pith. sign in

arxiv: 2606.11970 · v1 · pith:XIWKAM4Bnew · submitted 2026-06-10 · 📡 eess.SP

Low-Density EEG for Seizure Detection: Evaluating CNN-RNN Architectures on a Behind-the-Ear Montage Setup

Annika Stiehl , Patrick Wingert , Nicolas Weeger , Nicole Ille , Christian Uhl , Stefan Gei{\ss}els\"oder This is my paper

Pith reviewed 2026-06-27 08:49 UTC · model grok-4.3

classification 📡 eess.SP
keywords seizure detectionlow-density EEGbehind-the-ear montageCNN-RNN hybridwearable EEGepilepsy monitoringtemporal spectral fusion
0
0 comments X

The pith

A hybrid CNN-Merged model detects seizures from simulated behind-the-ear EEG at 85.89 percent ROC AUC.

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

The paper tests multiple CNN-RNN hybrid architectures on a low-density behind-the-ear electrode setup to determine whether they can compensate for lost spatial information and still detect seizures reliably. It reports that one architecture combining temporal and spectral features reaches 85.89 percent ROC AUC and 79.11 percent balanced accuracy on held-out data while staying consistent across different reference choices. The work addresses the practical limits of wearable EEG, such as lower signal quality and fewer channels, by showing that performance can approach what full-scalp recordings achieve. If correct, the result indicates that automated detection becomes feasible in resource-limited wearable devices without requiring dense electrode arrays.

Core claim

The CNN-Merged model integrates temporal and spectral feature representations from CNN-RNN variants and delivers the highest performance on the behind-the-ear montage, attaining a ROC AUC of 85.89 percent and balanced accuracy of 79.11 percent on the held-out test set while remaining robust across reference montages and reducing the performance difference relative to conventional full-scalp setups.

What carries the argument

The CNN-Merged architecture, which fuses temporal and spectral feature streams to offset the spatial information loss in reduced electrode counts.

If this is right

  • Hybrid temporal-spectral models can narrow the accuracy gap between full-scalp and low-density wearable EEG for seizure detection.
  • The architecture maintains performance stability when the reference montage changes.
  • Automated detection becomes viable for patient-independent use in settings with limited electrode coverage.
  • Such models could support continuous monitoring that reduces clinician review time for seizure diaries.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Real behind-the-ear hardware recordings would provide the decisive next test of the simulation-based results.
  • The approach opens a path to compact, always-on wearable monitors that do not require hospital-grade electrode placement.
  • Further work could examine whether the same fusion strategy improves detection latency in streaming wearable data.

Load-bearing premise

Performance measured on signals simulated from full-scalp recordings will carry over to actual signals recorded directly by behind-the-ear wearable sensors.

What would settle it

Train and evaluate the same model on a dataset of real behind-the-ear EEG recordings from patients and check whether the ROC AUC remains near 85.89 percent.

read the original abstract

Epilepsy affects over 50 million individuals globally, underscoring the need for automated seizure detection systems that can alleviate clinicians workload and enhance the accuracy of patient seizure diaries. In wearable EEG applications, however, reliable detection remains challenging due to the limited spatial resolution of low-density electrode configurations, reduced signal-to-noise ratios, and the scarcity of diverse, publicly available training datasets. This study investigates the efficacy of hybrid deep learning architectures for automated seizure detection using a simulated behind-the-ear montage derived from the Temple University Seizure Corpus (TUSZ, v2.0.3). We conduct a systematic comparison of several CNN-RNN models, including LSTM- and GRU-based variants, across multiple EEG montages to evaluate their capacity to compensate for the loss of spatial information inherent to reduced electrode configurations. The proposed CNN-Merged model, which integrates temporal and spectral feature representations, demonstrates superior performance, achieving a ROC AUC of 85.89% and a balanced accuracy of 79.11% on the held-out test set. Furthermore, the model exhibits strong robustness across different reference montages, effectively bridging the performance gap between conventional full-scalp recordings and resource-constrained wearable systems. These findings substantiate the potential of hybrid deep learning models as a promising avenue toward robust, patient-independent seizure detection in low-density EEG applications.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript evaluates hybrid CNN-RNN architectures, including LSTM/GRU variants and a proposed CNN-Merged model that combines temporal and spectral features, for seizure detection on a behind-the-ear montage simulated by channel selection from the full-scalp TUSZ v2.0.3 corpus. It reports that the CNN-Merged model achieves ROC AUC 85.89% and balanced accuracy 79.11% on a held-out test set and claims robustness across reference montages that effectively bridges the performance gap between conventional full-scalp EEG and resource-constrained wearable systems.

Significance. If the simulation faithfully reproduces real behind-the-ear signal statistics, the systematic architecture comparison would provide evidence that hybrid models can compensate for reduced spatial resolution in low-density EEG, supporting wearable seizure detection. The explicit performance numbers on a held-out set and the cross-montage evaluation are concrete strengths.

major comments (2)
  1. [Abstract and Methods (montage construction)] Abstract and the montage-construction description in the methods: the behind-the-ear montage is obtained solely by selecting a subset of channels from full-scalp TUSZ recordings. No experiments compare the same model on simultaneously recorded full-scalp versus actual behind-the-ear electrode data, nor do they inject wearable-specific artifacts (motion, electrode drift, lower input impedance). This untested assumption is load-bearing for the central claim that the model bridges the performance gap to resource-constrained wearable systems.
  2. [Results] Results section reporting the 85.89% ROC AUC and 79.11% balanced accuracy: the metrics are presented without accompanying details on data preprocessing, exact simulation procedure, class-imbalance handling, or statistical significance testing. These omissions prevent independent verification of the superiority claim for the CNN-Merged model over the other CNN-RNN variants.
minor comments (1)
  1. [Abstract] The abstract and introduction could more explicitly qualify all performance claims as applying to a simulated rather than physically recorded behind-the-ear montage.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript evaluating hybrid CNN-RNN architectures for seizure detection on simulated behind-the-ear EEG montages from the TUSZ corpus. We address each major comment point by point below.

read point-by-point responses

  1. Referee: [Abstract and Methods (montage construction)] Abstract and the montage-construction description in the methods: the behind-the-ear montage is obtained solely by selecting a subset of channels from full-scalp TUSZ recordings. No experiments compare the same model on simultaneously recorded full-scalp versus actual behind-the-ear electrode data, nor do they inject wearable-specific artifacts (motion, electrode drift, lower input impedance). This untested assumption is load-bearing for the central claim that the model bridges the performance gap to resource-constrained wearable systems.

    Authors: We acknowledge that the behind-the-ear montage is constructed via channel selection from the full-scalp TUSZ v2.0.3 recordings, as no public seizure-annotated dataset provides simultaneous full-scalp and behind-the-ear data. This simulation permits direct, controlled comparison of the same models and subjects across montages. We will revise the abstract and methods to explicitly detail the channel selection procedure and add a dedicated limitations paragraph discussing unmodeled wearable artifacts (e.g., motion, impedance differences). The cross-montage robustness results nevertheless demonstrate that hybrid architectures can mitigate spatial information loss, supporting the potential applicability to resource-constrained systems. revision: partial

  2. Referee: [Results] Results section reporting the 85.89% ROC AUC and 79.11% balanced accuracy: the metrics are presented without accompanying details on data preprocessing, exact simulation procedure, class-imbalance handling, or statistical significance testing. These omissions prevent independent verification of the superiority claim for the CNN-Merged model over the other CNN-RNN variants.

    Authors: Preprocessing (0.5–50 Hz bandpass filtering, z-score normalization per channel) and the montage simulation (selection of behind-the-ear channels from the 10-20 system) are described in the Methods. Class imbalance is mitigated via the balanced accuracy metric and stratified train/validation/test splits. We will add a new subsection to Results reporting statistical comparisons (bootstrap 95% CI and paired tests) between the CNN-Merged model and baselines to enable verification of the reported superiority on the held-out test set. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical results on held-out test set from standard ML evaluation

full rationale

The paper reports ROC AUC and balanced accuracy as direct empirical measurements from training CNN-RNN variants on a held-out portion of the TUSZ-derived simulated montage data. No equations, parameters, or claims reduce by construction to fitted inputs (e.g., no prediction of a quantity that was itself used to fit the model). No self-citations, uniqueness theorems, or ansatzes are invoked as load-bearing steps. The simulation via channel selection is an explicit methodological choice whose validity is an external assumption, not a definitional loop. This matches the default case of a self-contained empirical study with no circular reduction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that the simulated montage faithfully represents real low-density wearable EEG and that the TUSZ corpus is representative for patient-independent evaluation.

axioms (1)
  • domain assumption The simulated behind-the-ear montage derived from full-scalp recordings accurately represents signals from actual wearable behind-the-ear devices
    The study relies on this to claim applicability to resource-constrained wearable systems.

pith-pipeline@v0.9.1-grok · 5797 in / 1175 out tokens · 22410 ms · 2026-06-27T08:49:43.871330+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

18 extracted references

  1. [1]

    Diagnostic yield and limitations of in-hospital documentation in patients with epilepsy,

    A. Schulze-Bonhage, E. Bruno, A. Brandt, A. Shek, P. Viana, M. Heers, E. Martinez-Lizana, D. Altenm ¨uller, M. P. Richardson, and V . San Antonio-Arce, “Diagnostic yield and limitations of in-hospital documentation in patients with epilepsy,”Epilepsia, vol. 64, 2022

    2022

  2. [2]

    Brain epileptic seizure detection using joint cnn and exhaustive feature selection with rnn-blstm classifier,

    C. S. Prasanna, M. Z. U. Rahman, and M. Bayleyegn, “Brain epileptic seizure detection using joint cnn and exhaustive feature selection with rnn-blstm classifier,”IEEE Access, vol. 11, pp. 97 990–98 004, 2023

    2023

  3. [3]

    Seizure detection using ultra-long-term subcutaneous elec- troencephalography: A deep learning cnn-bilstm approach,

    S. Park, J. S. Clark, P. F. Viana, J. Cui, J. Duun-Henriksen, J. Mandrekar, N. Gregg, V . Kremen, G. A. Worrell, M. P. Richardson, and B. H. Brinkmann, “Seizure detection using ultra-long-term subcutaneous elec- troencephalography: A deep learning cnn-bilstm approach,”Epilepsia, Oct. 2025

    2025

  4. [4]

    Lookaroundnet: Extending tempo- ral context with transformers for clinically viable eeg seizure detection,

    P. Sverrisson and S. Gudmundsson, “Lookaroundnet: Extending tempo- ral context with transformers for clinically viable eeg seizure detection,” 2026, preprint

    2026

  5. [5]

    Beniczky, S

    S. Beniczky, S. Wiebe, J. Jeppesen, W. O. Tatum, M. Brazdil, Y . Wang, S. T. Herman, and P. Ryvlin, “Automated seizure detection using wearable devices: A clinical practice guideline of the international league against epilepsy and the international federation of clinical neurophysiology,”Epilepsia, vol. 62, no. 3, pp. 632–646, Mar. 2021

    2021

  6. [6]

    Multimodal detection of typical absence seizures in home environment with wearable electrodes,

    C. Chatzichristos, L. Swinnen, J. Macea, M. Bhagubai, W. Van Paess- chen, and M. De V os, “Multimodal detection of typical absence seizures in home environment with wearable electrodes,”Frontiers in Signal Processing, vol. 2, Oct. 2022

    2022

  7. [7]

    Visual seizure annotation and au- tomated seizure detection using behind-the-ear electroencephalographic channels,

    K. Vandecasteele, T. De Cooman, J. Dan, E. Cleeren, S. Van Huffel, B. Hunyadi, and W. Van Paesschen, “Visual seizure annotation and au- tomated seizure detection using behind-the-ear electroencephalographic channels,”Epilepsia, vol. 61, no. 4, pp. 766–775, Mar. 2020

    2020

  8. [8]

    Enhancing seizure detection accuracy in wearable eeg de- vices using deep learning algorithms,

    M. Alarfaj, M. H. Al-Adhaileh, M. I. Uddin, M. Adnan, and T. H. H. Aldhyani, “Enhancing seizure detection accuracy in wearable eeg de- vices using deep learning algorithms,”Journal of Disability Research, vol. 4, no. 4, 2025

    2025

  9. [9]

    Ai-based decision support for epilepsy monitor- ing,

    M. Claro Bhagubai, “Ai-based decision support for epilepsy monitor- ing,” Ph.D. dissertation, KU Leuven, Arenberg Doctoral School, 2025

    2025

  10. [10]

    Using a standalone ear-eeg device for focal-onset seizure detection,

    M. Joyner, S.-H. Hsu, S. Martin, J. Dwyer, D. F. Chen, R. Sameni, S. H. Waters, K. Borodin, G. D. Clifford, A. I. Levey, J. Hixson, D. Winkel, and J. Berent, “Using a standalone ear-eeg device for focal-onset seizure detection,”Bioelectronic Medicine, vol. 10, no. 1, Feb. 2024

    2024

  11. [11]

    Neurowave-net: Enhancing epileptic seizure detection from eeg brain signals via advanced convolutional and long short-term memory networks,

    M. M. Hassan, R. Haque, S. M. S. Islam, H. Meshref, R. Alroobaea, M. Masud, and A. K. Bairagi, “Neurowave-net: Enhancing epileptic seizure detection from eeg brain signals via advanced convolutional and long short-term memory networks,”AIMS Bioengineering, vol. 11, no. 1, pp. 85–109, 2024

    2024

  12. [12]

    A cnn-lstm hybrid network for automatic seizure detection in eeg signals,

    S. Shanmugam and S. Dharmar, “A cnn-lstm hybrid network for automatic seizure detection in eeg signals,”Neural Computing and Applications, vol. 35, no. 28, pp. 20 605–20 617, Jul. 2023

    2023

  13. [13]

    Seizeit2: Wearable dataset of patients with focal epilepsy,

    M. Bhagubai, C. Chatzichristos, L. Swinnen, J. Macea, J. Zhang, L. Lagae, K. Jansen, A. Schulze-Bonhage, F. Sales, B. Mahler, Y . Weber, W. Van Paesschen, and M. De V os, “Seizeit2: Wearable dataset of patients with focal epilepsy,”Scientific Data, vol. 12, no. 1, Jul. 2025

    2025

  14. [14]

    The temple university hospital seizure detection corpus,

    V . Shah, E. von Weltin, S. Lopez, J. R. McHugh, L. Veloso, M. Gol- mohammadi, I. Obeid, and J. Picone, “The temple university hospital seizure detection corpus,”Frontiers in Neuroinformatics, vol. 12, 2018

    2018

  15. [15]

    Towards automated eeg-based epilepsy detection using deep convolu- tional autoencoders,

    A. Stiehl, N. Weeger, C. Uhl, D. Bechtold, N. Ille, and S. Geißels ¨oder, “Towards automated eeg-based epilepsy detection using deep convolu- tional autoencoders,” in2025 47th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). IEEE, Jul. 2025, pp. 1–6

    2025

  16. [16]

    Pyseizure: A single machine learning classifier framework to detect seizures in diverse datasets,

    B. Chybowski, S. Abdullateef, H. Haule, A. Gonzalez-Sulser, and J. Escudero, “Pyseizure: A single machine learning classifier framework to detect seizures in diverse datasets,” 2025, preprint

    2025

  17. [17]

    Eegwavenet: Multiscale cnn- based spatiotemporal feature extraction for eeg seizure detection,

    P. Thuwajit, P. Rangpong, P. Sawangjai, P. Autthasan, R. Chaisaen, N. Banluesombatkul, P. Boonchit, N. Tatsaringkansakul, T. Sud- hawiyangkul, and T. Wilaiprasitporn, “Eegwavenet: Multiscale cnn- based spatiotemporal feature extraction for eeg seizure detection,”IEEE Transactions on Industrial Informatics, vol. 18, no. 8, pp. 5547–5557, 2022

    2022

  18. [18]

    Exploring the interpretability of eeg-inception convolutional neural networks for epilepsy prediction,

    G. Zhang, T. Wang, J. Guo, Z. Yang, Y . Wu, and G. Kang, “Exploring the interpretability of eeg-inception convolutional neural networks for epilepsy prediction,” inICASSP 2025 - 2025 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, Apr. 2025, pp. 1–5

    2025