Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

SHAP-AAD: DeepSHAP-Guided Channel Reduction for EEG Auditory Attention Detection

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A Shapley-value attribution method can rank EEG channels by importance, letting a 32-channel system match the accuracy of a 64-channel setup at detecting which speaker a listener attends to.

desk verdict The idea is a reasonable new application of DeepSHAP to EEG channel selection for AAD, but the reported accuracy numbers are not trustworthy because the evaluation leaks information through overlapping windows and test-set channel selection. read the letter →

arxiv 2507.03814 v1 pith:XMN2BY4P submitted 2025-07-04 eess.SP

Rayan Salmi , Guorui Lu , Qinyu Chen This is my paper · ORCID
classification eess.SP
keywords electroencephalographyauditoryattentiondetectionchannelreductionDeepSHAPShapleyvaluestemporalconvolutionalnetworkexplainableAIEEG
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to show that explainable-AI feature attribution can replace expensive 64-channel EEG setups for auditory attention detection, the task of deciding which of two competing speakers a listener is following. It ranks EEG channels with DeepSHAP applied to a CNN trained on alpha-power scalp maps, keeps the top channels, and trains a lightweight temporal convolutional network on the raw signals of those channels. On a public 18-subject dataset, the 32-channel system reaches 79.21% average accuracy against 81.06% for the full 64-channel setup, and some subjects retain satisfactory accuracy with 8 channels. If this holds, hearing-aid-style devices could use far fewer electrodes without a large loss in detection quality.

What carries the argument

The load-bearing object is the DeepSHAP importance map: DeepSHAP combines Shapley-value game theory with DeepLIFT backpropagation to assign each pixel of a 32x32 alpha-power topographic image a contribution score, which is averaged over samples and cross-validation folds and mapped back to electrode locations. These scores produce a global channel ranking from which the top-k electrodes are kept; a two-layer dilated temporal convolutional network then classifies attention from raw waveforms of only those channels. The ranking is the mechanism that turns an explainability output into a channel-selection procedure.

What would settle it

A reader could settle the transfer claim by training the same TCN on the top-32 channels from the DeepSHAP image-CNN ranking, on 32 channels chosen by greedy backward elimination on the TCN itself, and on random 32-channel subsets; the central claim fails if the greedy set does not clearly beat the DeepSHAP set or if random sets match it.

Watch

Extended reading notes

Core claim

The paper's central claim is that DeepSHAP-derived channel importance transfers across representations: a CNN that classifies attention from 32x32 topographic images of alpha-band power assigns per-pixel Shapley values that, averaged across samples and cross-validation folds, rank the 64 scalp channels, and the top-k of that ranking can be fed as raw time series to a compact two-layer dilated TCN that detects attention with accuracy close to the full montage. Quantitatively, the paper reports 81.06% mean accuracy with 64 channels and 79.21% with 32, with further degradation at 16 and 8 channels and notable subject-to-subject variation. The paper argues that moderate reduction to 32 channels preserves most discriminative information and that the TCN avoids the interpolation artifacts image representations would introduce after channel removal.

Load-bearing premise

The channel ranking produced by a model that looks at maps of brain-wave power over the scalp must transfer to a different model that reads the raw waveform of each electrode; if that transfer fails, the reported accuracy of the reduced-channel system is not explained by the selection method.

Editorial extensions

If this is right

  • A 32-channel wearable EEG device could plausibly support auditory attention detection with accuracy within roughly two percentage points of a full 64-channel laboratory setup.
  • Channel sets can be chosen per subject from the global importance map, so individual head geometries or attention-related topographies can guide electrode placement.
  • Model size and computation drop modestly with channel count, from 0.18M parameters and 220 MFLOPs at 64 channels to 0.15M and 183 MFLOPs at 32, making real-time inference easier on embedded hardware.
  • The two-stage design separates selection, image-based CNN, from classification, waveform-based TCN, so either stage could be swapped independently in later systems.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The authors' transfer assumption could be tested directly by computing DeepSHAP on the TCN over raw waveforms instead of on the image CNN; if the rankings disagree, the reported accuracy of the reduced TCN may be an artifact of the particular CNN used for selection.
  • The same selection pipeline could be applied to other EEG tasks, such as sleep staging or motor imagery, wherever a stable topographic summary exists, but its success would hinge on the same cross-representation transfer.
  • A subject-specific rather than global ranking might shrink the required channel count further, since the paper's own data show large inter-subject variability in how much accuracy drops at 8 or 16 channels.
  • The paper does not compare against exhaustive or greedy channel selection on the TCN itself, so the claim that DeepSHAP adds value over random or heuristic reduction is not yet established.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes SHAP-AAD, a two-stage framework for EEG-based auditory attention detection (AAD). In the first stage, a CNN trained on topographic alpha-band power images is interpreted with DeepSHAP to rank the 64 EEG channels by importance, and the top-k channels are selected. In the second stage, a compact temporal convolutional network (TCN) is trained directly on raw EEG waveforms from the selected channels. Experiments on the public DTU dataset report that using 32 channels yields average accuracy of 79.21% versus 81.06% with all 64 channels, and that some subjects retain high accuracy with as few as 8 channels. The central claim is that DeepSHAP-guided channel reduction can match the accuracy of the full-channel setup while reducing sensor count.

Significance. The problem addressed is practically relevant: reducing EEG channel count could improve wearability and comfort of AAD systems. The use of DeepSHAP for explainability-driven channel selection appears novel in this application, and the lightweight TCN architecture is a reasonable choice for deployment. The paper is clearly structured and uses a public dataset, which materially aids reproducibility. However, the empirical evidence is compromised by two independent forms of leakage—overlapping decision windows randomly split across train and test, and SHAP-based channel selection performed on test data—so the reported accuracies are not unbiased estimates of prospective performance. No error bars or statistical significance tests are provided. Consequently, the central claim is not established in the current manuscript; the value of the work hinges on a re-evaluation with a clean experimental protocol.

major comments (3)
  1. [III.A and IV.A] The evaluation protocol leaks temporal information: Section III.A segments the continuous EEG into 10-s windows with 50% overlap, and Section IV.A splits the data into training (80%), validation (10%), and test (10%) using random cross-validation without grouping by trial. As a result, overlapping windows from the same 50-s trial can appear in both training and test folds, and the reported test accuracy is inflated by strong temporal autocorrelation. All accuracy figures, including the headline 79.21% vs. 81.06% comparison in Section IV.C and Fig. 6, are therefore not valid estimates of generalization. A trial-disjoint split (e.g., grouping all windows of a 50-s trial into the same fold) is required before any conclusion about the method's performance can be drawn.
  2. [III.B.3] The channel selection procedure uses the test set: the text states that 'after each cross-validation fold, a 100 test samples of the test set were used to compute the SHAP values,' and these values are averaged across folds to produce the global importance map that determines the top-k channels. Because the test data are used to choose the channels on which the TCN is later trained and evaluated, the evaluation is not independent of the feature-selection step. The reported reduced-channel accuracy is optimistically biased, and this bias is load-bearing for the paper's central claim. The channel ranking must be derived exclusively from training data (e.g., a nested cross-validation loop or a separate selection set).
  3. [III.B vs. III.C] The channel ranking is computed from the alpha-power topographic CNN, but the final classifier is a TCN operating on raw waveforms. The paper asserts that channels important for the image-based CNN are the same channels that maximize the raw-waveform TCN's accuracy, but no evidence is provided for this transfer. Since the two models use fundamentally different input representations, this assumption is load-bearing: without it, the reported performance of the reduced-channel TCN is not explained by the SHAP-based selection. A concrete control experiment, such as comparing the selected channels against random or fixed channel subsets on the TCN, or computing SHAP values on the TCN itself, is needed to validate the transfer.
minor comments (5)
  1. [IV.C] The text says 'five different EEG channel configurations: 64, 60, 48, 32, 16, and 8 channels,' which lists six configurations; moreover, the 60-channel setting is not discussed in the results. Please correct the count and either add or remove the 60-channel configuration.
  2. [III.B.3] The phrase 'a 100 test samples' should be '100 test samples.' Also, it is unclear whether the 200 training samples used as the DeepSHAP background are drawn from each fold's training set or from the full training set, and whether the 100 test samples are the same across folds; please clarify.
  3. [Figures 3 and 4] The SHAP importance maps in Figures 3 and 4 would benefit from axis labels, a colorbar with units, and a clear statement of the subject and windowing parameters used; currently the figures are described only in the caption. In Figure 5, the subject indices on the x-axis should be explicitly defined.
  4. [IV.C and Table I] No comparison with existing channel-selection methods for AAD (e.g., the approaches in Refs. [13]–[15]) is provided. Since the introduction positions the work against these methods, the experimental section should include them as baselines to substantiate the claimed advantage.
  5. [IV.A] Only mean accuracies over 10 folds are reported, with no standard deviations or confidence intervals. Even after fixing the leakage issues, the claim that 32 channels are 'comparable' to 64 channels would be much stronger if accompanied by measures of variability across folds and subjects.

Circularity Check

2 steps flagged · score 6.0 of 10

Test-set SHAP values are used to choose channels before measuring TCN accuracy on the same test folds, making the 32-channel result partially fitted; overlapping windows add further train/test leakage.

  1. fitted input called prediction [Section III.B.3; Section III.C; Section IV.A]
    "After each cross-validation fold, a 100 test samples of the test set were used to compute the SHAP values for all input pixels. The SHAP values across all samples and cross-validation folds were averaged (absolute mean) to generate a global importance map."

    This global importance map selects the top-k channels used to train the TCN (Section III.C). Section IV.A then reports mean test accuracy for the reduced-channel TCN on the same cross-validation folds. The channel configuration whose accuracy is reported was therefore chosen using test-set signals on which that accuracy is measured, and the TCN is evaluated against those same test labels. The 79.21% vs 81.06% comparison is not an unbiased estimate; the 32-channel result is partly fitted to the test data, and averaging SHAP across folds does not remove the contamination because every fold's test samples contribute to the shared channel ranking.

  2. other [Section III.A; Section IV.A]
    "The continuous EEG data was segmented into overlapping decision windows of 10 seconds, with a 50% overlap to increase the number of training samples. Random cross-validation was used to split the data into training (80%), validation (10%), and test (10%) sets."

    The random cross-validation split is applied to overlapping windows without trial-disjoint grouping. Each 50-second trial generates 10-second windows with 50% overlap, so adjacent windows share five seconds of raw EEG and the same attended-speaker label. The same underlying trial can thus appear in both training and test folds, meaning the reported reduced-channel accuracy is partly a measure of the model having already seen near-identical data in training. This contaminates the 32- vs 64-channel comparison independently of the SHAP selection step.

full rationale

The paper has no formal derivation chain and no load-bearing self-citations; the central claim is empirical. The mathematical content of DeepSHAP and the TCN is not circular: the ranking is computed on a separate alpha-power CNN, and the TCN is a distinct architecture on raw waveforms. However, the evaluation protocol contains a genuine circular element: the channel ranking is fitted to test-set samples and the resulting channel configuration is then evaluated on those same test folds, so the headline accuracy is not an independent measure of the selected channels. A second leak comes from random cross-validation on overlapping windows, which allows the same trial to appear in both training and test folds. Together these make the reported 32-channel result partially fitted rather than a clean prospective prediction. Because the central claim depends on this contaminated estimate, the circularity score is 6; the result is not forced by definition or by a self-citation chain, so it does not reach 8 or 10.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new theoretical entities. Its load-bearing assumptions are domain-level (alpha power relevance, DeepSHAP fidelity, transferability of channel rankings across model classes) plus a violated statistical-independence assumption in the evaluation. The central claim depends on these assumptions.

free parameters (4)
  • top-k channel counts = 48, 32, 16, 8
    Hand-selected reduction levels; the choice of which k is 'comparable' is subjective.
  • alpha-band frequency range = 8-14 Hz
    Fixed from domain knowledge; not optimized on data.
  • decision window length and overlap = 10 s, 50% overlap
    Chosen without reported ablation; affects the amount of training data and the leakage risk.
  • SHAP background sample size = 200 training samples
    Arbitrary choice; no sensitivity analysis.
assumptions (4)
  • domain assumption Alpha-band (8-14 Hz) power in EEG contains the discriminative information for auditory attention.
    The channel importance is derived from a CNN trained only on topographic alpha-power maps; if other frequency bands carry the attention signal, the ranking may miss informative channels.
  • domain assumption DeepSHAP values computed on the trained CNN faithfully estimate the true contribution of each EEG channel to the CNN's predictions.
    DeepSHAP is an approximation of Shapley values; its accuracy for this model and data is not validated.
  • ad hoc to paper Channels ranked as important for the alpha-power image CNN are also the channels that maximize the raw-waveform TCN's accuracy.
    This transferability assumption is the paper's weakest premise; the selection model and the final classifier use different input representations, and no evidence is given that the rankings agree.
  • domain assumption Cross-validation folds built from overlapping windows are independent samples.
    The 50% window overlap means adjacent windows share raw EEG data, so random splits leak information between training and testing; the paper does not account for this.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SHAP-AAD: DeepSHAP-Guided Channel Reduction for EEG Auditory Attention Detection." pith.science (2026). https://pith.science/paper/XMN2BY4P

@misc{pith2026250703814,
  author       = {Pith},
  title        = {Pith review of: SHAP-AAD: DeepSHAP-Guided Channel Reduction for EEG Auditory Attention Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XMN2BY4P}},
  note         = {Machine review of arXiv:2507.03814}
}
read the original abstract

Electroencephalography (EEG)-based auditory attention detection (AAD) offers a non-invasive way to enhance hearing aids, but conventional methods rely on too many electrodes, limiting wearability and comfort. This paper presents SHAP-AAD, a two-stage framework that combines DeepSHAP-based channel selection with a lightweight temporal convolutional network (TCN) for efficient AAD using fewer channels.DeepSHAP, an explainable AI technique, is applied to a Convolutional Neural Network (CNN) trained on topographic alpha-power maps to rank channel importance, and the top-k EEG channels are used to train a compact TCN. Experiments on the DTU dataset show that using 32 channels yields comparable accuracy to the full 64-channel setup (79.21% vs. 81.06%) on average. In some cases, even 8 channels can deliver satisfactory accuracy. These results demonstrate the effectiveness of SHAP-AAD in reducing complexity while preserving high detection performance.

Figures

Figures reproduced from arXiv: 2507.03814 by the authors.

Figure 1
Figure 1. The overall workflow of the proposed SHAP-AAD framework. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Topographical map of alpha power distribution across [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Mean SHAP values for each pixel in all the topographic [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Ranked SHAP value of different EEG channels on Subject 1 (10-s window), ranking in descending order. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Subject-wise AAD accuracy across the DTU dataset under different EEG channel reductions (10-s window). [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 6
Figure 6. Figure 6: Average AAD accuracy across the DTU dataset under [PITH_FULL_IMAGE:figures/full_fig_p004_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Efficient EEG Seizure Detection Using INT8 Quantization, Channel Pruning, and Spiking Neural Networks

    eess.SP 2026-07 conditional novelty 4.0 of 10

    On a shared 1D-CNN baseline for CHB-MIT seizure detection, INT8 quantization cut model size from 1.63 to 0.44 MB and latency by 2.8x with preserved AUC, while SNN conversion was 288x slower on CPU.

Reference graph

Works this paper leans on

23 extracted references · 22 canonical work pages · cited by 1 Pith paper

  1. [1]

    The cocktail party problem,

    S. Haykin and Z. Chen, “The cocktail party problem,” Neural compu- tation, vol. 17, no. 9, pp. 1875–1902, 2005

    work page 1902

  2. [2]

    Hearing loss in adults,

    L. L. Cunningham and D. L. Tucci, “Hearing loss in adults,” New England Journal of Medicine , vol. 377, no. 25, pp. 2465–2473, 2017

    work page 2017

  3. [3]

    Sensorineural hearing loss degrades behavioral and physiological measures of human spatial selective auditory attention,

    L. Dai, V . Best, and B. G. Shinn-Cunningham, “Sensorineural hearing loss degrades behavioral and physiological measures of human spatial selective auditory attention,” Proceedings of the National Academy of Sciences, vol. 115, no. 14, pp. E3286–E3295, 2018

    work page 2018

  4. [4]

    Comparison of two-talker attention decoding from eeg with nonlinear neural networks and linear methods,

    G. Ciccarelli, M. Nolan, J. Perricone, P. T. Calamia, S. Haro, J. O’sullivan, N. Mesgarani, T. F. Quatieri, and C. J. Smalt, “Comparison of two-talker attention decoding from eeg with nonlinear neural networks and linear methods,” Scientific reports, vol. 9, no. 1, p. 11538, 2019

    work page 2019

  5. [5]

    Seeing helps hearing: A multi-modal dataset and a mamba- based dual branch parallel network for auditory attention decoding,

    C. Fan, H. Zhang, Q. Ni, J. Zhang, J. Tao, J. Zhou, J. Yi, Z. Lv, and X. Wu, “Seeing helps hearing: A multi-modal dataset and a mamba- based dual branch parallel network for auditory attention decoding,” Information Fusion, p. 102946, 2025

    work page 2025

  6. [6]

    A neural-inspired architec- ture for eeg-based auditory attention detection,

    S. Cai, P. Li, E. Su, Q. Liu, and L. Xie, “A neural-inspired architec- ture for eeg-based auditory attention detection,” IEEE Transactions on Human-Machine Systems, vol. 52, no. 4, pp. 668–676, 2022

    work page 2022

  7. [7]

    Low-power and lightweight spiking transformer for eeg-based auditory attention detection,

    Y . Lan, Y . Wang, Y . Zhang, and H. Zhu, “Low-power and lightweight spiking transformer for eeg-based auditory attention detection,” Neural Networks, vol. 183, p. 106977, 2025

    work page 2025

  8. [8]

    A comprehensive review of eeg-based brain–computer interface paradigms,

    R. Abiri, S. Borhani, E. W. Sellers, Y . Jiang, and X. Zhao, “A comprehensive review of eeg-based brain–computer interface paradigms,” Journal of Neural Engineering, vol. 16, no. 1, p. 011001, jan

    work page

Show all 23 references
  1. [9]

    Eeg-based detection of the locus of auditory attention with convolutional neural networks,

    S. Vandecappelle, L. Deckers, N. Das, A. H. Ansari, A. Bertrand, and T. Francart, “Eeg-based detection of the locus of auditory attention with convolutional neural networks,” Elife, vol. 10, p. e56481, 2021

    2021

  2. [10]

    A gru–cnn model for auditory attention detection using microstate and recurrence quantification analysis,

    M. EskandariNasab, Z. Raeisi, R. A. Lashaki, and H. Najafi, “A gru–cnn model for auditory attention detection using microstate and recurrence quantification analysis,” Scientific Reports, vol. 14, no. 1, p. 8861, 2024

    2024

  3. [11]

    Darnet: Dual attention refinement network with spatiotemporal construction for au- ditory attention detection,

    S. Yan, C. Fan, H. Zhang, X. Yang, J. Tao, and Z. Lv, “Darnet: Dual attention refinement network with spatiotemporal construction for au- ditory attention detection,” Advances in Neural Information Processing Systems, vol. 37, pp. 31 688–31 707, 2024

    2024

  4. [12]

    Eeg-based auditory attention detection via frequency and channel neural attention,

    S. Cai, E. Su, L. Xie, and H. Li, “Eeg-based auditory attention detection via frequency and channel neural attention,” IEEE Transactions on Human-Machine Systems, vol. 52, no. 2, pp. 256–266, 2021

    2021

  5. [13]

    The effect of miniaturization and galvanic separation of eeg sensor devices in an auditory attention detection task,

    A. M. Narayanan and A. Bertrand, “The effect of miniaturization and galvanic separation of eeg sensor devices in an auditory attention detection task,” in 2018 40th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC) . IEEE, 2018, pp. 77–80

    2018

  6. [14]

    Decoding the attended speech stream with multi-channel eeg: implications for online, daily-life applications,

    B. Mirkovic, S. Debener, M. Jaeger, and M. De V os, “Decoding the attended speech stream with multi-channel eeg: implications for online, daily-life applications,” Journal of neural engineering , vol. 12, no. 4, p. 046007, 2015

    2015

  7. [15]

    Analysis of miniaturization effects and channel selection strategies for eeg sensor networks with application to auditory attention detection,

    A. M. Narayanan and A. Bertrand, “Analysis of miniaturization effects and channel selection strategies for eeg sensor networks with application to auditory attention detection,” IEEE Transactions on Biomedical Engineering, vol. 67, no. 1, pp. 234–244, 2019

    2019

  8. [16]

    Explaining models by propagating shapley values of local components,

    H. Chen, S. Lundberg, and S.-I. Lee, “Explaining models by propagating shapley values of local components,” in Explainable AI in Healthcare and Medicine: Building a Culture of Transparency and Accountability . Springer, 2020, pp. 261–270

    2020

  9. [17]

    A value for n-person games,

    L. S. Shapley et al., “A value for n-person games,” 1953

    1953

  10. [18]

    The explanation game: Explaining machine learning models using shapley values,

    L. Merrick and A. Taly, “The explanation game: Explaining machine learning models using shapley values,” in Machine Learning and Knowledge Extraction: 4th IFIP TC 5, TC 12, WG 8.4, WG 8.9, WG 12.9 International Cross-Domain Conference, CD-MAKE 2020, Dublin, Ireland, August 25–...

    2020

  11. [19]

    Learning important features through propagating activation differences,

    A. Shrikumar, P. Greenside, and A. Kundaje, “Learning important features through propagating activation differences,” in International conference on machine learning , 2017, pp. 3145–3153

    2017

  12. [20]

    Eeg and audio dataset for auditory attention decoding,

    S. A. Fuglsang, D. D. Wong, and J. Hjortkjær, “Eeg and audio dataset for auditory attention decoding,” Mar. 2018. [Online]. Available: https://doi.org/10.5281/zenodo.1199011

    2018 doi

  13. [21]

    Noise-robust cortical tracking of attended speech in real-world acoustic scenes,

    S. A. Fuglsang, T. Dau, and J. Hjortkjær, “Noise-robust cortical tracking of attended speech in real-world acoustic scenes,” NeuroImage, vol. 156, pp. 435–444, 2017

    2017

  14. [22]

    A comparison of regularization methods in forward and backward models for auditory attention decoding,

    D. D. Wong, S. A. Fuglsang, J. Hjortkjær, E. Ceolini, M. Slaney, and A. De Cheveigne, “A comparison of regularization methods in forward and backward models for auditory attention decoding,” Frontiers in neuroscience, vol. 12, p. 531, 2018

    2018

  15. [2019]

    Available: https://dx.doi.org/10.1088/1741-2552/aaf12e

    [Online]. Available: https://dx.doi.org/10.1088/1741-2552/aaf12e

    doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.