Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Mamba-based State Space Models (SSMs) provide linear time sequence modeling but rely on dense floating-point multiply–accumulate (MAC) operations for recurrent updates. In continuous biomedical monitoring, this arithmetic density directly translates into energy cost, limiting practical deployment. SpikeMamba reformulates selective SSM dynamics in the spike domain, rewriting the recurrent update itself as an event-triggered state transition rather than a continuous dense computation. Recurrent dynamics are implemented using Leaky Integrate-and-Fire membrane equations, replacing MAC-heavy transitions with binary accumulations that occur only when spikes are emitted. A lightweight gating mechanism further suppresses updates during temporally inactive segments. Across five biomedical benchmarks, including ECG (PTB-XL) and EEG (CHB-MIT), the model operates below a 15% average firing rate while maintaining accuracy comparable to strong ANN baselines. Energy estimates under a 45nm CMOS model indicate clear reductions relative to dense ANN-Mamba implementations. Our code is available here.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4698_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/Brain-Inspired-Computing-Lab/SpikeMamba

Link to the Dataset(s)

N/A

BibTex

@InProceedings{LeeSi_SpikeMamba_MICCAI2026,
        author = { Lee, Si Yong AND Lee, Ryangjin AND Park, Hawon AND Kim, Yoora AND Kang, Byungkon AND Yang, Yoon Seok},
        title = { { SpikeMamba: Spike-Driven State Space Models for Energy-Efficient Biomedical Sequence Modeling } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16896},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    “SpikeMamba” Mamba is a selective SSM dynamics and the recurrent dynamics are implemented in this paper using Leaky Integrate-and- Fire membrane equations, with the goal of replacing MAC-heavy transitions with binary accumulations that occur only when spikes are emitted.

  • 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.

    Spike mamba layer models and operations.

  • 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.

    Temporal Spike Encoder is not clear. Function “SiLU” is not defined. Forward pass in the model is explained but loss function and training is not clearly defined.

  • Please rate the clarity and organization of this paper

    Poor

  • 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 provide sufficient information for 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?

    For the dataset “ADFTD (3-Class)” Medformer method seems to perform better.

  • 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

    For the dataset “ADFTD (3-Class)” Medformer method seems to perform better. The reason should be discussed. Is this performance related to the dataset properties?

  • 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?

    Forward pass in the model is explained but loss function and training is not clearly defined.

  • Reviewer confidence

    Somewhat confident (2)

  • [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 tests suggested by reviewers are done and results are provided.



Review #2

  • Please describe the contribution of the paper

    The paper propses SpikeMamba, a computationally efficient state-space model(SSM), that processes time-series biosignals efficiently by eliminating continuous dense updates. New model parts Leaky-integrate-and-fire (LIF) nueroon and state-dependent gate enable this by utilizing smart thresholding operation to convert the model into event-triggered updates.

  • 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.
    • Well-written paper which was super easy to follow and interesting to read.
    • The proposed SpikeMamba reduced energy consumption by several folds which is significant improvement
    • SpikeMamba matches performance with other energy-hungry models across 5 datasets
    • Wonderful step towards more efficient models
  • 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.

    • Experiments are not cross-validated across datasets: All models perform similarly across datasets, which may indicate over-optimization to a single split. Hyperparameters (e.g., threshold levels) could be tuned to that split, affecting both performance and energy results.

    • Statistical significance of performance differences: It would strengthen the work to report statistical significance via cross-validation. Ideally, show that SpikeMamba performs within the standard deviation of larger models while achieving significantly lower (e.g., 10×) energy consumption.

    • Segment-level split in EEG dataset could lead to data leakage: If segments from the same subject appear in both train and test sets, results may be inflated. Please clarify whether subject-wise splits were enforced, and why ECG uses subject-independent splits but EEG does not.

    • Inadequate baselines: EEG experiments include only one transformer-based baseline, which is insufficient. More baselines are needed, and a vanilla Mamba (non-spiking) model is an essential missing comparison.

    • No discussion on effects on non-spike phenomena: While ECG and some seizure types are spike-driven, many neurological conditions (e.g., Alzheimer’s in the paper, and others like schizophrenia and autism) are not. Others are also connectivity-based that requires looking across channels which transformers effectively deliver. It is unclear how the SpikeMamba handles such cases. Does latent thresholding mitigate this, or risk losing important non-spike information? Are depth-wise convolutions sufficient to capture cross-channel information?

  • 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
    • Future work can analyze energy vs performance using line plots
    • Analysis of spiking activity vs degree of severity of diagnosis would be interesting
    • Effect of noisy spikes and how to remove it may also be interesting. For example, muscle and eye movement spikes
    • Real-time inference on edge-devices and its efficiency
    • Analyze if the model can pick up slowing activity that signals abnormality in EEG
    • Ablation on sensitivity to threshold level
  • 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?

    It is a strong paper that takes a significant step toward reducing energy consumption and computational cost of deep networks for biosignals without meaningful loss in performance. It is well-suited for the conference and likely to spark interesting discussion. A stronger experimental framework would have led to a higher score.

  • 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.

    It is a crucial research study on reducing energy consumption and has potential to spark good discussion. Overall, my assessment is still remains to be borderline. Some of my concerns regarding cross-validation and hyper-parameters were successfully resolved by the authors. However, three main concerns remain unresolved, two of which require new results to be included to solve. 1.Questionable experimental setup with EEG dataset: It looks like the segment-level train-test splits (subject dependent) was chosen to elevate sensitivity scores of the proposed model by avoiding severe class-imbalance and other baselines were omitted for failing to detect seizures even after providing the slight advantage by manipulating data splits. 2.Inadequate baselines and ablations: Despite the promise to add new baseline and ablation results, the current manuscript does not support the performance strength of the proposed framework. 3.Discussion around adaptation to non-spike phenomena: The received response from authors did not adequately resolve this issue. It is essential to include additional comments in the paper to discuss the intuition behind the methodology, which the authors have not currently agreed to do so.



Review #3

  • Please describe the contribution of the paper

    The paper proposes SpikeMamba, an energy-efficient State Space Model (SSM) architecture designed for biomedical time-series analysis. The primary contribution is mapping the Mamba selective scan mechanism into the spike domain. By utilizing Leaky Integrate-and-Fire (LIF) neuron dynamics, the model converts continuous dense multiply-accumulate (MAC) operations into sparse, event-driven accumulations (AC). Additionally, the authors introduce a State-Dependent Gate to suppress state updates during temporally inactive segments. The model is evaluated on five biomedical benchmarks across ECG and EEG modalities.

  • 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.Formulating SSMs (Mamba) within the spike domain is an interesting and original approach. Replacing MAC-heavy transitions with binary accumulations via LIF membrane equations provides an theoretical solution for sequential modeling on resource-constrained devices. 2.The experimental design evaluates the proposed architecture across five distinct benchmarks spanning two critical biomedical modalities (ECG and EEG). This demonstrates the general applicability of the method. 3.The theoretical operation profiling is impressive, showing that over 96% of dense MAC operations can be converted into energy-efficient AC operations.

  • 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.Missing Critical Ablation Studies The State-Dependent Gate is highlighted as a primary contribution (Contribution 3) for suppressing redundant updates. However, section 4.1 explicitly states that All variants use the state-dependent gate. There is no ablation study isolating the impact of this gate. It is currently impossible to determine how much of the energy savings or potential accuracy degradation is attributable to the core LIF dynamics versus this specific gating mechanism. 2.Inconsistent Baseline Comparisons In Table 2, the proposed method is compared against a robust set of baselines (iTransformer, MTST, Reformer, Medformer). However, in Table 3, the comparison is restricted only to Medformer. The omission of the other state-of-the-art baselines makes the evaluation appear incomplete and raises concerns about how SpikeMamba truly stacks up against contemporary models on these specific tasks. 3.Lack of Hardware or Simulated Validation While the energy efficiency claims are a core motivation, they are entirely theoretical—derived from FLOPs counting and a standard 45 nm CMOS energy model. For a paper targeting edge deployment, the lack of physical neuromorphic hardware validation (e.g., on Loihi 2) or cycle-accurate simulation weakens the practical impact. It ignores potential real-world bottlenecks such as memory access energy, hardware routing overhead, and latency. 4.Hyperparameter Sensitivity The architecture introduces several key hyperparameters, such as the Spike Firing Threshold and the learnable gate threshold. The paper lacks an analysis of how sensitive the model’s accuracy and spike firing rate are to these specific values, which is important for understanding the robustness of the SNN conversion. 5.Doubts exist regarding the reproducibility of the experiment. Please make the necessary codes and data publicly available.

  • 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 provide sufficient information for 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?

    My overall recommendation is a Weak Reject, primarily driven by critical gaps in the experimental validation and ablation studies, despite the paper’s highly promising theoretical premise.

    The paper presents an original and theoretically elegant approach by formulating State Space Models (Mamba) within the spike domain. The idea of replacing MAC-heavy transitions with binary accumulations via LIF dynamics is a strong theoretical step forward for sequence modeling on resource-constrained devices. Furthermore, the broad evaluation across five distinct ECG and EEG benchmarks, alongside the impressive theoretical operation profiling (>96% MAC-to-AC conversion), highlights the model’s potential applicability and efficiency.

    However, several major factors currently pull the paper below the acceptance threshold:

    1.Unverified Core Contributions: The State-Dependent Gate is claimed as a primary contribution, yet the complete absence of ablation studies makes it impossible to disentangle its specific impact on energy savings and accuracy from the core LIF dynamics. 2.Incomplete Evaluation: The baseline comparisons are noticeably inconsistent. Dropping the robust set of baselines used in Table 2 down to just a single baseline (Medformer) in Table 3 makes the evaluation incomplete and leaves the model’s true comparative standing ambiguous for those specific tasks. 3.Lack of Real-World Hardware Validation: The impressive energy claims are purely theoretical. Without physical neuromorphic hardware validation or cycle-accurate simulation, the paper overlooks critical real-world edge deployment bottlenecks like memory access energy, hardware routing overhead, and latency. 4.Robustness and Reproducibility Concerns: The omission of a hyperparameter sensitivity analysis (particularly regarding the Spike Firing and gate thresholds) leaves questions about the robustness of the SNN conversion. Additionally, the lack of publicly available code and data hinders the reproducibility of these highly specific claims.

    Summary for Rebuttal: I am open to raising my score to an Accept if the authors can use the rebuttal period to provide the missing ablation studies on the gating mechanism, complete the baseline comparisons for Table 3, and provide a clear commitment to open-sourcing their code and data for reproducibility.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Author Feedback

We respectfully thank all reviewers for their constructive feedback. We address the concerns below in a consolidated manner, grouping related issues across reviewers. [Clarity] (R1-W1,W2): SiLU(x)=x·σ(x) (Elfwing et al., 2018); will be added in Eq.(2). Training details (weighted CE, AdamW lr=1e-3, cosine annealing, early stopping, surrogate gradient) will be consolidated in the revision. [Statistical Robustness] (R2-W1,W2): We performed 10-run bootstrap resampling on all models and report mean±std. For example, SpikeMamba-L on PTB achieves 91.72±0.39% vs Medformer 83.50%. The full mean±std results were omitted due to page limits and will be included in the revision. [State-Dependent Gate Ablation] (R3-W1): We ablated the gate by comparing SpikeMamba-L with and without the State-Dependent Gate across three datasets: ADFTD (53.62% vs 53.22% Acc), PTB (91.72% vs 90.79% Acc), PTB-XL (74.16% vs 73.97% Acc). The gated variant consistently outperforms by selectively suppressing redundant spikes during inactive segments. [Hyperparameter Sensitivity Ablation] (R2-W1,W2; R3-W1,W4): All gate parameters (κ, δ, λ₁, λ₂, τ_g) are learnable and adapt during training. We swept V_th ∈ {0.5, 0.75, 1.0, 1.5, 2.0} and κ ∈ {0.05, 0.1, 0.3, 0.4, 0.45, 0.5} on PTB-Medium. V_th=1.0 yields the best accuracy (89.00%) and F1 (92.27%); performance degrades symmetrically for lower/higher values. κ ∈ [0.01, 0.3] maintains stable accuracy (87.98–89.07%). Increasing κ reveals the gate’s energy–accuracy trade-off: at κ=0.5, energy drops to 22.11 mJ, a 39.5% reduction vs. the no-gate baseline (36.55 mJ) at the cost of 4.07% accuracy and 2.91% F1.At κ=0.45, energy is 23.65 mJ with only 2.61% accuracy and 2.26% F1 trade-off, demonstrating the gate as a controllable energy–performance knob beyond core LIF dynamics. These results will be included in the revision. [Baseline Comparisons] (R2-W4; R3-W2): We include parameter-matched Vanilla Mamba (8L-404D, ~8.6M params) as a direct non-spiking baseline across all five datasets. At identical parameter count, SpikeMamba-L outperforms Vanilla Mamba on PTB (91.72% vs 89.16%) and matches on PTB-XL (74.16% vs 74.23%). On CHB-MIT, SpikeMamba-L achieves substantially higher seizure sensitivity (89.15% vs 71.26%). Across all datasets, SpikeMamba consumes 30–40× less energy (e.g., 322.91 vs 12,184.98 mJ on PTB). For Table 3, we evaluated iTransformer, MTST, and Reformer on CHB-MIT and MIT-BIH. On CHB-MIT, iTransformer and MTST failed to detect seizures (sensitivity 0% and F1≈0%). Reformer achieved 80.26% accuracy with 78.21% sensitivity, but SpikeMamba-M outperformed it in accuracy (88.57%), F1 (80.02%), and AUROC (94.48% vs 88.19%) at 10.5× lower energy (56.86 vs 598.41 mJ). On MIT-BIH, all models achieved comparable accuracy (98.46–99.02%), with SpikeMamba-M consuming 5.6–44.8× less energy. These results will be included in the revised manuscript. [EEG Data Split] (R2-W3): CHB-MIT uses segment-level split (6:2:2) as few seizure events per subject make subject-wise splitting impractical. PTB, PTB-XL, and ADFTD use subject-independent splits following Medformer. [ADFTD & Non-Spike Phenomena] (R1-W4; R2-W5): Medformer’s F1 advantage on ADFTD (50.65% vs 46.58%) reflects its cross-resolution routing for heterogeneous EEG patterns. SpikeMamba-L matches accuracy (53.62% vs 53.27%) at 44× lower energy (262 vs 11,590 mJ). The LIF neuron operates in latent space; depthwise convolutions and learnable leak coefficient α capture cross-channel and non-spike temporal dynamics. [Hardware & Future Work] (R2-Future; R3-W3): Current energy estimates are theoretical. We plan deployment on MindCore which is an FPGA-based neuromorphic hardware system to validate real-world energy/latency. Remaining suggestions will be addressed in the revised Future Work section. [Reproducibility] (R1,R2,R3): Full source code, hyper-parameters, and preprocessing scripts will be released 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.

    Please prioritize clarifying the method and training procedure, including the temporal spike encoder, SiLU definition, loss function, and optimization details. Additionally, address concerns about experimental rigor, especially subject-wise splitting and possible EEG leakage, cross-validation or statistical significance, missing baselines such as vanilla Mamba, and inconsistent baseline comparisons across tables.

  • 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 paper proposes SpikeMamba, a computationally efficient state-space model for time-series biosignals that avoids continuous dense updates and introduces new architectural components, including state-dependent gates. The proposed design is promising not only for biosignal analysis but also as a potentially extensible approach for other sequential modeling applications. A key strength of SpikeMamba is its substantial reduction in energy consumption, achieving several-fold improvements while maintaining competitive performance against more energy-intensive models across five datasets. During review, several concerns were raised regarding experimental rigor, method clarification, and training details. In the rebuttal, the authors provided a comprehensive response, clarifying important aspects including statistical robustness, state-dependent gate ablations, baseline comparisons, and hyperparameter sensitivity analysis. These clarifications were recognized by the reviewers, some of whom increased their scores and recommended acceptance. Overall, the paper presents a strong and practically relevant contribution, and is recommended for acceptance.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Reject

  • Please justify your recommendation.

    After considering the rebuttal and updated reviews, I recommend rejection. Although the paper addresses an important topic, reviewers still raised several substantial concerns after the rebuttal. In particular, some key issues would require additional experimental evidence that cannot be adequately resolved at this stage. Therefore, I do not think the paper meets the acceptance threshold.



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.

    This paper proposes SpikeMamba, an event-driven state-space architecture for energy-efficient biomedical sequence modeling. The idea of integrating spiking dynamics with state-space modeling is interesting and relevant for resource-constrained biomedical monitoring. The submitted manuscript evaluates the method across five ECG and EEG benchmarks and reports competitive predictive performance together with substantial theoretical energy reductions.

    The reviewers raised valid concerns regarding the completeness of baseline comparisons, the lack of dedicated ablation for the state-dependent gate, and the theoretical nature of the energy evaluation. Nevertheless, the rebuttal clarifies the training setup and provides a reasonable response to the major concerns. Overall, I consider the work novel and promising enough for acceptance, while recognizing that the experimental validation could be strengthened in future work.



back to top