acceptodds
Under review as a conference paper at ICLR 2027

BioSQ: Frequency-Aware Neural Compression of Biosignals

Abstract

Neural codecs have recently emerged as effective compression models for time-series data, offering a compelling alternative to classical algorithms. For the most part, modern neural codecs are designed to compress audio recordings, such as music and speech, with their architectures and quantization strategies targeted to those modalities. Therefore, extending neural codecs to other modalities poses specific challenges that require careful consideration. In this work, we evaluate the suitability of neural compressors to a variety of biosignals: scalp electroencephalogram (EEG), intracranial EEG, and electrocardiogram (ECG). In particular, we find that global fidelity metrics, commonly used to evaluate compression performance, are insufficient to fully characterize the reconstruction of biosignals. Remarkably, our analysis shows that global fidelity can obscure frequency-localized distortions with disproportionate effects on downstream accuracy. Guided by these findings, we systematically revisit the main components of neural compressors, evaluating alternative quantization strategies and encoder–decoder architectures, and combine them with classical information-theoretic insights. On the one hand, we observe that implicit quantization schemes outperform classical vector quantization. In fact, we introduce residual binary spherical quantization (RBSQ), which extends binary spherical quantization with multiple residual stages, and show that it notably improves fidelity. On the other hand, we find that asymmetric encoder-decoder architectures, while improving reconstruction of audio data, notably underperform on biosignals. Based on our thorough analysis, we introduce BioSQ as the state-of-the-art (SoTA) neural codec for biosignals. We further reduce the bitrate of BioSQ through entropy regularization and arithmetic coding, promoting stable and uniform symbol distributions. BioSQ outperforms SoTA neural biosignal compressors in reconstruction fidelity across 4 physiological datasets and compression ratios from 4× to 128×, reducing reconstruction error by up to 81%. These improvements are also reflected in downstream tasks, where BioSQ better preserves performance on seizure detection, motor-imagery classification, and QRS detection by up to 25% as compression becomes more restrictive. Overall, our results show that neural codec design choices developed for audio do not transfer uniformly to physiological signals, and establish BioSQ as an effective approach to general-purpose biosignal compression.

Primary area:
learning on time series and dynamical systems
Keywords:
Neural compression, Biosignal compression, Vector quantization, Binary Spherical Quantization, Frequency-band analysis

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_KF24OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper evaluates neural compression methosd for EEG, iEEG, and ECG signals. The authors argue that global fidelity metrics such as PRD or RMSE offer limited insight into the error distribution across particular frequency bands, that may affect downstream physiological analysis. They demonstrate this with a band-swapping experiment on CHB-MIT seizure detection at 64× compression. They therefore utilize a bandwise PRD as a main diagnostic to compare quantizers and encoder-decoder variants and report that lookup-free quantizers beat codebook-based ones and a symmetric convolutional encoder-decoder is the best tradeoff between accuracy and cost.

Based on these findings, the authors introduce their codec BioSQ, which pairs a symmetric convolutional encoder-decoder with Residual Binary Spherical Quantization (RBSQ). RBSQ is a multi-stage residual extension of BSQ trained with per-stage entropy regularization and followed by arithmetic coding of the binary indices. They evaluate reconstruction fidelity of BioSQ against BrainCodec and classical compression baselines on four public datasets at compression ratios from 4x to 128x, with BioSQ reporting lower PRD than the baselines. They also demonstrate better downstream task performance of BioSQ compared against BrainCodec. The authors claim that BioSQ is the first neural codec effective across both brain and cardiac signals.

Critical strengths

1. Relevant, underexplored problem. Long-term EEG, iEEG and ECG monitoring produces large data volumes, and neural compression of biosignals has received far less attention than audio or images. Testing which audio-codec design choices transfer to physiological signals is a worthwhile question.

2. Clinically motivated evaluation. The paper convincingly demonstrates that global PRD can mask band-specific distortions. The band-swap experiment (Fig. 2) shows that perturbations with similar effects on global PRD can affect downstream seizure detection very differently.

3. Broad empirical coverage. The evaluation spans four public datasets, three modalities (scalp EEG, iEEG, ECG), compression ratios from 4× to 128×, and three downstream task types (seizure detection, motor imagery, QRS detection). This provides a useful evaluation across different physiological signals.

4. Systematic comparison of quantizers and architectures. The authors investigate RVQ, ERVQ, SNAC, BSQ, RFSQ, and RBSQ, along with alternative encoder-decoder designs. Both successful and unsuccessful architectural choices are reported. The appendix further includes useful analyses of codebook utilization, reconstruction limits, frequency-band specialization, and RBSQ stage configurations.

5. Simpler training. BioSQ does not require adversarial training and generally matches or outperforms BrainCodec, especially at high compression ratios, where downstream gains are largest.

Critical weaknesses

1. Frequency awareness is limited to evaluation and model selection. Despite the title, BioSQ does not explicitly incorporate frequency-dependent compression mechanisms, such as band-weighted losses or frequency-adaptive bit allocation. The frequency-aware analysis is valuable, but the final codec is not itself explicitly frequency-adaptive.

2. The "first neural codec across brain and cardiac signals" claim is overstated. NeuroRVQ (arXiv:2510.13068) reconstructs EEG, ECG and EMG with RVQ tokenizers and releases ECG weights. PhysioOmni (arXiv:2504.19596) and Hypnos (arXiv:2606.09605) also tokenize EEG and ECG. Though these works do not report compression ratio and PRD, so a narrower claim may hold, but they should be cited and discussed.

3. Frequency-aware evaluation is missing from the final quantitative comparison. Although the paper argues that global PRD can mask frequency-localized errors, band-wise PRD is used only to compare BioSQ variants internally, while the main comparison against prior methods (Fig. 4) rely solely on global PRD. No band-wise comparison against BrainCodec is provided, even though it could explain the cases where BrainCodec achieves lower global PRD but BioSQ performs better downstream.

4. Limited statistical reporting. The results include means and medians but no confidence intervals, variability across training runs, or significance tests. This is particularly relevant for the smaller improvements in Table 1 and the BCI IV-2a downstream results. More information about random seeds and experimental variability is needed.

5. Minor presentation and reproducibility issues.

  • The architecture is called SEANet in Fig. 1 but described as 2D (K,1) convolutions in Sec. 3.
  • F = 16 (Sec. 4) is undefined.
  • The classical baseline is labeled NLSPIHT in the CHB-MIT and SWEC-ETHZ panels of Fig. 4 but SPIHT in the BCI IV-2a panel, while Sec. 4.2 lists only NLSPIHT.
  • In Figs. 3 and 14 (top row), the proposed quantizer is labeled “BioSQ” rather than “RBSQ”. BioSQ denotes the full system, while the other legend entries are quantizers.
  • Figs. 3 and 14 do not state the dataset used, or whether band PRD values are means or medians and over which subjects or segments.
Questions for authors

1. Given the motivation and title, did the authors explore making the codec itself frequency-aware, for example through band-weighted reconstruction losses or frequency-adaptive bit allocation?

2. How do you reconcile the claim of being the first neural codec effective across brain and cardiac signals with NeuroRVQ, PhysioOmni and Hypnos? Please cite them and consider comparing against NeuroRVQ's public ECG tokenizer.

3. Could the authors provide band-wise PRD comparisons between BioSQ and BrainCodec on the EEG and iEEG datasets, particularly at operating points where BrainCodec has lower global PRD but BioSQ achieves better downstream performance?

4. Is a single model trained across all modalities and compression ratios, or are separate models trained per dataset and compression ratio? Could the authors also provide random seeds or variability estimates?

5. Please reconcile the abstract's "up to 25%" with the conclusion's 12%, 2% and 22%, and state the dataset and compression ratio at which the "81%" reduction is obtained.

4 comments

We thank the Reviewer for highlighting the importance of the problem which is the focus of this work, and for appreciating our broad and thorough evaluation.

Following the Reviewer's suggestions, we now clarify the practical relevance of our frequency aware evaluation, and improve the positioning of our paper with respect to the existing literature. We also add EMG as an additional modality to strengthen the generality of our model, and improve the clarity of our manuscript.

We address each point in detail below.


W1. Frequency-aware pipeline

We thank the Reviewer for raising this point, which allows us to better clarify the novelty of BioSQ and especially of its frequency-based analysis. As described in the manuscript, the entire development pipeline for BioSQ was guided by the evaluation on biologically relevant frequency bands rather than on the global reconstruction quality, with the end goal of improving reconstruction of real-world morphological features (see Appendix D) and classification performance.

This frequency-aware setup directly allowed us to identify highly relevant and novel effects, such as the decoupling between global reconstruction quality (PRD) and downstream classification accuracy (see Section 5.1). We now better highlight the value of this contribution in the Introduction and in Section 5.1. The frequency-based analysis also guided our choice of quantizer, which settled on an augmented version of BSQ, RBSQ, which achieved the best reconstruction fidelity at lower, more biologically relevant frequencies, in exchange for lower fidelity at higher frequencies (see Figure 3 at CR=64).

Following the Reviewer's suggestion, we now directly manipulate the band fidelity of BioSQ during training by adding a band-restricted spectral term to the training loss with weight λB\lambda_B (full details in Appendix H). As the gradient balancer rescales the gradient of each term at the output, the band term receives a share λB/(2+λB)\lambda_B/(2+\lambda_B) of the gradient norm: 33% for λB=1\lambda_B=1 and 50% for λB=2\lambda_B=2. We report the results for β and γ-low on CHB-MIT at CR=16 below:

β PRD (13-30 Hz)γ-low PRD (30-70 Hz)γ-high PRD (70-120 Hz)Median PRD
Baseline (λB=0\lambda_B=0)36.348.651.410.28
β band, λB=1\lambda_B=116.356.966.911.09
β band, λB=2\lambda_B=214.157.977.611.44
γ-low band, λB=1\lambda_B=145.342.778.911.03
γ-low band, λB=2\lambda_B=260.643.097.312.31

As expected, differential weighting helps the targeted frequency band at the cost of the others. Such a solution may be attractive for a task-specific compressor, but the band worth favouring depends on the task. Indeed, in seizure detection (Fig. 2), the largest gain comes from β. We now repeat the same band-swap analysis on motor imagery (BCI IV-2a, CR=64): we replace one band of the generalist reconstruction by the reconstruction of a specialist trained only on that band, and report the change in accuracy. A summary Table is below:

Band replacedΔ accuracy (points, 95% CI)
δ-0.07 [-0.60, +0.45]
θ+0.19 [-0.07, +0.48]
α+1.41 [+0.59, +2.27]
β+0.55 [+0.08, +1.01]
γ-low-0.10 [-0.42, +0.22]
γ-high-0.12 [-0.41, +0.15]

In motor imagery, the largest gain comes from α, with smaller effects for the other bands. A codec weighted towards β would therefore be tuned to one task and not to the other. Therefore, we keep band agnostic training for BioSQ and describe this finding in detail in Appendix E. This additional contribution suggested by the Reviewer may certainly improve task performance for practitioners and is of high interest.

W2. Related works

We thank the Reviewer for pointing out these highly relevant methods. Indeed, all three papers are interesting and we now include them in the Related Works section. We highlight in the Table below the key similarities and differences with our BioSQ:

NeuroRVQPhysioOmniHypnosBioSQ
ArchitectureTokenizerTokenizerTokenizerNeural compressor
QuantizerMulti-branch RVQ, one tokenizer per modalityMultimodal VQRVQRBSQ (8 stages × 4 bits)
SignalsEEG, ECG, EMGEEG, ECG, EOG, EMGPolysomnography (incl. ECG)EEG, iEEG, ECG, EMG
Reported metricsValidation MSENone for reconstructionSNR, correlationCR, PRD, band PRD, downstream tasks
Compression ratio / bit rateNot reportedNot reportedNot reported4× to 128× (2,048 to 64 bit/s per channel)
ECG nominal rate (our calculation)~2,000 bit/s per channelNot available~44 bit/s2,048 bit/s (4×), 64 bit/s (128×)
Entropy codingNoNoNoPer-stage arithmetic coding
Frequency-band evaluationNoNoNoYes

See the rest of the response in the following comment.

It must be noted that all three works are designed to be tokenizers for foundation models, rather than biosignal compressors. Therefore, the design choices and evaluation results are notably different from ours, and the models are much larger in both parameters and training dataset. By our calculation from their published configurations, the nominal rate of the ECG tokenizer of NeuroRVQ is about 2,000 bit/s, similar to BioSQ at CR 4×, while that of Hypnos is about 44 bit/s, below BioSQ at CR 128× (about 64 bit/s).

Following the Reviewer's suggestion, we evaluate the publicly released EEG tokenizers of NeuroRVQ and Hypnos on the same CHB-MIT test windows as BioSQ and BrainCodec, and compute the PRD from their reconstructions. Both tokenizers are used as released, without retraining. Both models low-pass the input at 45 Hz, so we report the PRD against the original signal and the PRD after band-passing both signals to 0.5-45 Hz.

Codecbit/s per channelPRD, full bandPRD, 0.5-45 Hz
BioSQ (CR 128×)6443.538.5
Hypnos8848.833.1
BioSQ (CR 64×)12830.326.8
NeuroRVQ41648.627.0
BioSQ (CR 16×)51210.37.6

In the 0.5-45 Hz band, NeuroRVQ at 416 bit/s reaches the PRD that BioSQ reaches at 128 bit/s, and Hypnos at 88 bit/s is comparable to BioSQ at the same rate. On the full band both tokenizers are worse, since they do not reconstruct frequencies above 45 Hz. We emphasize that these models are tokenizers for foundation models: they were trained on different data and sampling rates, and were not designed for fidelity at a fixed bit rate. We therefore read this comparison as a reference point and not as a claim of superiority. We add the results and the evaluation protocol to Appendix I.

For ECG, the publicly released NeuroRVQ tokenizer (2,080 bit/s per lead), evaluated on a sample of 80 MIT-BIH windows shared with BioSQ and BrainCodec, reaches a PRD of 1.5%. This is similar to the 1.6% of BioSQ at CR 64× (128 bit/s), which uses 16 times fewer bits, and higher than the 0.55% of BioSQ at CR 16× (512 bit/s). We include this comparison in Appendix I.

To further strengthen the generality of BioSQ, we also add electromyography (EMG) as a fourth modality. We use the NinaPro DB2 dataset [4] and train one BioSQ and one BrainCodec model per compression ratio, with the same pipeline and configuration as for the other datasets. The Table below shows the median PRD over held-out subjects at the same nominal rate, without arithmetic coding, as in our answer to Reviewer T4sm (W1).

DatasetCodec4× ↓16× ↓64× ↓128× ↓
NinaPro DB2 (EMG)BrainCodec2.3716.0156.9472.22
NinaPro DB2 (EMG)BioSQ4.5014.7053.1670.09

As on SWEC-ETHZ and MIT-BIH, BrainCodec is better at 4×, and BioSQ is better from 16× on. We read this as evidence that the approach extends to EMG, and add the results to Appendix K.

Finally, we now better clarify the role of BioSQ in the Introduction as specifically targeting compression in real-world clinical and industry application, rather than dataset tokenization.

W3. Frequency-aware evaluation

We show the results of the quantitative frequency-aware evaluation in Sections 5.1 and 5.2. Specifically, Figure 3 (top row) compares band-wise PRD on CHB-MIT between RVQ (the quantizer of BrainCodec), and RBSQ with the same encoder and decoder. RBSQ's PRD is lower in every band at 4×, 16× and 64×. Following the Reviewer's suggestion, we now add the same comparison for the iEEG dataset, shown below at CR 4× and 16× and added to Appendix G in full. Band and global PRD are computed in the same way, on the same test windows for both codecs, so the global values may differ slightly from the per-subject medians of Figure 4.

PRD (%)Globalδθαβγ-lowγ-high
CR 4×, BrainCodec0.70.10.20.40.92.97.1
CR 4×, BioSQ1.40.40.51.01.94.911.2
CR 16×, BrainCodec4.51.01.32.66.721.847.1
CR 16×, BioSQ4.20.60.81.63.516.254.4

At CR 4×, BrainCodec's PRD is lower in every band, in line with Figure 4. Errors up to β are below 2%, and both codecs reach the F1 of the original signal (Figure 5). At CR 16×, the global PRD is similar, while BioSQ is better from δ to γ-low, most clearly in β, and worse in γ-high, which holds a small part of the signal energy. While the per-band analysis does improve our understanding of the compression performance, we cannot claim that it fully explains this behaviour. We now state the dataset and the aggregation in the captions of Figures 3 and 14.

See the rest of the response in the following comment.

W4. Statistical reporting

Following the Reviewer's suggestion, we now add subject-level statistics for Figure 4. The Table below shows the median PRD over held-out subjects, the median paired difference between BioSQ and BrainCodec with its 95% bootstrap interval and the p-value of a paired Wilcoxon signed-rank test.

DatasetCRBrainCodecBioSQDifferencep
CHB-MIT4×6.261.20−4.9 [−5.7, −2.7]<0.001
CHB-MIT16×15.0410.28−5.0 [−6.0, −3.5]<0.001
CHB-MIT64×33.5830.34−3.6 [−4.0, −3.0]<0.001
SWEC-ETHZ4×0.781.39+0.5 [+0.4, +0.6]<0.001
SWEC-ETHZ16×4.204.25+0.2 [−0.2, +0.3]0.34
SWEC-ETHZ64×16.4915.99−0.8 [−1.7, −0.4]<0.001
SWEC-ETHZ128×30.9327.81−2.7 [−3.2, −2.5]<0.001

BioSQ is significantly better on CHB-MIT at all compression ratios, and on SWEC-ETHZ from 64×. BrainCodec is better on SWEC-ETHZ at 4×, and the two are tied at 16×. Moreover, we also add the same analysis to the downstream task on BCI IV-2a. The Table below shows the mean accuracy over held-out subjects, as in Figure 5, and the mean paired difference in accuracy points with its 95% interval across subjects.

CRBrainCodecBioSQDifference
4×73.773.6−0.03 [−0.31, +0.24]
16×73.273.3+0.08 [−0.79, +0.95]
64×70.772.4+1.72 [−2.20, +5.64]
128×64.265.7+1.59 [+0.70, +2.48]

The two codecs are equivalent up to 16×, with differences below 0.1 points. We now add these results to Appendix J and better clarify the positioning of BioSQ with respect to the SoTA.

W5. Minor presentation issues

We thank the Reviewer for helping us improve our manuscript. We now address each and update the manuscript accordingly:

  1. SEANet is a 1D convolutional encoder-decoder. In Section 3 we write its convolutions as 2D with kernel (K,1) over the channel and time axes. This is equivalent to applying the same 1D convolution over time to each channel independently, as stated in Section 4. We now mention this explicitly in Section 3.
  2. F is the number of channels of the first convolution. It doubles after each downsampling layer up to a maximum of 256. We now define it in Section 4.
  3. The classical baseline indeed differs between datasets. In CHB-MIT and SWEC-ETHZ it is NLSPIHT [1], with results taken from [2]. For BCI IV-2a no NLSPIHT results are available, and the panel shows the SPIHT results reported by [3] together with AAC. We now use a label that makes this distinction clear, correct the caption of Figure 4, and list both baselines in Section 4.2.
  4. The Reviewer is indeed correct, in the top row of Figures 3 and 14 the encoder-decoder is fixed and only the quantizer changes, so the proposed quantizer should be labeled RBSQ. We now relabel it. BioSQ in the bottom row denotes the full system, RBSQ with the SEANet encoder-decoder.
  5. Figures 3 and 14 show CHB-MIT. Band PRD is the PRDb of Eq. 3 computed over all test windows, so it is neither a mean nor a median across subjects. We now state the dataset, the aggregation and the test subjects in the captions.

Q1. Frequency-based losses

We address this in W1.

Q2. Related works

We address this in W2.

Q3. Band-wise PRD comparisons

We address this in detail in W3.

Q4. Per-modality model

We train a separate model for each dataset and each compression ratio, using the same pipeline and configuration for each. Each model is evaluated on held-out subjects that are disjoint from its training subjects. We now make this explicit in Section 4. Variability across held-out subjects is reported in our answer to W4. Training one model per dataset and compression ratio over multiple random seeds is computationally expensive. These runs are in progress, and we will add the results once they finish.

Q5. Manuscript improvements

We thank the Reviewer for pointing out this improvement. The 81% is the reduction of the median PRD on CHB-MIT at CR = 4×, from 6.26% for BrainCodec to 1.20% for BioSQ. We now state the dataset and the compression ratio in the Abstract and in the Conclusion. The Abstract and the Conclusion report different values for the same quantity, the downstream gain over BrainCodec. The relative improvement over BrainCodec at 128× is 12% (seizure detection, CHB-MIT), 5% (seizure detection, SWEC-ETHZ), 2% (motor-imagery classification) and 22% (QRS detection). We now replace “up to 25%” with “up to 22%”.

[1] Xu, Gaowei, et al IEEE Signal Processing Letters (2015). [2] Carzaniga, Francesco, et al ICLR (2025). [3] Nguyen, Binh, et al. IJCNN (2017). [4] Atzori, Manfredo, et al Scientific Data 1, 140053 (2014).

We greatly appreciate the time and effort dedicated by the Reviewer to improve our work. As the discussion period is coming to an end, we kindly ask the Reviewer to acknowledge whether we have successfully addressed their outstanding concerns. We remain available to further discuss any remaining question or weakness in our work.

Reviewer_cSvFOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies neural compression of biosignals. The paper first shows hidden frequency-localized reconstruction errors of global fidelity metrics by swapping frequency bands between a full-band model and band-specific models on CHB-MIT at 64× compression. Then authors propose BioSQ, which is composed of (i) RBSQ which applies BSQ over multiple residual stages, (ii) per-stage entropy and bit-balance regularization; (iii) a symmetric convolutional encoder–decoder and (iv) static per-stage arithmetic coding. BioSQ is evaluated on four public datasets covering global and band-wise PRD, morphological features and downstream tasks. The paper reports up to 81% lower PRD and up to 25% better downstream performance than BrainCodec.

Critical strengths

S1. The idea and writing is overall clear and explicit. The motivation is reasonable and the results proves the proposed method outperforms other existing works in most cases.

S2. The experimental results are strong and insightful. The paper reports more detailed metrics and multiple complementary perspectives.

Critical weaknesses

W1. The novelty is limited. L187 said "iterative quantization benefits lookup-free methods" but gave no explanation.

W2. The title of the paper mentioned "Frequency-aware", however, the proposed method doesn't utilize frequency band-wise training method. Subsequent findings of β\beta bands were not reflected in the processing pipeline.

W3. There are some writing contradictions in the paper:

  1. L030 claimed they achieved sota results. But L396 gave results which showed BrainCodec achieved better performance at low CR.
  2. L037 said they improved downtask performance by up to 25%, but in Sec 6 this number turned into 22%.
  3. L031 said that BioSQ prompts uniform symbol distributions, however, uniform distributions will result in the entropy which is to the maximum possible number of bits, leaving little room for gains from arithmetic coding.
  4. L049–050) identified spectral distortion caused by strided/transposed convolutions; yet, BioSQ employs precisely these two types of operations.
Questions for authors

Clarification requests

  1. Because the number of stages appears to strongly affect PRD at low compression ratios. Could the authors report RVQ with the same stage structure (8 stages × 16-entry codebooks), and RBSQ with 4 stages × 8 bits?
  2. Fig. 4 plots BioSQ at its effective compression ratio after arithmetic coding, but BrainCodec at its nominal ratio. Could the authors show both codecs at nominal rate, or both with arithmetic coding applied?
  3. The residuals are not normalized, and the entropy term is also calculated based on the unnormalized vv; consequently, the spherical constraint loses the role it plays in BSQ. So please clarify the effect of spherical constrains.
  4. In the Eq.2, the entropy term could be trivially minimized by scaling up WinW_{in}, without any change to the quantized output. In addition, the per-bit balance term does not prevent symbol-level collapse. Could the authors clarify how these issues are avoided?
  5. Given the paper’s own findings on the β band and on band-specialized models, did the authors consider band-weighted spectral losses or frequency-dependent bit allocation?
4 comments

We thank the Reviewer for recognizing the strength of our experiments and the insightfulness of our approach.

Following the Reviewer's suggestions, we now clarify the novelty of BioSQ and the usefulness of the frequency-based analysis. We also broaden our ablations and improve the writing of the manuscript.


W1/2. Methodological novelty

We thank the Reviewer for raising these points, which allow us to better clarify the novelty of BioSQ and especially of its frequency-based analysis. As described in the manuscript, the entire development pipeline for BioSQ was guided by the evaluation on biologically relevant frequency bands rather than on the global reconstruction quality, with the end goal of improving reconstruction of real-world morphological features (see Appendix D) and classification performance.

This frequency-aware setup directly allowed us to identify highly relevant and novel effects, such as the decoupling between global reconstruction quality (PRD) and downstream classification accuracy (see Section 5.1). We now better highlight the value of this contribution in the Introduction and in Section 5.1. The frequency-based analysis also guided our choice of quantizer, which settled on an augmented version of BSQ, RBSQ, which achieved the best reconstruction fidelity at lower, more biologically relevant frequencies, in exchange for lower fidelity at higher frequencies (see Figure 3 at CR=64).

Following the Reviewer's suggestion, we now directly manipulate the band fidelity of BioSQ during training by adding a band-restricted spectral term to the training loss with weight λB\lambda_B (full details in Appendix H). As the gradient balancer rescales the gradient of each term at the output, the band term receives a share λB/(2+λB)\lambda_B/(2+\lambda_B) of the gradient norm: 33% for λB=1\lambda_B=1 and 50% for λB=2\lambda_B=2. We report the results for β and γ-low on CHB-MIT at CR=16 below:

β PRD (13-30 Hz)γ-low PRD (30-70 Hz)γ-high PRD (70-120 Hz)Median PRD
Baseline (λB=0\lambda_B=0)36.348.651.410.28
β band, λB=1\lambda_B=116.356.966.911.09
β band, λB=2\lambda_B=214.157.977.611.44
γ-low band, λB=1\lambda_B=145.342.778.911.03
γ-low band, λB=2\lambda_B=260.643.097.312.31

As expected, differential weighting helps the targeted frequency band at the cost of the others. Such a solution may be attractive for a task-specific compressor, but the band worth favouring depends on the task. Indeed, in seizure detection (Fig. 2), the largest gain comes from β. We now repeat the same band-swap analysis on motor imagery (BCI IV-2a, CR=64): we replace one band of the generalist reconstruction by the reconstruction of a specialist trained only on that band, and report the change in accuracy. A summary Table is below:

Band replacedΔ accuracy (points, 95% CI)
δ-0.07 [-0.60, +0.45]
θ+0.19 [-0.07, +0.48]
α+1.41 [+0.59, +2.27]
β+0.55 [+0.08, +1.01]
γ-low-0.10 [-0.42, +0.22]
γ-high-0.12 [-0.41, +0.15]

In motor imagery, the largest gain comes from α, with smaller effects for the other bands. A codec weighted towards β would therefore be tuned to one task and not to the other. Therefore, we keep band agnostic training for BioSQ and describe this finding in detail in Appendix E. This additional contribution suggested by the Reviewer may certainly improve task performance for practitioners and is of high interest.

Finally, L187 ("iterative quantization benefits lookup-free methods") refers to our finding that adding residual stages improves reconstruction in this family. We now add single-stage FSQ to Appendix G, next to the RFSQ, BSQ and RBSQ already reported there, and rephrase L187 to state this finding and point to Appendix G. The results are in the Table below, comparing single-stage and residual versions under the same nominal budget of 32 bits per frame on CHB-MIT: FSQ against residual FSQ (RFSQ, 4 × 8 bits), and BSQ against RBSQ (8 × 4 bits).

CRFSQ SNR (dB) ↑RFSQ SNR (dB) ↑BSQ SNR (dB) ↑RBSQ SNR (dB) ↑
4×19.322.522.636.8
8×16.217.818.625.4
16×12.713.514.618.5
64×6.37.69.310.5

Adding residual stages improves reconstruction in both pairs: from FSQ to RFSQ (+0.8 to +3.2 dB) and from BSQ to RBSQ (+1.2 to +14.2 dB). The gain is much larger for BSQ, and RBSQ is the best quantizer at every compression ratio. As can be seen, RBSQ not only improves on BSQ in the key biologically relevant frequencies as mentioned before, but also achieves broadly superior performance in classical signal fidelity metrics.

See the rest of the response in the following comments.

Continuation of the previous comment.

W3. Manuscript clarifications

We thank the Reviewer for helping us improve our manuscript. We now address each and update the manuscript accordingly:

  1. In Section 5.1 we have shown that global PRD and downstream performance are not necessarily coupled, and this is the case also for BioSQ and BrainCodec at low compression ratios. BrainCodec has a lower PRD on SWEC-ETHZ and MIT-BIH at 4×, yet BioSQ matches or outperforms it downstream on every dataset (Section 5.4). We clarify this aspect in the Abstract.
  2. The Abstract and Section 6 refer to the same results in Figure 5, but the Abstract was inconsistent. The gains at 128× are 12% (seizure detection F1), 2% (motor-imagery accuracy) and 22% (QRS detection sensitivity), so the correct “up to” value is 22%. We correct the Abstract and the Introduction to “up to 22%” and state that this is a relative gain at 128× compression, where reconstruction fidelity matters most.
  3. The Reviewer is correct that a perfectly uniform distribution leaves no room for arithmetic coding. The regularization promotes balanced bit usage, while arithmetic coding exploits the redundancy that remains, so the two mechanisms are complementary. This is why the gain of arithmetic coding is small where the symbol distribution is close to uniform, on CHB-MIT (3.77–3.96 bits per symbol) and on SWEC-ETHZ from 16×, and large on ECG, where symbol usage is strongly non-uniform and the effective compression ratio increases by up to 1.9× (Appendix C). We now rephrase the Abstract accordingly.
  4. The Introduction raises the spectral effect of strided and transposed convolutions as an open question, and our exploration answers it for biosignals. First, without quantization the same convolutional encoder-decoder reconstructs CHB-MIT with a PRD below 0.11% up to 16× (Appendix E), so these operations are not the main source of error. Second, the Vocos variant (Section 5.2), which predicts the STFT and reconstructs the signal with an inverse transform instead of transposed convolutions, underperforms across nearly all bands and compression ratios. The WavTokenizer decoder improves fidelity, but with 17× more parameters and 9× lower throughput. We therefore keep the convolutional design as the best accuracy-cost trade-off, and now state this explicitly in the Introduction.

Q1. RVQ vs RBSQ

The comparison of RBSQ with 4, 8 and 16 stages on CHB-MIT, at the same budget of 32 bits per frame, is in Appendix F (Figure 13). The 4-stage variant uses 4 stages of 8 bits. The number of stages matters mainly at CR 4×, where the 8-stage configuration is clearly better. The differences become small from CR 16×. Following the Reviewer's suggestion, we now train RVQ with 8 stages of 16 codes (32 bits per frame), with the same recipe and test subjects. The Table below shows the median PRD over test subjects, as in Figure 4.

CRRVQ, 4 stages × 256 codes (paper)RVQ, 8 stages × 16 codesRBSQ, 4 stages × 8 bitsRBSQ, 8 stages × 4 bits (paper)
4×6.267.923.001.20
16×15.0414.4912.1310.28
64×33.5834.2329.7830.34

RVQ with 8 stages does not improve over RVQ with 4 stages: it is close at CR 16× and 64×, and worse at CR 4×. RBSQ with 4 stages of 8 bits, matching the stage structure of the 4-stage RVQ, is already below both RVQ variants at every CR, and it is close to the 8-stage RBSQ from CR 16×. The number of stages therefore does not explain the advantage of RBSQ. We add this result to Appendix F.

Q2. RBSQ at nominal rate

Following the Reviewer's suggestion, we now ablate arithmetic coding from BioSQ and compare it with BrainCodec. We add the full results in Appendix C and present a summary Table here for ease of reading:

DatasetCodec4× ↓16× ↓64× ↓128× ↓
CHB-MITBrainCodec6.2615.0433.5846.77
CHB-MITBioSQ1.2010.2830.3443.53
BCI IV-2aBrainCodec2.5119.8642.0563.05
BCI IV-2aBioSQ1.8219.0640.3057.44
SWEC-ETHZBrainCodec0.784.2016.4930.93
SWEC-ETHZBioSQ1.394.2515.9927.81
MIT-BIHBrainCodec0.180.691.983.31
MIT-BIHBioSQ0.360.551.593.17

As can be seen, BioSQ without arithmetic coding still consistently achieves lower PRD than BrainCodec. On the two EEG datasets, CHB-MIT and BCI IV-2a, BioSQ is better at every compression ratio. On SWEC-ETHZ and MIT-BIH, BioSQ is better from 16x on (on SWEC-ETHZ it is within 0.05 points at 16x, and markedly better afterwards). We report the subject-level statistics of these comparisons in our answer to Reviewer KF24 (W4).

See the rest of the response in the following comment.

Q3. Spherical constraints

The residuals and the entropy term are indeed not normalized. In RBSQ, vkv_k is normalized only before binarization (Eq. 1). Since the sign does not depend on the norm, the forward pass is the same with or without the normalization. The constraint therefore acts only on the straight-through gradient, from which it removes the component along vkv_k. Following the Reviewer's suggestion, we now train RBSQ without the normalization in the backward pass, passing the gradient directly to vkv_k. Both variants use the same training and evaluation protocol, with the same data, hyperparameters, seed and test subjects. The Table below shows the median PRD over test subjects, where each entry is a single run:

CRRBSQ (spherical)RBSQ without normalization
4×1.20100.0
16×10.28100.0
64×30.34100.0

Without the normalization, training does not converge: the PRD stays at 100%, so the reconstruction carries no signal. We add this result to Appendix F.

Q4. Entropy and balance terms

The Reviewer is indeed correct: the quantized output depends only on the sign of WinrW_in r, so scaling WinW_in does not change it. To check that our trained models do not exploit this, we now pass held-out CHB-MIT test windows through the final RBSQ models and measure, in each of the 8 stages: (i) the fraction of bits that change if the projection or the residual is normalized, (ii) the norm of v=Winrv = W_in r and the fraction of coordinates with ∣vi∣>4|v_i| > 4 (σ(4) ≈ 0.98, so the bit is almost certain), and (iii) the mean binary entropy H(σ(vi))H(\sigma(v_i)) of Eq. 2, which is maximal (ln⁡2≈0.693\ln 2 \approx 0.693) when the bits are not confident. If WinW_{in} had been inflated, projections would be large, many coordinates would exceed 4, and the entropy would be near zero. We observe the opposite: the entropy is close to its maximum, so the term is not being exploited.

CRBits changed by normalization90th percentile of |W_in r| (max over stages)Fraction of |v_i| > 4Entropy per bit (max = ln 2 ≈ 0.693)
4×2 / 11,796,4800.6100.691
16×0 / 2,949,1200.6500.691
64×0 / 737,2801.6500.686
128×0 / 368,6401.6200.686

The observation on symbol-level collapse is also correct: balancing each bit does not prevent the 16 symbols of a stage from concentrating, since bits can be balanced individually and still be correlated. We therefore measure symbol usage directly (Appendix C, Figure 9), with the perplexity ratio 2H/162^H/16 and the fraction of unused symbols, averaged over the 8 stages. On CHB-MIT the usage is close to uniform (0.89, 0.97 and 0.96 at 4×, 16× and 64×) with no unused symbols. On SWEC-ETHZ it is less uniform at 4× (0.64, with 12% unused) and close to uniform at 16× and 64× (0.94 and 0.90). On MIT-BIH the concentration is strongest: 0.32 at 4× and 0.58 at 64×, with 16% to 27% of the symbols unused. Symbol-level concentration is therefore not prevented in general: it appears at low compression ratios and on ECG, and it is absent on EEG and iEEG at higher compression ratios. Where it appears, arithmetic coding recovers the redundancy (W3.3; up to 1.9× on ECG). We now state this explicitly in Section 3.

Q5. Band-weighted losses

We now directly evaluate a frequency-based training regime, which we describe in detail in W1/2 and Appendix H. Overall, this new contribution suggested by the Reviewer is interesting, but deteriorates BioSQ's generality on downstream tasks.

We greatly appreciate the time and effort dedicated by the Reviewer to improve our work. As the discussion period is coming to an end, we kindly ask the Reviewer to acknowledge whether we have successfully addressed their outstanding concerns. We remain available to further discuss any remaining question or weakness in our work.

Reviewer_T4smOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies neural lossy compression of biosignals. The proposed codec, BioSQ, is built on an SEANet-style convolutional encoder/decoder architecture with Residual Binary Spherical Quantization (RBSQ), which is comparable to RVQ-based BrainCodec. An arithmetic coder is further applied to decrease the total bitrate.

Critical strengths
  1. Systematic component comparison: The paper compares a broad set of quantizers and backbones.
  2. Evaluating on three modalities and four datasets with task-based metrics (such as seizure and QRS detection, and motor imagery) shows the impact of neural compression on downstream tasks.
Critical weaknesses
  1. Arithmetic coding is a lossless post-coder, and it could be used in any neural codec. But in experimental comparisons, the compression ratio results in Figure 4 seem to be not so fair, since the AC isn't applied to other neural codecs.

  2. Limited methodological novelty. The claimed 'frequency-aware' isn't reflected in the method, since the proposed RBSQ is a composition of BSQ with residual quantization, analogous to existing residual FSQ, residual LFQ, residual VQ variants.

Questions for authors
  1. The codec is trained without adversarial training, is there any detailed comparison about the pro/cons of it?
  2. Unlike the audio/video codec, which is more for real-time transmission, I'm not so familiar with the main purpose of biosignal compression.
3 comments

We thank the Reviewer for highlighting the thoroughness of our evaluations and systematic analysis of all the components of BioSQ.

Following the Reviewer's suggestions, we now compare BioSQ and BrainCodec at the same nominal rate, and clarify the practical purpose of biosignal compression and the novelty of our frequency-based analysis. We address each point in detail below.

W1. Arithmetic coding

The Reviewer is indeed correct that arithmetic coding can in principle be applied to any compressor. This compelling property is one of the reasons we implemented it in our BioSQ pipeline, finding notable advantage in its use. Following the Reviewer's suggestion, we now ablate this component of BioSQ and compare it with BrainCodec. We add the full results in Appendix C and present a summary Table here for ease of reading:

DatasetCodec4× ↓16× ↓64× ↓128× ↓
CHB-MITBrainCodec6.2615.0433.5846.77
CHB-MITBioSQ1.2010.2830.3443.53
BCI IV-2aBrainCodec2.5119.8642.0563.05
BCI IV-2aBioSQ1.8219.0640.3057.44
SWEC-ETHZBrainCodec0.784.2016.4930.93
SWEC-ETHZBioSQ1.394.2515.9927.81
MIT-BIHBrainCodec0.180.691.983.31
MIT-BIHBioSQ0.360.551.593.17

As can be seen, BioSQ without arithmetic coding still consistently achieves lower PRD than BrainCodec. On the two EEG datasets, CHB-MIT and BCI IV-2a, BioSQ is better at every compression ratio. On SWEC-ETHZ and MIT-BIH, BioSQ is better from 16x on (on SWEC-ETHZ it is within 0.05 points at 16x, and markedly better afterwards). We report the subject-level statistics of these comparisons in our answer to Reviewer KF24 (W4).

Moreover, we already quantify in Appendix C (Figure 10) how much arithmetic coding would save on BrainCodec's RVQ. On EEG and iEEG, it saves 3–12% of the bits for RVQ and 1–18% for RBSQ, so arithmetic coding does not systematically favour either quantizer. The full BioSQ, with arithmetic coding, remains the best choice overall.

W2. Methodological novelty

We thank the Reviewer for raising this point, which allows us to better clarify the novelty of BioSQ and especially of its frequency-based analysis. As described in the manuscript, the entire development pipeline for BioSQ was guided by the evaluation on biologically relevant frequency bands rather than on the global reconstruction quality, with the end goal of improving reconstruction of real-world morphological features (see Appendix D) and classification performance.

This frequency-aware setup directly allowed us to identify highly relevant and novel effects, such as the decoupling between global reconstruction quality (PRD) and downstream classification accuracy (see Section 5.1). We now better highlight the value of this contribution in the Introduction and in Section 5.1. The frequency-based analysis also guided our choice of quantizer, which settled on an augmented version of BSQ, RBSQ, which achieved the best reconstruction fidelity at lower, more biologically relevant frequencies, in exchange for lower fidelity at higher frequencies (see Figure 3 at CR=64).

Following the Reviewer's suggestion, we now directly manipulate the band fidelity of BioSQ during training by adding a band-restricted spectral term to the training loss with weight λB\lambda_B (full details in Appendix H). As the gradient balancer rescales the gradient of each term at the output, the band term receives a share λB/(2+λB)\lambda_B/(2+\lambda_B) of the gradient norm: 33% for λB=1\lambda_B=1 and 50% for λB=2\lambda_B=2. We report the results for β and γ-low on CHB-MIT at CR=16 below:

β PRD (13-30 Hz)γ-low PRD (30-70 Hz)γ-high PRD (70-120 Hz)Median PRD
Baseline (λB=0\lambda_B=0)36.348.651.410.28
β band, λB=1\lambda_B=116.356.966.911.09
β band, λB=2\lambda_B=214.157.977.611.44
γ-low band, λB=1\lambda_B=145.342.778.911.03
γ-low band, λB=2\lambda_B=260.643.097.312.31

As expected, differential weighting helps the targeted frequency band at the cost of the others. Such a solution may be attractive for a task-specific compressor, but the band worth favouring depends on the task. Indeed, in seizure detection (Fig. 2), the largest gain comes from β. We now repeat the same band-swap analysis on motor imagery (BCI IV-2a, CR=64): we replace one band of the generalist reconstruction by the reconstruction of a specialist trained only on that band, and report the change in accuracy. A summary Table is below:

Band replacedΔ accuracy (points, 95% CI)
δ-0.07 [-0.60, +0.45]
θ+0.19 [-0.07, +0.48]
α+1.41 [+0.59, +2.27]
β+0.55 [+0.08, +1.01]
γ-low-0.10 [-0.42, +0.22]
γ-high-0.12 [-0.41, +0.15]

See next comment

Continuation of previous comment

In motor imagery, the largest gain comes from α, with smaller effects for the other bands. A codec weighted towards β would therefore be tuned to one task and not to the other. Therefore, we keep band agnostic training for BioSQ and describe this finding in detail in Appendix E. This additional contribution suggested by the Reviewer may certainly improve task performance for practitioners and is of high interest.

Lastly, the Reviewer is correct that residual quantization is well known, and has been combined with FSQ and LFQ. To our knowledge, we are the first to study a residual extension of BSQ for biosignal compression, and to find that it notably improves fidelity. We systematically evaluate RVQ, ERVQ, SNAC, BSQ, RFSQ and RBSQ, and RBSQ has the lowest mean band PRD at every compression ratio (Appendix G). Notably, the residual extension is not equally beneficial for every lookup-free quantizer: RFSQ improves on BSQ only in the lower bands, and is worse overall from 16× (Appendix G). To further illustrate the benefits of this iterative quantization, the Table below compares single-stage BSQ (32 bits) with RBSQ (8 × 4 bits) on CHB-MIT under the same nominal bit budget:

CRBSQ SNR (dB) ↑RBSQ SNR (dB) ↑Improvement (dB) ↑
4×22.636.8+14.2
8×18.625.4+6.8
16×14.618.5+3.9
64×9.310.5+1.2

As can be seen, RBSQ not only improves on BSQ in the key biologically relevant frequencies as mentioned before, but also achieves broadly superior performance in classical signal fidelity metrics.

Q1. Adversarial training

While adversarial training has shown positive effects in reconstructing audio and other modalities [1], the same has not been found for biosignals [2]. In particular, previous work has shown that GANs only mildly improve PRD at the cost of increased spurious high-frequency activity, which has been rated negatively by expert clinicians. Moreover, adversarial training introduces additional computational requirements, such as a second network, a second optimizer, and extra adversarial and feature-matching losses, each with its own weight, which are not justified by the performance. Therefore, following the established literature, we do not train BioSQ adversarially, and indeed our results outperform the GAN SoTA at no additional cost (as mentioned by Reviewer KF24, Strength 5). We now update Section 4 to clarify this aspect.

Q2. Practical purpose

There are a multitude of applications for biosignal compression, either online or static, in real-world environments such as hospitals and industry institutions. We focus here on two main ones: reducing wireless transmission costs and storage requirements for long-term physiological monitoring.

In wearable and implantable devices, wireless communication can consume orders of magnitude more power than sensing and computation [3]. Compression can therefore improve energy efficiency, as demonstrated in wireless EEG systems [4], while also enabling high-resolution, multichannel streaming over bandwidth-limited wireless protocols [5].

Compression also facilitates long-term storage of physiological recordings, enabling retrospective and longitudinal analyses to monitor disease-related changes and treatment response [6,7]. We now add this motivation to the Introduction.

[1] Défossez, Alexandre, et al. "High Fidelity Neural Audio Compression." Transactions on Machine Learning Research (2023).

[2] Carzaniga, Francesco, et al. "The Case for Cleaner Biosignals: High-fidelity Neural Compressor Enables Transfer from Cleaner iEEG to Noisier EEG." ICLR (2025).

[3] Chatterjee, Baibhab, et al. "Bioelectronic Sensor Nodes for the Internet of Bodies." Annual Review of Biomedical Engineering (2023).

[4] Dufort y Álvarez, Guillermo, et al. "Wireless EEG System Achieving High Throughput and Reduced Energy Consumption Through Lossless and Near-Lossless Compression." IEEE Transactions on Biomedical Circuits and Systems (2018).

[5] Nguyen, Anh Tuan, et al. "Lossless Compression Methods for Wireless, Real-Time Streaming of Neural Data From Neural Interface Devices." IEEE Sensors Journal (2026).

[6] Gollwitzer, Stephanie, et al. "The long-term course of temporal lobe epilepsy: From unilateral to bilateral interictal epileptiform discharges in repeated video-EEG monitorings." Epilepsy & Behavior (2017).

[7] Rocamora, Rodrigo, et al. "The spectrum of indications for ultralong-term EEG monitoring." Seizure (2024).

We greatly appreciate the time and effort dedicated by the Reviewer to improve our work. As the discussion period is coming to an end, we kindly ask the Reviewer to acknowledge whether we have successfully addressed their outstanding concerns. We remain available to further discuss any remaining question or weakness in our work.

Comments on the paper

2 comments

We are deeply grateful to all Reviewers for their dedicated service and for providing thoughtful and interesting feedback. We have taken all the reviews to heart and are updating the manuscript to reflect all the changes made following the Reviewers' suggestions, with changes highlighted in blue.

In addition to the individual responses, we wish to highlight here the main improvements to our work.

Most importantly, we now show directly that the frequency-aware analysis can be used to control the codec: weighting a frequency band during training improves that band at the cost of the others, with the band worth favouring depending on the task. This can be used to create task-specific compressors, while we keep the general BioSQ band-agnostic.

  1. Frequency-aware pipeline and training. The entire development pipeline for BioSQ was guided by the frequency-aware evaluation, with the end goal of improving reconstruction of real-world morphological features and classification performance. This frequency-aware setup directly allowed us to identify relevant and novel effects, such as the decoupling between global reconstruction quality and downstream classification accuracy. As suggested by Reviewers T4sm, cSvF and KF24, we now directly condition the training of BioSQ based on a specific frequency band. On CHB-MIT at CR=16, targeting β reduces its PRD from 36.3 to 14.1, and targeting γ-low reduces its PRD from 48.6 to 42.7, in both cases at the cost of the other bands:
Trainingβ PRD (13-30 Hz)γ-low PRD (30-70 Hz)Median PRD
Baseline36.348.610.28
β band, λB=2\lambda_B=214.157.911.44
γ-low band, λB=1\lambda_B=145.342.711.03

We also repeat the band-swap analysis of Figure 2 on motor imagery. Here, the band of interest is α, whose replacement by a specialist raises accuracy by 1.4 points (95% CI [+0.6, +2.3]), whereas seizure detection is driven by β. Since the relevant band changes with the task, a fixed frequency weighting would not transfer across tasks, and we keep band-agnostic training for general BioSQ. This discussion and results are added to Appendices H and E.

  1. Residual extension of BSQ. As suggested by Reviewers T4sm and cSvF, we compare single-stage and residual versions of the lookup-free quantizers under the same budget of 32 bits per frame on CHB-MIT (Reviewer cSvF, W1/2). Adding residual stages improves reconstruction in both pairs (FSQ to RFSQ, and BSQ to RBSQ), and RBSQ is the best quantizer at every compression ratio. We also ablate the number of stages with RVQ and RBSQ at the same budget (Reviewer cSvF, Q1), and the spherical normalization (Q3).

  2. Comparison with BrainCodec at the same nominal rate. As suggested by Reviewers T4sm and cSvF, we ablate arithmetic coding from BioSQ (Reviewer T4sm, W1; Reviewer cSvF, Q2). Without arithmetic coding, BioSQ still achieves lower PRD than BrainCodec in 13 of the 16 tested settings: it is better at every ratio on CHB-MIT and BCI IV-2a, from 64× on SWEC-ETHZ and from 16× on MIT-BIH. We also quantify the symbol usage of RBSQ and the bits saved by arithmetic coding on both quantizers (Reviewer cSvF, Q4).

  3. Statistical reporting. As suggested by Reviewer KF24 (W4), we report subject-level statistics with 95% intervals and paired tests. BioSQ is significantly better than BrainCodec on CHB-MIT at all compression ratios and on SWEC-ETHZ from 64×, BrainCodec is better on SWEC-ETHZ at 4×, and on BCI IV-2a the two are equivalent up to 16× and BioSQ is better at 128× (+1.6 accuracy points, 95% CI [+0.7, +2.5]). Results over multiple random seeds are running and will be added.

  4. Positioning with respect to recent work. As suggested by Reviewer KF24 (W2), we now discuss NeuroRVQ, PhysioOmni and Hypnos, and compare their key design choices with BioSQ in a table. We also evaluate the released tokenizers of NeuroRVQ and Hypnos on EEG, and of NeuroRVQ on ECG, on the same windows as BioSQ. In the 0.5-45 Hz band, NeuroRVQ needs 416 bit/s to reach the PRD that BioSQ reaches at 128 bit/s, and on ECG it needs 2,080 bit/s for the PRD of BioSQ at 128 bit/s. We read these results as reference points, since these models are tokenizers for foundation models and are not designed for fidelity at a fixed bit rate. We also add EMG as a fourth modality (Appendix K), where BioSQ is better than BrainCodec from 16× on.

We have reworked all relevant aspects of the manuscript as suggested by the Reviewers, and are convinced that our work is now better and clearer thanks to all the feedback. We remain available to address any further concerns and are looking forward to engage in interesting discussions with all the Reviewers.

To all Reviewers, Area Chairs, Senior Area Chairs, and Program Chairs,

we are deeply grateful to all Reviewers for the interesting and thoughtful reviews. We are glad that Reviewer cSvF found that our rebuttal clarified most of their points, and are convinced that our final clarification addresses the two remaining.

Regrettably, we were unable to connect with Reviewer T4sm and Reviewer KF24 for further discussion. Nonetheless, we found their initial assessment interesting and thought-provoking, and provided an in-depth analysis and response to each of the points raised, in particular:

  • Extended frequency-aware pipeline. While the frequency-aware analysis and evaluation remains an integral contribution of our work, we now extend it to include frequency-aware training as well. We find that band-specific BioSQ-α\alpha (and β\beta, γ\gamma, etc...) provide the strongest task-specific performance at the cost of general capabilities. Therefore, the two approaches are complementary and highlighted as such in the manuscript.
  • Importance of RBSQ. In addition to the original 5 competing quantizers, we now compare single-stage and residual versions of the lookup-free quantizers: RBSQ remains the highest performing quantizer at every compression ratio.
  • Stronger statistical validation. We extend our evaluation to report subject-level statistics with paired tests, and find that BioSQ is better than the baseline BrainCodec on all datasets, significantly so as the compression ratio raises. Moreover, BioSQ remains consistenly superior even with no arithmetic coding.

We thank once again the Reviewers for the productive discussion, and believe that our work has been much improved during this rebuttal period. We are confident that our responses satisfactorily address all concerns highlighted by all Reviewers.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.