.

Nevertheless, in the revised version of the manuscript, we have made our perspective more explicit. We now emphasise that brain drift may represent a key factor underlying the efficacy gap between surgery and neuromodulation. When resective surgery is successful, the epileptogenic network is removed, effectively eliminating the source of pathological activity. In contrast, neuromodulation must continuously regulate pathological activity while simultaneously adapting to ongoing changes within the epileptogenic network itself. Most current algorithms, including those implemented in approved devices, do not explicitly account for such drift, which may fundamentally limit their long-term effectiveness. We have clarified and expanded this argument in the revised text (lines 338–363) to better address the reviewer’s concern.

 

Additional unprompted revisions:

In addition to the changes made in direct response to the reviewers’ comments, we have implemented unprompted revisions to further improve the clarity and completeness of the manuscript. Although not the primary focus of this review, which centres on closed-loop neuromodulation, we now also briefly mention the most recently approved open-loop neuromodulation therapy (epicranial focal cortex stimulation) when introducing neuromodulation (line 58), for completeness.

Furthermore, Figure 6 (lines 432-433) has been updated to match the visual style of the newly introduced figures. The content of the figure remains unchanged. 

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

The author has made the revisions as required, and the current version can be accepted.