pith. sign in

arxiv: 2412.17842 · v2 · submitted 2024-12-18 · 📡 eess.SP · cs.LG

Canine EEG Helps Human: Cross-Species and Cross-Modality Epileptic Seizure Detection via Multi-Space Alignment

Z. Wang , S. Li , Dongrui Wu This is my paper

Pith reviewed 2026-05-23 06:55 UTC · model grok-4.3

classification 📡 eess.SP cs.LG
keywords cross-species EEGseizure detectiondomain adaptationepilepsyknowledge distillationcanine EEGmulti-space alignmentEEG-based detection
0
0 comments X

The pith

Cross-species and cross-modality EEG alignment enables seizure detection above 90% AUC with limited target data.

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

The paper establishes that a multi-space alignment framework can transfer knowledge from canine to human EEG and across surface to intracranial recordings for better epileptic seizure detection. This matters because traditional models are limited by scarce labeled data in specific species or modalities, while combining data from dogs and humans could expand training resources. The approach uses domain adaptation and knowledge distillation to align the signals, achieving over 90 percent AUC in cross cases. Experiments on multiple datasets show improvements over within-species models. It also hints at generalizability to other brain-computer interface tasks.

Core claim

By employing deep learning techniques, including domain adaptation and knowledge distillation, the framework aligns cross-species and cross-modality EEG signals to enhance the detection capability beyond traditional within-species and within-modality models. Experiments on multiple surface and intracranial EEG datasets of humans and canines demonstrated substantial improvements in the detection accuracy, achieving over 90% AUC scores for cross-species and cross-modality seizure detection with extremely limited labeled data from the target species/modality. This is the first study that demonstrates the effectiveness of integrating heterogeneous data from different species and modalities to改善e

What carries the argument

Multi-space alignment approach based on domain adaptation and knowledge distillation that aligns EEG signals across species and modalities.

If this is right

  • Detection accuracy improves beyond within-species and within-modality baselines.
  • Over 90% AUC is reached even with extremely limited labeled data from the target domain.
  • The method is the first to integrate heterogeneous data from different species and modalities for EEG seizure detection.
  • The approach may generalize to other brain-computer interface paradigms.
  • Combining data from different species and modalities can increase the amount of training data for large EEG models.

Where Pith is reading between the lines

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

  • Pooling animal and human EEG could help overcome data scarcity for training models on rare neurological events.
  • The same alignment strategy might extend to other cross-species medical signal tasks such as ECG or EMG analysis.
  • If successful on additional datasets, the method could support pre-training human diagnostic tools on abundant animal recordings.
  • It raises the question of whether similar multi-space techniques apply to non-EEG modalities like MEG or fNIRS.

Load-bearing premise

EEG signals recorded from canines and humans share enough common structure across species and recording modalities for domain adaptation and knowledge distillation to produce effective knowledge transfer despite biological and technical differences.

What would settle it

A controlled test showing no AUC improvement or outright performance drop when the alignment framework is applied to new cross-species or cross-modality EEG test sets would falsify the central claim.

Figures

Figures reproduced from arXiv: 2412.17842 by Dongrui Wu, S. Li, Z. Wang.

Figure 1
Figure 1. Figure 1: Evidences for cross-species and cross-modality f [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Ablation study, t-SNE feature visualization, and parameter sensitivity analysis. a, Ablation study on the Kaggle iEEG dataset, including two tasks of canine-to-human and human-to-canine, and two scenarios of unsupervised cross-species transfer and semi-supervised cross-species transfer. b, Feature t-SNE visualizations on the Canine-to-Human cross-species transfer task of Kaggle, CHSZ and NICU datasets. c, … view at source ↗
Figure 3
Figure 3. Figure 3: Effect of Input Space Alignment via the Proposed Re [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
read the original abstract

Epilepsy significantly impacts global health, affecting about 65 million people worldwide, along with various animal species. The diagnostic processes of epilepsy are often hindered by the transient and unpredictable nature of seizures. Here we propose a multi-space alignment approach based on cross-species and cross-modality electroencephalogram (EEG) data to enhance the detection capabilities and understanding of epileptic seizures. By employing deep learning techniques, including domain adaptation and knowledge distillation, our framework aligns cross-species and cross-modality EEG signals to enhance the detection capability beyond traditional within-species and with-modality models. Experiments on multiple surface and intracranial EEG datasets of humans and canines demonstrated substantial improvements in the detection accuracy, achieving over 90% AUC scores for cross-species and cross-modality seizure detection with extremely limited labeled data from the target species/modality. To our knowledge, this is the first study that demonstrates the effectiveness of integrating heterogeneous data from different species and modalities to improve EEG-based seizure detection performance. The approach may also be generalizable to different brain-computer interface paradigms, and suggests the possibility to combine data from different species/modalities to increase the amount of training data for large EEG models.

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

3 major / 2 minor

Summary. The paper proposes a multi-space alignment framework that combines domain adaptation and knowledge distillation to transfer seizure-detection knowledge from canine to human EEG and across surface/intracranial modalities, claiming that this yields >90% AUC on cross-species and cross-modality tasks even with extremely limited target-domain labels and constitutes the first demonstration of heterogeneous species/modality data improving EEG seizure detection.

Significance. If the central claim holds, the result would be significant because it directly addresses data scarcity in clinical EEG by showing that canine recordings can augment human models and that cross-modality alignment is feasible; the work also supplies a concrete, falsifiable test (cross-species AUC with limited labels) that could be replicated on other BCI tasks.

major comments (3)
  1. [Abstract, §4] Abstract and §4 (Experiments): the reported >90% AUC on cross-species/cross-modality tasks is presented without baseline comparisons, statistical tests, or dataset descriptions (number of seizures, subjects, recording lengths), so it is impossible to determine whether the scores reflect genuine transfer or dataset-shift artifacts.
  2. [§3, §4] §3 (Method) and §4: no domain-discrepancy metrics (e.g., MMD or CORAL before/after alignment), no t-SNE or alignment visualizations, and no ablation that removes the multi-space alignment component are provided; without these, the claim that improvements arise from seizure-relevant invariants rather than generic statistics cannot be evaluated.
  3. [§4] §4: the assumption that seizure semiology and EEG signatures survive lissencephalic vs. gyrencephalic cortex differences and scalp vs. iEEG spatial filtering is load-bearing for the central claim yet is not tested by any control experiment that isolates biological vs. technical domain shift.
minor comments (2)
  1. [§3] Notation for the three alignment spaces (species, modality, joint) is introduced without an explicit diagram or equation that shows how the losses are combined.
  2. [Abstract, §5] The statement that the approach 'may also be generalizable to different brain-computer interface paradigms' is unsupported by any experiment outside seizure detection.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the insightful comments. We will revise the manuscript to include the requested details, metrics, visualizations, ablations, and discussions to strengthen the presentation of our results. We respond to each major comment below.

read point-by-point responses

  1. Referee: [Abstract, §4] Abstract and §4 (Experiments): the reported >90% AUC on cross-species/cross-modality tasks is presented without baseline comparisons, statistical tests, or dataset descriptions (number of seizures, subjects, recording lengths), so it is impossible to determine whether the scores reflect genuine transfer or dataset-shift artifacts.

    Authors: We agree that more details are needed to substantiate the claims. In the revised manuscript, we will expand §4 to include full dataset descriptions with the number of seizures, subjects, and recording lengths. We will also add baseline comparisons using standard seizure detection methods and report statistical significance tests for the AUC improvements to confirm they are not due to dataset-shift artifacts. revision: yes

  2. Referee: [§3, §4] §3 (Method) and §4: no domain-discrepancy metrics (e.g., MMD or CORAL before/after alignment), no t-SNE or alignment visualizations, and no ablation that removes the multi-space alignment component are provided; without these, the claim that improvements arise from seizure-relevant invariants rather than generic statistics cannot be evaluated.

    Authors: We concur that these elements are essential for validating the alignment mechanism. We will incorporate domain discrepancy metrics such as MMD before and after alignment, t-SNE visualizations of the aligned spaces, and ablation studies removing the multi-space alignment to demonstrate that the performance gains stem from seizure-relevant features. revision: yes

  3. Referee: [§4] §4: the assumption that seizure semiology and EEG signatures survive lissencephalic vs. gyrencephalic cortex differences and scalp vs. iEEG spatial filtering is load-bearing for the central claim yet is not tested by any control experiment that isolates biological vs. technical domain shift.

    Authors: The central experiments already cross both biological (canine vs human cortex) and technical (surface vs intracranial) domains, and the high performance indicates that the learned representations capture invariants across these shifts. We will add a dedicated discussion in the revised paper addressing how the results support survival of the signatures and acknowledge the lack of an isolated control as a limitation, while noting that such isolation may require future work with matched datasets. revision: partial

Circularity Check

0 steps flagged

No significant circularity; standard ML alignment techniques

full rationale

The paper applies established domain adaptation and knowledge distillation to align cross-species/cross-modality EEG signals. No equations, derivations, or self-citations are shown that reduce predictions to fitted parameters by construction, nor any self-definitional loops or uniqueness theorems imported from the authors' prior work. The reported AUC improvements are presented as empirical outcomes on external datasets rather than quantities forced by the method's own definitions.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no access to methods, equations, or results sections prevents identification of free parameters, axioms, or invented entities.

pith-pipeline@v0.9.0 · 5744 in / 1151 out tokens · 31957 ms · 2026-05-23T06:55:29.920356+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

82 extracted references · 82 canonical work pages · 1 internal anchor

  1. [1]

    An electric-field-responsive paramagnetic contrast agent enhances the visualization of epileptic foci in mouse models of drug-resistant epilepsy

    Wang C, Sun W, Zhang J et al. An electric-field-responsive paramagnetic contrast agent enhances the visualization of epileptic foci in mouse models of drug-resistant epilepsy . Nature Biomedical Engineering 2021; 5: 278–289

    work page 2021

  2. [2]

    Dentate gyrus mossy cells control spontaneous convulsive seizures and spatial memory .Science 2018; 359: 787–790

    Bui AD, Nguyen TM, Limouse C et al. Dentate gyrus mossy cells control spontaneous convulsive seizures and spatial memory .Science 2018; 359: 787–790

    work page 2018

  3. [3]

    Cycles in epilepsy

    Karoly PJ, Rao VR, Gregg NM et al. Cycles in epilepsy . Nature Reviews Neurology 2021; 17: 267–284

    work page 2021

  4. [4]

    Long-term outcomes in epilepsy surgery: antiepileptic drugs, mortality , cognitive and psychosocial aspects

    T ´ellez-Zenteno JF , Dhar R, Hernandez-Ronquillo L et al. Long-term outcomes in epilepsy surgery: antiepileptic drugs, mortality , cognitive and psychosocial aspects. Brain 2007; 130: 334–345

    work page 2007

  5. [5]

    The impact of epilepsy on patients’ lives

    Kerr MP . The impact of epilepsy on patients’ lives. Acta Neurologica Scandinavica 2012; 126: 1–9

    work page 2012

  6. [6]

    Automated recognition of epilepsy from EEG signals using a combining space– time algorithm of CNN-LSTM

    Wang X, Wang Y , Liu D et al. Automated recognition of epilepsy from EEG signals using a combining space– time algorithm of CNN-LSTM. Scientific Reports 2023; 13: 14876

    work page 2023

  7. [7]

    Weak supervision as an efficient approach for automated seizure detection in elec- troencephalography .NPJ Digital Medicine 2020; 3: 59

    Saab K, Dunnmon J, R ´e C et al. Weak supervision as an efficient approach for automated seizure detection in elec- troencephalography .NPJ Digital Medicine 2020; 3: 59

    work page 2020

  8. [8]

    Human intracranial EEG: Promises and limitations

    Parvizi J and Kastner S. Human intracranial EEG: Promises and limitations. Nature Neuroscience 2018; 21: 474

    work page 2018

  9. [9]

    Continuous and routine EEG in intensive care: Utilization and outcomes, United States 2005–2009

    Ney JP , van der Goes DN, Nuwer MR et al. Continuous and routine EEG in intensive care: Utilization and outcomes, United States 2005–2009. Neurology 2013; 81: 2002– 2008

    work page 2005

  10. [10]

    A guide to deep learning in healthcare

    Esteva A, Robicquet A, Ramsundar B et al. A guide to deep learning in healthcare. Nature Medicine 2019; 25: 24–29

    work page 2019

  11. [11]

    Automated EEG analysis of epilepsy: A review

    Acharya UR, Sree SV , Swapna G et al. Automated EEG analysis of epilepsy: A review. Knowledge-Based Systems 2013; 45: 147–165

    work page 2013

  12. [12]

    Wavelet-based EEG processing for computer-aided seizure detection and epilepsy diagnosis

    Faust O, Acharya UR, Adeli H et al. Wavelet-based EEG processing for computer-aided seizure detection and epilepsy diagnosis. Seizure 2015; 26: 56–64. Page 14 of 22

    work page 2015

  13. [13]

    Application of entropies for automated diagnosis of epilepsy using EEG signals: A review

    Acharya UR, Fujita H, Sudarshan VK et al. Application of entropies for automated diagnosis of epilepsy using EEG signals: A review. Knowledge-Based Systems 2015; 88: 85–96

    work page 2015

  14. [14]

    Deep convolutional neural network for the automated detection and diagnosis of seizure using EEG signals

    Acharya UR, Oh SL, Hagiwara Y et al. Deep convolutional neural network for the automated detection and diagnosis of seizure using EEG signals. Computers in Biology and Medicine 2018; 100: 270–278

    work page 2018

  15. [15]

    Optimized deep neural network architecture for robust detection of epilep - tic seizures using EEG signals

    Hussein R, Palangi H, Ward RK et al. Optimized deep neural network architecture for robust detection of epilep - tic seizures using EEG signals. Clinical Neurophysiology 2019; 130: 25–37

    work page 2019

  16. [16]

    Continuous seizure detection based on T ransformer and long-term iEEG.IEEE Journal of Biomedical and Health Informatics 2022; 26: 5418–5427

    Sun Y , Jin W, Si X et al. Continuous seizure detection based on T ransformer and long-term iEEG.IEEE Journal of Biomedical and Health Informatics 2022; 26: 5418–5427

    work page 2022

  17. [17]

    Unsupervised domain adapta- tion for cross-patient seizure classification

    Wang Z, Zhang W, Li S et al. Unsupervised domain adapta- tion for cross-patient seizure classification. Journal of Neu- ral Engineering 2023; 20: 066002

    work page 2023

  18. [18]

    Lightweight convolution trans- former for cross-patient seizure detection in multi-chann el EEG signals

    Rukhsar S and Tiwari AK. Lightweight convolution trans- former for cross-patient seizure detection in multi-chann el EEG signals. Computer Methods and Programs in Biomedicine 2023; 242: 107856

    work page 2023

  19. [19]

    Cross-patient automatic epilep- tic seizure detection using patient-adversarial neural ne t- works with spatio-temporal EEG augmentation

    Zhang Z, Ji T , Xiao M et al. Cross-patient automatic epilep- tic seizure detection using patient-adversarial neural ne t- works with spatio-temporal EEG augmentation. Biomedical Signal Processing and Control 2024; 89: 105664

    work page 2024

  20. [20]

    Lipid-accumulated reactive astrocytes promote disease progression in epilepsy .Nature Neuroscience 2023; 26: 542–554

    Chen ZP , Wang S, Zhao X et al. Lipid-accumulated reactive astrocytes promote disease progression in epilepsy .Nature Neuroscience 2023; 26: 542–554

    work page 2023

  21. [21]

    A cross- species approach to disorders affecting brain and be- haviour

    Devinsky O, Boesch JM, Cerda-Gonzalez S et al. A cross- species approach to disorders affecting brain and be- haviour. Nature Reviews Neurology 2018; 14: 677–686

    work page 2018

  22. [22]

    Obtaining human ischemic stroke gene expression biomarkers from animal models: a cross- species validation study .Scientific Reports 2016; 6: 29693

    Wang Y and Cai Y . Obtaining human ischemic stroke gene expression biomarkers from animal models: a cross- species validation study .Scientific Reports 2016; 6: 29693

    work page 2016

  23. [23]

    The role of astrocytes in seizure generation: Insights from a novel in vitro seizure model based on mitochondrial dysfunction

    Chan F , Lax NZ, Voss CM et al. The role of astrocytes in seizure generation: Insights from a novel in vitro seizure model based on mitochondrial dysfunction. Brain 2019; 142: 391–411

    work page 2019

  24. [24]

    Electroen- cephalography in dogs with epilepsy: Similarities between human and canine findings

    Berendt M, Høgenhaven H, Flagstad A et al. Electroen- cephalography in dogs with epilepsy: Similarities between human and canine findings. Acta Neurologica Scandinav- ica 1999; 99: 276–283

    work page 1999

  25. [25]

    Circadian and mul- tiday seizure periodicities, and seizure clusters in canin e epilepsy .Brain Communications 2020; 2: fcaa008

    Gregg NM, Nasseri M, Kremen V et al. Circadian and mul- tiday seizure periodicities, and seizure clusters in canin e epilepsy .Brain Communications 2020; 2: fcaa008

    work page 2020

  26. [26]

    Electroen- cephalography findings in healthy and Finnish Spitz dogs with epilepsy: Visual and background quantitative analysi s

    Jeserevics J, Viitmaa R, Cizinauskas S et al. Electroen- cephalography findings in healthy and Finnish Spitz dogs with epilepsy: Visual and background quantitative analysi s. Journal of Veterinary Internal Medicine 2007; 21: 1299– 1306

    work page 2007

  27. [27]

    Canine status epilep- ticus: Proof of principle studies

    Leppik IE, Patterson E, Hardy B et al. Canine status epilep- ticus: Proof of principle studies. Epilepsia 2009; 50

    work page 2009

  28. [28]

    Canine electroencephalo- graphic recording technique: Findings in normal and epileptic dogs

    Pellegrino FC and Sica RE. Canine electroencephalo- graphic recording technique: Findings in normal and epileptic dogs. Clinical Neurophysiology 2004; 115: 477– 487

    work page 2004

  29. [29]

    Integrating brain im- plants with local and distributed computing devices: A next generation epilepsy management system

    Kremen V , Brinkmann BH, Kim I et al. Integrating brain im- plants with local and distributed computing devices: A next generation epilepsy management system. IEEE Journal of T ranslational Engineering in Health and Medicine 2018; 6: 1–12

    work page 2018

  30. [30]

    Canine epilepsy: What can we learn from hu- man seizure disorders? The Veterinary Journal 2006; 172: 207–217

    Chandler K. Canine epilepsy: What can we learn from hu- man seizure disorders? The Veterinary Journal 2006; 172: 207–217

    work page 2006

  31. [31]

    Comparative clinical and electroencephalographic studies of canine epilepsy .Epilepsia 1970; 11: 281–292

    Holliday T , Cunningham J and Gutnick M. Comparative clinical and electroencephalographic studies of canine epilepsy .Epilepsia 1970; 11: 281–292

    work page 1970

  32. [32]

    Investigation of the presence of specific neural antibodies in dogs with epilepsy or dyskinesia using murine and human assays

    Hemmeter L, Bien CG, Bien CI et al. Investigation of the presence of specific neural antibodies in dogs with epilepsy or dyskinesia using murine and human assays. Journal of Veterinary Internal Medicine 2023; 37: 1409–1417

    work page 2023

  33. [33]

    Analysis of plasma-derived exosomal microRNAs as po- tential biomarkers for canine idiopathic epilepsy

    Garc´ ıa-Gracia M, Moreno-Martinez L, Hernaiz A et al. Analysis of plasma-derived exosomal microRNAs as po- tential biomarkers for canine idiopathic epilepsy . Animals 2024; 14: 252

    work page 2024

  34. [34]

    Crowdsourc- ing seizure detection: Algorithm development and valida- tion on human implanted device recordings

    Baldassano SN, Brinkmann BH, Ung H et al. Crowdsourc- ing seizure detection: Algorithm development and valida- tion on human implanted device recordings. Brain 2017; 140: 1680–1691

    work page 2017

  35. [35]

    EPILEPSIAE–A European epilepsy database

    Ihle M, Feldwisch-Drentrup H, T eixeira CA et al. EPILEPSIAE–A European epilepsy database. Computer Methods and Programs in Biomedicine 2012; 106: 127– 138

    work page 2012

  36. [36]

    A dataset of neonatal EEG recordings with seizure annotations

    Stevenson NJ, T apani K, Lauronen L et al. A dataset of neonatal EEG recordings with seizure annotations. Scien- tific Data 2019; 6: 1–8

    work page 2019

  37. [37]

    T ransfer learning for brain-computer inter - faces: A Euclidean space data alignment approach

    He H and Wu D. T ransfer learning for brain-computer inter - faces: A Euclidean space data alignment approach. IEEE T rans. on Biomedical Engineering 2020; 67: 399–410

    work page 2020

  38. [38]

    A survey of unsupervised deep domain adaptation

    Wilson G and Cook DJ. A survey of unsupervised deep domain adaptation. ACM T rans. on Intelligent Systems and T echnology2020; 11: 1–46

    work page

  39. [39]

    Distilling the knowledge i n a neural network

    Hinton G, Vinyals O and Dean J. Distilling the knowledge i n a neural network. In: Workshop on Proc. Advances in Neu- ral Information Processing Systems, Montr ´eal, Canada, Dec. 2015

    work page 2015

  40. [40]

    Visualizing data using t-SNE

    Van der Maaten L and Hinton G. Visualizing data using t-SNE. Journal of Machine Learning Research 2008; 9: 2579–2605

    work page 2008

  41. [41]

    A unified approach to interpret- ing model predictions

    Scott M, Su-In L et al. A unified approach to interpret- ing model predictions. Advances in Neural Information Pro- cessing Systems 2017; 30: 4765–4774

    work page 2017

  42. [42]

    From local explana- tions to global understanding with explainable ai for trees

    Lundberg SM, Erion G, Chen H et al. From local explana- tions to global understanding with explainable ai for trees . Nature Machine Intelligence 2020; 2: 56–67

    work page 2020

  43. [43]

    T ransfer learning for motor im- agery based brain-computer interfaces: A tutorial

    Wu D, Jiang X and Peng R. T ransfer learning for motor im- agery based brain-computer interfaces: A tutorial. Neural Networks 2022; 153: 235–253

    work page 2022

  44. [44]

    A system- atic evaluation of euclidean alignment with deep learning for EEG decoding

    Junqueira B, Aristimunha B, Chevallier S et al. A system- atic evaluation of euclidean alignment with deep learning for EEG decoding. Journal of Neural Engineering 2024; 21: 036038

    work page 2024

  45. [45]

    Affective brain–computer inter- faces (abcis): A tutorial

    Wu D, Lu BL, Hu B et al. Affective brain–computer inter- faces (abcis): A tutorial. Proceedings of the IEEE 2023; 111: 1314–1332

    work page 2023

  46. [46]

    Deep transfer learning with joint adaptation networks

    Long M, Zhu H, Wang J et al. Deep transfer learning with joint adaptation networks. In: Proc. Int’l Conf. Machine Learning, Sydney , Australia, Aug. 2017

    work page 2017

  47. [47]

    Deep subdomain adapta- tion network for image classification

    Zhu Y , Zhuang F , Wang J et al. Deep subdomain adapta- tion network for image classification. IEEE T rans. on Neural Networks and Learning Systems 2020; 32: 1713–1722. Page 15 of 22

    work page 2020

  48. [48]

    EEGNet: A compact convolutional neural network for EEG-based brain-computer interfaces

    Lawhern VJ, Solon AJ, Waytowich NR et al. EEGNet: A compact convolutional neural network for EEG-based brain-computer interfaces. Journal of Neural Engineering 2018; 15: 056013

    work page 2018

  49. [49]

    Leakage and the repro- ducibility crisis in machine-learning-based science

    Kapoor S and Narayanan A. Leakage and the repro- ducibility crisis in machine-learning-based science. Pat- terns 2023; 4: 100804

    work page 2023

  50. [50]

    Learning transferable fea- tures with deep adaptation networks

    Long M, Cao Y , Wang J et al. Learning transferable fea- tures with deep adaptation networks. In: Proc. Int’l Conf. Machine Learning, Lille, France, Jul. 2015

    work page 2015

  51. [51]

    A kernel two- sample test

    Gretton A, Borgwardt KM, Rasch MJ et al. A kernel two- sample test. Journal of Machine Learning Research 2012; 13: 723–773

    work page 2012

  52. [52]

    Minimum class confusion for versatile domain adaptation

    Jin Y , Wang X, Long M et al. Minimum class confusion for versatile domain adaptation. In: Proc. European Conf. Computer Vision, Glasgow, UK, Aug. 2020

    work page 2020

  53. [53]

    Source data-absent unsuper- vised domain adaptation through hypothesis transfer and labeling transfer

    Liang J, Hu D, Wang Y et al. Source data-absent unsuper- vised domain adaptation through hypothesis transfer and labeling transfer. IEEE T rans. Pattern Analysis and Ma- chine Intelligence 2022; 44: 8602–8617

    work page 2022

  54. [54]

    Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer

    Zagoruyko S and Komodakis N. Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer. In: Proc. Int’l Con f. Learning Representations, T oulon, France, Apr. 2017

    work page 2017

  55. [55]

    Like What You Like: Knowledge Distill via Neuron Selectivity Transfer

    Huang Z and Wang N. Like what you like: Knowl- edge distill via neuron selectivity transfer ; ArXiv prepri nt arXiv:1707.01219, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  56. [56]

    Similarity-preserving knowledge dist il- lation

    T ung F and Mori G. Similarity-preserving knowledge dist il- lation. In: Proc. IEEE/CVF Int’l Conf. on Computer Vision, Seoul, Korea, Oct. 2019

    work page 2019

  57. [57]

    Relational knowledge distilla- tion

    Park W, Kim D, Lu Y et al. Relational knowledge distilla- tion. In: Proc. IEEE/CVF Conf. Computer Vision and Pat- tern Recognition, Long Beach, CA, USA, June 2019

    work page 2019

  58. [58]

    Learning deep representations with probabilistic knowledge transfer

    Passalis N and T efas A. Learning deep representations with probabilistic knowledge transfer. In: Proc. European Conf. on Computer Vision, Munich, Germany , Sep. 2018

    work page 2018

  59. [59]

    Correlation congruence for knowl- edge distillation

    Peng B, Jin X, Liu J et al. Correlation congruence for knowl- edge distillation. In: Proc. IEEE/CVF Int’l Conf. on Com- puter Vision, Seoul, Korea, Oct. 2019

    work page 2019

  60. [60]

    Crowdsourc- ing reproducible seizure forecasting in human and canine epilepsy .Brain 2016; 139: 1713–1722

    Brinkmann BH, Wagenaar J, Abbot D et al. Crowdsourc- ing reproducible seizure forecasting in human and canine epilepsy .Brain 2016; 139: 1713–1722. Page 16 of 22 APPENDIX A: ALGORITHMS UNDER COMP ARISON The proposed MSA framework is illustrated in Figure A1. ResizeNet (Figure A1b) incorporates a Transformer encoder, a linear layer, and two reshape modules...

    work page 2016

  61. [61]

    Deep Adaptation Network (DAN) [50], which achieves feature alignment by min- imizing the maximum mean discrepancies (MMD) [51] in the feature space

    work page

  62. [62]

    Joint Adaptation Network (JAN) [46] and Deep Subdomain Adaptation Network (DSAN) [47], which construct transfer networks by aligning class-conditioned sub- domain distributions based on local MMD in the feature space, utilizing either model prediction probabilities or pseudo-labels

    work page

  63. [63]

    Minimum class confusion (MCC) [52], which integrates a weighted prediction entropy with class correlation to reduce class confusion in the output space

    work page

  64. [64]

    Six popular and diverse knowledge distilla- tion approaches were combined with the pro- posed ResizeNet to further improve the align- ment performance:

    Source hypothesis transfer (SHOT) [53], which utilizes information maximization for conditional entropy minimization and label marginal entropy regularization in the output space. Six popular and diverse knowledge distilla- tion approaches were combined with the pro- posed ResizeNet to further improve the align- ment performance:

    work page

  65. [65]

    Attention Transfer (A T) [54], which utilizes an attention mechanism to enhance the stu- dent network’s performance by mapping at- tention from the teacher’s feature maps

    work page

  66. [66]

    Neuron Selectivity Transfer (NST) [55], which treats knowledge transfer as a distri- bution matching problem by aligning neuron selectivity patterns between networks

    work page

  67. [67]

    Similarity-Preserving (SP) [56], which en- sures the preservation of pairwise similarities between activations in the teacher network to maintain relational integrity in the student network

    work page

  68. [68]

    Relational Knowledge Distillation (RKD) [57], which focuses on transferring mutual re- lations of data samples using distance-wise and angle-wise distillation losses

    work page

  69. [69]

    Probabilistic Knowledge Transfer (PKT) [58], which aligns the probability distribu- tions in the feature space, rather than directly mapping the features

    work page

  70. [70]

    #$%"&'()%*%+*#,- ! !

    Correlation Congruence (CC) [59], which transfers correlations between instances us- ing a generalized kernel method based on the Taylor series expansion. The classification accuracies (%) of the base- line methods and the proposed ResizeNet+ ap- proaches are summarized in Table A1. APPENDIX B: DAT ASET INFORMATION The Kaggle dataset includes data from eig...

    work page

  71. [71]

    Since there were no channel differences between the NICU and CHSZ datasets, the proposed ResizeNet+MSA approach was not applied in this setting

    Cross-dataset, within-species (human-to- human) and within-modality transfer: In this scenario, models were trained on human sEEG data from the NICU or CHSZ dataset and tested on another human sEEG dataset. Since there were no channel differences between the NICU and CHSZ datasets, the proposed ResizeNet+MSA approach was not applied in this setting

    work page

  72. [72]

    Input discrepancies exist in this setting, making alignment necessary

    Cross-dataset, within-species (human-to- human) and cross-modality transfer: Models were trained on human iEEG or sEEG datasets and tested on human datasets of a different modality (e.g., iEEG to sEEG, or sEEG to iEEG). Input discrepancies exist in this setting, making alignment necessary

    work page

  73. [73]

    This setting involves both multi- Page 19 of 22 Table B1

    Cross-dataset, multi-species [(canine+human)-to-human] and multi- modality transfer: Canine iEEG and human iEEG/sEEG datasets were combined as the training set, with CHSZ used as the test dataset. This setting involves both multi- Page 19 of 22 Table B1. Summary of the four epilepsy datasets. Dataset EEG Type # Patients # Channels Sampling rate Signal len...

    work page

  74. [74]

    The cross-dataset within-species transfer per- formance was better than basic cross-species transfer without additional techniques

    work page

  75. [75]

    ResizeNet+MSA demonstrated its effective- ness in cross-dataset cross-modality experi- ments

    work page

  76. [76]

    Combining multi-species data in the training set and testing on a human dataset resulted in the best performance

    work page

  77. [77]

    UDA approaches were not always effec- tive

    For the within-modality scenario, NICU and CHSZ datasets have the same num- ber and placement of channels, so there were no data heterogeneous discrepancies. UDA approaches were not always effec- tive. Compared to the cross-modality sce- nario, within-modality transfer achieved bet- ter performance, e.g., 85. 17% > 84. 37% and

    work page

  78. [78]

    24%, indicating that cross- modality transfer (with data heterogeneous discrepancy) is indeed more challenging

    87% > 67. 24%, indicating that cross- modality transfer (with data heterogeneous discrepancy) is indeed more challenging

    work page

  79. [79]

    While UDA approaches may be effective in some cases, the proposed ResizeNet frame- work consistently outperformed others

    For the cross-modality scenario, iEEG and sEEG signals, collected from different de- vices, introduce significant data heterogene- ity. While UDA approaches may be effective in some cases, the proposed ResizeNet frame- work consistently outperformed others. No- tably, ResizeNet+MSA achieved the highest performance among all approaches

    work page

  80. [80]

    With the increasing number of training data, the performance improved, e.g., 84

    For the multi-modality scenario, we further combined two-species datasets to build the training set. With the increasing number of training data, the performance improved, e.g., 84. 37% < 86. 33% < 87. 05%. With the help of ResizeNet+MSA, the best perfor- mance was achieved on the CHSZ dataset, e.g., 91.34%, surpassing models trained on single-species sin...

    work page

Showing first 80 references.