Pith. sign in

REVIEW 3 major objections 4 minor 50 references

TransformEEG: Towards Improving Model Generalizability in Deep Learning-based EEG Parkinson's Disease Detection

T0 review · 3 major / 4 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 TransformEEG, a hybrid convolutional-transformer with a depthwise convolutional tokenizer, reports the highest median balanced accuracy and lowest variability among eight EEG deep learning models on a 290-subject Parkinson's disease…

desk verdict A serious, well-run benchmark whose headline claim is not yet supported: the young healthy cohort in ds004148 is never tested as a confound, and the paper's own two-dataset result reverses the ranking. read the letter →

arxiv 2507.07622 v1 pith:UUGDULS2 submitted 2025-07-10 cs.LG cs.AI

Federico Del Pup , Riccardo Brun , Filippo Iotti , Edoardo Paccagnella , Mattia Pezzato , Sabrina Bertozzo , Andrea Zanola , Louis Fabrice Tshimanga
show 2 more authors
Henning Müller Manfredo Atzori
This is my paper · ORCID
classification cs.LGcs.AI
keywords EEGParkinson'sdiseasedeeplearningtransformerdepthwiseconvolutiongeneralizabilitynestedcross-validationbalancedaccuracy
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 argues that poor generalizability in EEG deep learning models comes in part from how tokens are built for attention layers. EEGNet-style encoders mix channels before attention, so the transformer sees redundant projections. TransformEEG replaces this with a depthwise convolutional tokenizer that keeps channel-specific local temporal features separate, letting self-attention do the channel mixing. On four harmonized public datasets with 290 subjects (140 Parkinson's patients, 150 healthy controls), evaluated with a nested leave-N-subjects-out cross-validation, TransformEEG reports the highest median balanced accuracy (78.45% baseline; 80.10% with data augmentation and threshold correction) and the lowest interquartile range (6.37% and 5.74%). The paper concludes that this design produces more consistent, less skewed results and more reliable Parkinson's disease detection from EEG.

What carries the argument

The central object is the depthwise convolutional tokenizer, a two-block stack of depthwise 1D convolutions with batch normalization, ELU activation, average pooling (kernel 4, stride 2), dropout, and residual connections. It maps an input window $(B,C,L)$ to $(B,C\times D,\lfloor (L+(S-K)(1+S))/S^2\rfloor)$—for this paper, $(64,32,2000)$ to $(64,128,498)$—while keeping each feature tied to one EEG channel. The attached transformer encoder (two layers, single attention head, no positional embeddings, no class token) then performs all cross-channel mixing, and a global-average-pooled MLP produces the Parkinson's disease probability.

What would settle it

Train TransformEEG and the baselines on the two age-matched datasets only (81 subjects, as in the paper's scalability analysis) and compare per-split balanced accuracy; if TransformEEG's median no longer exceeds ShallowNet's 78.73% baseline, or if the aggregated four-dataset advantage disappears when the young healthy cohort is removed, the claim that the architecture improves generalizability for Parkinson's disease detection would be falsified.

Watch

Extended reading notes

Core claim

The central claim is that a hybrid architecture with deliberately separated roles—depthwise convolutions extract channel-specific local temporal features, and self-attention mixes across channels—generalizes better across subjects than existing EEG deep learning models. The evidence is an aggregated benchmark of four datasets with 290 subjects; across 100 nested cross-validation splits, TransformEEG ranks first in median balanced accuracy and interquartile range, and is the only model whose test accuracies are almost all above 70% when optimal augmentation is used. The authors attribute the gain to the tokenizer design: maintaining channel identity in the tokens enables more effective feature mixing inside the transformer encoder, reducing overfitting and split-to-split variability.

Load-bearing premise

The benchmark assumes that combining the four datasets yields a clean Parkinson's-versus-healthy contrast, so the model is not latching onto age or dataset identity rather than disease.

Editorial extensions

If this is right

  • On the aggregated four-dataset benchmark, TransformEEG reports the highest median balanced accuracy among eight models: 78.45% baseline and 80.10% with data augmentation plus threshold correction.
  • Its interquartile range is the lowest in every training configuration (6.37% baseline, 5.74% with augmentation and threshold correction), which the paper interprets as more consistent and less skewed predictions across subject splits.
  • With optimal augmentation, 99 of 100 nested cross-validation test accuracies are above 70%, the only model in the comparison to achieve this.
  • When training data shrink to 81 subjects, all models including TransformEEG show increased variability, indicating that the architecture's advantage requires sufficiently large, multi-center data.
  • Supplementary ablations show that positional embeddings, class tokens, additional attention heads, spectral-fusion branches, and common spatial pattern scaling do not improve TransformEEG.

Reading between the lines

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

  • Editorial inference: if the four-dataset advantage is partly driven by the young healthy cohort in one dataset, the generalizability claim should be re-tested on age-matched data; the paper's own two-dataset results already show ShallowNet ahead (78.73% vs 72.09% median baseline).
  • Editorial inference: the same depthwise tokenizer design could be transferred to other EEG pathology tasks where channel-level local features matter, which would test whether the benefit is task-specific or general.
  • Editorial inference: the paper's Augmentation Relative Improvement Score could be reused as a standard metric for selecting EEG data augmentations in other benchmark studies.
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 / 4 minor

Summary. This paper proposes TransformEEG, a hybrid convolutional-transformer architecture for Parkinson's disease detection from EEG, with a depthwise convolutional tokenizer intended to preserve channel-specific features before self-attention. The authors aggregate four public OpenNeuro datasets into 290 subjects (140 PD, 150 healthy), harmonize them with BIDSAlign, and evaluate with a 10-outer, 10-inner Nested-Leave-N-Subjects-Out (N-LNSO) scheme against seven baseline models. They report that TransformEEG achieves the highest median balanced accuracy (78.45% baseline; 80.10% with augmentation and threshold correction) and the lowest interquartile range, and conclude that it yields more reliable PD detection. The paper also contributes an ARIS score for data-augmentation selection, a seed-sensitivity analysis, and extensive supplementary ablations.

Significance. The study is unusually transparent: code is released, the architectural tables are detailed, the N-LNSO evaluation is subject-disjoint and nested, the baseline set spans a wide parameter-count range, and the supplementary material reports F1/kappa, seed sensitivity, window-length/overlap sweeps, and several negative ablations. If the causal claim about the tokenizer and the 'reliable PD detection' claim were supported, this would be a useful contribution to EEG-based PD classification and to the generalizability literature. However, two load-bearing weaknesses—an age/dataset confound in the aggregated benchmark and seed-dependent ranking—currently prevent the conclusions from being accepted as stated; the significance of the paper therefore depends on resolving these issues.

major comments (3)
  1. [Section 2.1, Table 1, Section 3.4] The aggregated benchmark contains a class–age confound. ds004148 contributes 60 healthy subjects with mean age 20.0±1.9 years, while the PD groups in ds002778, ds003490, and ds004584 have mean ages 63.3, 69.7, and 68.5 years and their healthy controls are age-matched (63.5, 69.3, 70.9). Because N-LNSO partitions by subject, a model can separate classes by age-related EEG properties or recording site rather than by PD pathology. The article never reports per-dataset accuracy or a leave-dataset-out analysis. The paper's own Table 3 provides the natural check: on the two age-matched datasets (ds002778+ds003490, 81 subjects), TransformEEG's baseline median balanced accuracy is 72.09%, below ShallowNet's 78.73%, reversing the four-dataset ranking (78.45% vs 77.72%). Section 4 interprets this result only as a sample-size effect ('an interquartile range increase of 9.56% and the median balanced accuracy drop of 6.01%'), but it is equally or more plausibly a removal of the young healthy cohort. Please add per-dataset results, an age/domain-confounding analysis (e.g., evaluate with ds004148 held out, or match healthy ages), and adjust the Abstract and Section 5 claims accordingly.
  2. [Section A.3, Supplementary Table 2] The seed analysis in Section A.3 (Supplementary Table 2) does not support the statement that 'the study's conclusions are not influenced by the choice of the seed.' At seed 12, TransformEEG's median balanced accuracy is 76.10%, lower than ShallowNet (77.19%) and EEGConformer (77.21%); at seed 1 it is 77.33%, lower than ShallowNet (77.49%). Only seed 42 gives the ordering reported in Section 3.1. Since the main text reports only seed 42 (Section 2.4.3), the claim that TransformEEG 'ranks first' in median accuracy is not stable across seeds. Please report per-seed rankings or a summary across seeds, and temper the conclusion that TransformEEG 'achieves the highest median' to reflect the observed seed-to-seed variability.
  3. [Section 2.3.1, Section 3.1, Supplementary A.5] The paper's architectural claim—that the depthwise convolutional tokenizer 'enables more effective feature mixing within the self-attention layers'—is not tested by a controlled ablation. The comparison in Section 3.1 pits TransformEEG against EEGConformer and ATCNet, which differ in tokenizer design, transformer depth, attention heads, and classification head. Supplementary A.5 ablates positional encoding, class token, and number of heads, but not the tokenizer. A direct ablation that keeps the transformer encoder and classification MLP fixed and swaps only the tokenizer (e.g., an EEGNet-style 2D convolutional tokenizer or a non-depthwise 1D tokenizer) is needed to support the central design rationale. Without it, the performance difference could be due to overall architecture scale or regularization, not to channel-specific tokenization.
minor comments (4)
  1. [Table 12, Figure 1] Table 12 and Figure 1 are inconsistent about the average-pooling stride: Figure 1 states 'Kernel = 4, Stride = 2', while Table 12 lists 'AvgPool1d [4]' with no stride, which in PyTorch defaults to stride 4; the output lengths 999/498 in Table 12 correspond to stride 2, so Table 12 should be corrected or clarified.
  2. [Section A.1] Section A.1 contains a duplicated word: 'an ensemble of of N_outer x N_inner accuracies' should be 'an ensemble of N_outer x N_inner accuracies.'
  3. [Section 3.4] Section 3.4 is titled 'Model scalability,' but the two-dataset comparison is also the only age-matched subset in the paper; it should be cross-referenced to the confound discussion and explicitly labeled as an age-matched check in the main text.
  4. [Section A.7.1] In the window-aggregation analysis, the 'minimal ratio of positive windows' should state explicitly whether this ratio is optimized on the validation set for each N-LNSO split or determined globally, since this affects the interpretation of the aggregated accuracy.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation; TransformEEG's benchmark advantage is an empirical result, with same-group methodology citations adding self-citation weight but no reduction.

full rationale

The paper's central claim is that TransformEEG achieves the highest median balanced accuracy and the lowest IQR among eight models under a 10-outer, 10-inner N-LNSO evaluation on four public OpenNeuro datasets. The reported metrics are computed on held-out test subjects and are not defined in terms of the model's own fitted parameters or outputs, so the outcome is not encoded in the architecture by construction. The depthwise convolutional tokenizer is a new architectural component, not an imported ansatz, and no equation in the paper defines the reported accuracies from the training inputs. The evaluation methodology relies on same-group earlier work (N-LNSO from Del Pup et al. 2025, BIDSAlign preprocessing, SelfEEG), and this self-citation weight is non-trivial; however, these citations supply a concrete cross-validation scheme and openly available code that the paper also describes and that is externally checkable, rather than a uniqueness theorem invoked to forbid alternatives. Threshold correction and per-model data-augmentation selection (ARIS) are performed on the validation folds only (Sections 2.4.2 and 3.3), so the final test numbers are not a fitted input renamed as a prediction. The age/dataset confound raised by ds004148 (60 healthy subjects aged 20.0 +/- 1.9 years versus PD groups aged 63-70 in Table 1) and the reversal in Table 3 (TransformEEG 72.09% baseline median vs ShallowNet 78.73% on the two age-matched datasets) is a real external-validity threat to the 'more reliable PD detection' conclusion, but it is a correctness risk, not a circularity: the benchmark does not define reliability in terms of the confound. No circular step can be exhibited with a specific reduction, so the score rests only on the presence of methodological self-citations that are not load-bearing for the central empirical comparison.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the assumption that harmonized multi-dataset aggregation produces a clean PD-vs-healthy task, that the ICLabel-based preprocessing preserves disease signal, and that the N-LNSO protocol yields comparable, unbiased estimates. The most fragile assumption is the absence of dataset/age confounds: ds004148 contributes only young healthy subjects, and no analysis controls for this. Free parameters include the per-model validation-tuned classification threshold and the per-model data augmentation composition selected from 100 candidates on the full corpus; these are legitimate tuning choices but they shape the reported absolute numbers. No invented entities are introduced.

free parameters (3)
  • Classification threshold per model = tuned on validation set, exact values not reported
    Section 3.3: the 0.5 default threshold is replaced by the value maximizing validation balanced accuracy for each model; used in the B+DA+T headline results.
  • Data augmentation composition per model = TransformEEG: masking + time reverse
    Section 2.4.2; chosen from 100 combinations using ARIS on the full dataset for each of the eight models, then reused in the B+DA and B+DA+T comparisons.
  • Window length and overlap = 16 seconds, 25% overlap
    Section A.6; selected via ablation as a compromise between median balanced accuracy and IQR; a design choice that shapes all reported numbers.
assumptions (4)
  • domain assumption EEG slowing (decreased alpha/beta power, increased delta/theta power) is a usable biomarker separating PD from healthy controls.
    Introduction cites refs [7,8]; the entire classification task presupposes PD-vs-healthy is decodable from resting-state EEG.
  • domain assumption ICLabel-based automatic component rejection with [90%,100%] noise confidence and [0%,10%] brain confidence preserves disease-relevant signal across all four datasets.
    Section 2.2; adopted from the authors' prior work [14] without independent verification in this study.
  • domain assumption The Nested-Leave-N-Subjects-Out scheme with 10 outer and 10 inner folds provides unbiased and comparable performance estimates across models.
    Section 2.4.1 and A.1; the method comes from the authors' own [17], and the paper notes it violates independence assumptions of standard statistical tests.
  • ad hoc to paper Harmonizing four datasets to 32 common channels and aggregating them introduces no systematic confounds (e.g., age or site) between the PD and healthy classes.
    Section 2.2/Table 1; ds004148 supplies 60 healthy subjects aged 20±1.9 while PD patients are 63+, so this assumption is load-bearing and unverified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TransformEEG: Towards Improving Model Generalizability in Deep Learning-based EEG Parkinson's Disease Detection." pith.science (2026). https://pith.science/paper/UUGDULS2

@misc{pith2026250707622,
  author       = {Pith},
  title        = {Pith review of: TransformEEG: Towards Improving Model Generalizability in Deep Learning-based EEG Parkinson's Disease Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UUGDULS2}},
  note         = {Machine review of arXiv:2507.07622}
}
read the original abstract

Electroencephalography (EEG) is establishing itself as an important, low-cost, noninvasive diagnostic tool for the early detection of Parkinson's Disease (PD). In this context, EEG-based Deep Learning (DL) models have shown promising results due to their ability to discover highly nonlinear patterns within the signal. However, current state-of-the-art DL models suffer from poor generalizability caused by high inter-subject variability. This high variability underscores the need for enhancing model generalizability by developing new architectures better tailored to EEG data. This paper introduces TransformEEG, a hybrid Convolutional-Transformer designed for Parkinson's disease detection using EEG data. Unlike transformer models based on the EEGNet structure, TransformEEG incorporates a depthwise convolutional tokenizer. This tokenizer is specialized in generating tokens composed by channel-specific features, which enables more effective feature mixing within the self-attention layers of the transformer encoder. To evaluate the proposed model, four public datasets comprising 290 subjects (140 PD patients, 150 healthy controls) were harmonized and aggregated. A 10-outer, 10-inner Nested-Leave-N-Subjects-Out (N-LNSO) cross-validation was performed to provide an unbiased comparison against seven other consolidated EEG deep learning models. TransformEEG achieved the highest balanced accuracy's median (78.45%) as well as the lowest interquartile range (6.37%) across all the N-LNSO partitions. When combined with data augmentation and threshold correction, median accuracy increased to 80.10%, with an interquartile range of 5.74%. In conclusion, TransformEEG produces more consistent and less skewed results. It demonstrates a substantial reduction in variability and more reliable PD detection using EEG data compared to the other investigated models.

Figures

Figures reproduced from arXiv: 2507.07622 by the authors.

Figure 1
Figure 1. Schematic representation of the TransformEEG architecture. TransformEEG consists of three modules: a depthwise convolutional tokenizer (A), a transformer encoder (B), and a classification MLP (C). The depthwise convolutional tokenizer creates EEG tokens describing local time portions of the input window with channel-specific features. The transformer encoder recombines the tokens with the self-attention mechanism. T… view at source ↗
Figure 2
Figure 2. Performance comparison of the selected models using a 10-outer, 10-inner N-LNSO cross-validation scheme. Models are organized in ascending order based on the number of learnable parameters from left to right. All models were trained without data augmentation or threshold correction to serve as a baseline reference. TransformEEG ranks first in median balanced accuracy and interquartile range. It also achieved the hig… view at source ↗
Figure 3
Figure 3. Effect of data augmentation on the performance of TransformEEG and the other selected EEG-based deep learning models. Panels A and B display the median and interquartile range of balanced accuracy, respectively, when different data augmentation compositions are incorporated during the training of TransformEEG. For each augmentation combination, an N-LNSO cross-validation was performed. Panel C shows the results of a… view at source ↗
Figures from the paper (5 more)
Figure 1
Figure 1. Figure 1: Scheme of the N-LNSO cross-validation method proposed in Del Pup et al., 2025. N-LNSO adds an additional nested level to Leave-N-Subjects-Out approaches to generate a set of train-validation-test splits. Models are trained on the training set, monitored with early stop…
Figure 3
Figure 3. Figure 3: Comparison between TransformEEG and variants that use a transformer encoder layer with more heads. F. Del Pup et al.: Preprint submitted to Elsevier Page 18 of 30 [PITH_FULL_IMAGE:figures/full_fig_p018_3.png]
Figure 4
Figure 4. Figure 4 [PITH_FULL_IMAGE:figures/full_fig_p021_4.png]
Figure 5
Figure 5. Figure 5: Effects of the addition of the PSD on the performance of TransformEEG. Panel A schematizes mid- and late-fusion strategies. Panel B shows the comparison between TransformEEG and variants that include the PSD branch. All models are trained with the masking plus flip hor…
Figure 6
Figure 6. Figure 6: TransformEEG balanced accuracy variation when trained with an increasing number of CSP filters. Panel A shows the balanced accuracy using a baseline pipeline without data augmentation and threshold correction. Panel b shows the balanced accuracy using a baseline pipeli…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 30 canonical work pages

  1. [1]

    Pringsheim, N

    T. Pringsheim, N. Jette, A. Frolkis, T. D. Steeves, The prevalence of Parkinson’s disease: A systematic review and meta-analysis, Mov. Disord. 29 (13) (2014) 1583–1590.doi:10.1002/mds.25945

    work page doi:10.1002/mds.25945 2014

  2. [2]

    Dauer, S

    W. Dauer, S. Przedborski, Parkinson’s disease: mechanisms and models, Neuron 39 (6) (2003) 889–909. doi:10.1016/ S0896-6273(03)00568-3

    work page 2003

  3. [3]

    neurochem

    S.Sveinbjornsdottir,TheclinicalsymptomsofParkinson’sdisease,J. neurochem. 139 (2016) 318–324.doi:10.1111/jnc.13691

    work page doi:10.1111/jnc.13691 2016

  4. [4]

    S.-Y. Lim, A. E. Lang, The nonmotor symptoms of Parkinson’s dis- ease—an overview, Mov. Disord. 25 (S1) (2010) S123–S130.doi: 10.1002/mds.22786

    work page doi:10.1002/mds.22786 2010

  5. [5]

    M. J. Armstrong, M. S. Okun, Diagnosis and treatment of Parkinson Disease:Areview,JAMA323(6)(2020)548–560. doi:10.1001/ jama.2019.22360

    arXiv 2020

  6. [6]

    Monte-Silva, Quantitative electroencephalography char- acteristicsforParkinson’sdisease:Asystematicreview,J.Parkinson’s Dis

    L.Shirahige,M.Berenguer-Rocha,S.Mendonça,S.Rocha,M.C.Ro- drigues, K. Monte-Silva, Quantitative electroencephalography char- acteristicsforParkinson’sdisease:Asystematicreview,J.Parkinson’s Dis. 10 (2) (2020) 455–470.doi:10.3233/JPD-191840

    work page doi:10.3233/jpd-191840 2020

  7. [7]

    V. J. Geraedts, L. I. Boon, J. Marinus, A. A. Gouw, J. J. van Hilten, C. J. Stam, M. R. Tannemaat, M. F. Contarino, Clinical correlates of quantitative EEG in Parkinson disease: A systematic review, Neurology 91 (19) (2018) 871–883.doi:10.1212/WNL. 0000000000006473

    doi:10.1212/wnl 2018

  8. [8]

    Soikkeli, J

    R. Soikkeli, J. Partanen, H. Soininen, A. Pääkkönen, P. Riekki- nen, Slowing of EEG in Parkinson’s disease, Electroencephalogr. Clin. Neurophysiol. 79 (3) (1991) 159–165. doi:10.1016/ 0013-4694(91)90134-P

    work page 1991

Show all 50 references
  1. [9]

    A. M. Maitin, J. P. Romero Muñoz, A. J. García-Tejedor, Survey of machine learning techniques in the analysis of EEG Signals for Parkinson’s disease: A systematic review, Appl. Sci. 12 (14) (2022). doi:10.3390/app12146967

    2022 doi

  2. [10]

    H. W. Loh, W. Hong, C. P. Ooi, S. Chakraborty, P. D. Barua, R. C. Deo, J. Soar, E. E. Palmer, U. R. Acharya, Application of deep learning models for automated identification of Parkinson’s disease: A review (2011–2021), Sensors 21 (21) (2021).doi:10.3390/ s21217034

    2021

  3. [11]

    S. K. Khare, V. Bajaj, U. R. Acharya, PDCNNet: An automatic frameworkforthedetectionofParkinson’sdiseaseusingEEGsignals, IEEESens.J.21(15)(2021)17017–17024. doi:10.1109/JSEN. 2021.3080135

    2021

  4. [12]

    Balaji, D

    E. Balaji, D. Brindha, V. K. Elumalai, R. Vikrama, Automatic and non-invasive Parkinson’s disease diagnosis and severity rating using LSTM network, Appl. Soft Comput. 108 (2021) 107463. doi: 10.1016/j.asoc.2021.107463

    2021

  5. [13]

    Miltiadous, E

    A. Miltiadous, E. Gionanidis, K. D. Tzimourta, N. Giannakeas, A. T. Tzallas, DICE-net: a novel convolution-transformer architecture for Alzheimer detection in EEG signals, IEEE Access (2023).doi: 10.1109/ACCESS.2023.3294618

    2023

  6. [14]

    Neural Syst

    F.DelPup,A.Zanola,L.FabriceTshimanga,A.Bertoldo,M.Atzori, The more, the better? Evaluating the role of EEG preprocessing for deep learning applications, IEEE Trans. Neural Syst. Rehabil. Eng. 33 (2025) 1061–1070.doi:10.1109/TNSRE.2025.3547616

    2025

  7. [15]

    Pion-Tonachini, K

    L. Pion-Tonachini, K. Kreutz-Delgado, S. Makeig, ICLabel: An au- tomated electroencephalographic independent component classifier, dataset, and website, NeuroImage 198 (2019) 181–197.doi:10. 1016/j.neuroimage.2019.05.026

    2019

  8. [16]

    T. R. Mullen, C. A. E. Kothe, Y. M. Chi, A. Ojeda, T. Kerth, S.Makeig,T.-P.Jung,G.Cauwenberghs,Real-timeneuroimagingand cognitive monitoring using wearable dry EEG, IEEE Trans. Biomed. Eng. 62 (11) (2015) 2553–2567. doi:10.1109/TBME.2015. 2481482

    2015 doi

  9. [17]

    Del Pup, A

    F. Del Pup, A. Zanola, L. F. Tshimanga, A. Bertoldo, L. Finos, M. Atzori, The role of data partitioning on the performance of EEG- based deep learning models in supervised cross-subject analysis: A preliminary study, Comput. Biol. Med. 196 (2025) 110608.doi: 10.1016/j.compbiom...

    2025

  10. [18]

    Y. Song, Q. Zheng, B. Liu, X. Gao, EEG Conformer: Convolu- tionalTransformerforEEGDecodingandVisualization,IEEETrans. Neural Syst. Rehabil. Eng. 31 (2023) 710–719.doi:10.1109/ TNSRE.2022.3230250

    2023

  11. [19]

    H.Altaheri,G.Muhammad,M.Alsulaiman,Physics-informedatten- tion temporal convolutional network for EEG-based motor imagery classification, IEEE Trans. Ind. Inform. 19 (2) (2023) 2249–2258. doi:10.1109/TII.2022.3197419

    2023

  12. [20]

    V. J. Lawhern, A. J. Solon, N. R. Waytowich, S. M. Gordon, C. P. Hung,B. J.Lance,EEGNet: acompactconvolutionalneural network for EEG-based brain–computer interfaces, J. Neural Eng. 15 (5) (2018) 056013. doi:10.1088/1741-2552/aace8c

    2018 doi

  13. [21]

    Del Pup, M

    F. Del Pup, M. Atzori, Toward improving reproducibility in neu- roimaging deep learning studies, Front. Neurosc. 18 (2024).doi: 10.3389/fnins.2024.1509358

    2024

  14. [22]

    C. J. Markiewicz, K. J. Gorgolewski, F. Feingold, R. Blair, Y. O. Halchenko, E. Miller, N. Hardcastle, J. Wexler, O. Esteban, M. Goncavles, A. Jwa, R. Poldrack, The OpenNeuro resource for sharing of neuroscience data, eLife 10 (2021) e71774.doi:10. 7554/eLife.71774. F. Del Pup...

    2021

  15. [23]

    Delorme, R

    C.R.Pernet,S.Appelhoff,K.J.Gorgolewski,G.Flandin,C.Phillips, A. Delorme, R. Oostenveld, EEG-BIDS, an extension to the brain imaging data structure for electroencephalography, Sci. Data 6 (1) (2019) 103. doi:10.1038/s41597-019-0104-8

    2019 doi

  16. [24]

    Y. Wang, W. Duan, D. Dong, L. Ding, X. Lei, A test-retest rest- ing and cognitive state EEG dataset (2022). doi:10.18112/ openneuro.ds004148.v1.0.1

    2022

  17. [25]

    doi:10.18112/openneuro.ds002778.v1.0.5

    A.P.Rockhill,N.Jackson,J.George,A.Aron,N.C.Swann,UCSan Diego resting state EEG data from patients with Parkinson’s disease (2021). doi:10.18112/openneuro.ds002778.v1.0.5

    2021 doi

  18. [26]

    doi:10.18112/openneuro.ds003490.v1.1.0

    J.F.Cavanagh,EEG:3-StimauditoryoddballandrestinParkinson’s (2021). doi:10.18112/openneuro.ds003490.v1.1.0

    2021 doi

  19. [27]

    doi:10.18112/openneuro.ds004584.v1.0

    A.Singh,R.Cole,A.Espinoza,J.Cavanagh,N.Narayanan,Resteyes open(2023). doi:10.18112/openneuro.ds004584.v1.0. 0

    2023 doi

  20. [29]

    A. J. Bell, T. J. Sejnowski, An information-maximization approach to blind separation and blind deconvolution, Neural Comput. 7 (6) (1995) 1129–1159. doi:10.1162/neco.1995.7.6.1129

    1995 doi

  21. [30]

    Paszke, S

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, et al., Pytorch: An imperativestyle,high-performancedeeplearninglibrary,Adv.Neural Inf. Process. Syst. 32 (2019). doi:10.48550/arXiv.1912. 01703

    2019 doi

  22. [31]

    Brookshire, J

    G. Brookshire, J. Kasper, N. M. Blauch, Y. C. Wu, R. Glatt, D. A. Merrill, S. Gerrol, K. J. Yoder, C. Quirk, C. Lucero, Data leakage in deep learning studies of translational EEG, Front. Neurosci. 18 (2024). doi:10.3389/fnins.2024.1373515

    2024

  23. [32]

    Zanola, L

    A. Zanola, L. F. Tshimanga, F. Del Pup, M. Baiesi, M. Atzori, xEEGNet: Towards explainable AI in EEG dementia classification, arXiv Preprint (2025).doi:10.48550/arXiv.2504.21457

    2025 doi

  24. [33]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. u. Kaiser, I. Polosukhin, Attention is all you need, in: AdvancesinNeuralInformationProcessingSystems,Vol.30,Curran Associates, Inc., 2017, pp. 5998–6008.doi:10.48550/arXiv. 1706.03762

    2017 doi arXiv

  25. [34]

    Devlin, M.-W

    J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, BERT: Pre-training of deep bidirectional transformers for language understanding, in: Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, V...

    2019

  26. [35]

    Del Pup, A

    F. Del Pup, A. Zanola, L. F. Tshimanga, P. E. Mazzon, M. Atzori, SelfEEG: A Python library for self-supervised learning in electroen- cephalography, J. Open Source Softw. 9 (95) (2024) 6224.doi: 10.21105/joss.06224

    2024 doi

  27. [36]

    6 (60) (2021) 3021.doi:10.21105/joss.03021

    M.L.Waskom,seaborn:statisticaldatavisualization,J.OpenSource Softw. 6 (60) (2021) 3021.doi:10.21105/joss.03021

    2021 doi

  28. [37]

    brain mapp

    R.T.Schirrmeister,J.T.Springenberg,L.D.J.Fiederer,M.Glasstet- ter,K.Eggensperger,M.Tangermann,F.Hutter,W.Burgard,T.Ball, Deep learning with convolutional neural networks for EEG decoding and visualization, Hum. brain mapp. 38 (11) (2017) 5391–5420. doi:10.1002/hbm.23730

    2017 doi

  29. [38]

    K. H. Cheah, H. Nisar, V. V. Yap, C.-Y. Lee, G. Sinha, Optimizing residual networks and VGG for classification of EEG signals: Identi- fyingidealchannelsforemotionrecognition,J.Healthc.Eng.2021(1) (2021) 5599615. doi:10.1155/2021/5599615

    2021 doi

  30. [39]

    Grandini, E

    M. Grandini, E. Bagli, G. Visani, Metrics for multi-class classi- fication: an overview, arXiv Preprint (2020). doi:10.48550/ arXiv.2008.05756

    2020 doi arXiv

  31. [40]

    Neurosci

    E.Lashgari,D.Liang,U.Maoz,Dataaugmentationfordeep-learning- based electroencephalography, J. Neurosci. Methods 346 (2020) 108885. doi:10.1016/j.jneumeth.2020.108885

    2020

  32. [41]

    Rommel, J

    C. Rommel, J. Paillard, T. Moreau, A. Gramfort, Data augmentation for learning predictive models on EEG: a systematic comparison, J. Neural Eng. 19 (6) (2022) 066020.doi:10.1088/1741-2552/ aca220

    2022 doi

  33. [42]

    Le Guennec, S

    A. Le Guennec, S. Malinowski, R. Tavenard, Data Augmentation for Time Series Classification using Convolutional Neural Networks, in: ECML/PKDD Workshop on Advanced Analytics and Learning on Temporal Data, Riva Del Garda, Italy, 2016, p. 8. URL https://shs.hal.science/halshs-01357973

    2016

  34. [43]

    Lemkhenter, P

    A. Lemkhenter, P. Favaro, Boosting generalization in bio-signal classification by learning the phase-amplitude coupling, in: Pattern Recognition,SpringerInternationalPublishing,Cham,2021,pp.72–

    2021

  35. [44]

    F. N. Fritsch, J. Butland, A method for constructing local monotone piecewisecubicinterpolants,SIAMJ.Sci.Stat.Comput.5(2)(1984) 300–304. doi:10.1137/0905021

    1984 doi

  36. [45]

    Kingma, J

    D. Kingma, J. Ba, Adam: A method for stochastic optimization, in: 3rd international conference on learning representations (ICLR), San Diega, CA, USA, 2015, p. 13.doi:10.48550/arXiv.1412. 6980

    2015 doi

  37. [46]

    Lee, H.-R

    H.-T. Lee, H.-R. Cheon, S.-H. Lee, M. Shim, H.-J. Hwang, Risk of data leakage in estimating the diagnostic performance of a deep-learning-based computer-aided system for psychiatric disorders, Sci. Rep. 13 (1) (2023) 16633. doi:10.1038/ s41598-023-43542-8

    2023

  38. [47]

    M. H. Rafiei, L. V. Gauthier, H. Adeli, D. Takabi, Self-supervised learning for electroencephalography, IEEE Trans. Neural Netw. Learn. Syst. 35 (2) (2024) 1457–1471.doi:10.1109/TNNLS. 2022.3190448

    2024

  39. [48]

    Babiloni, R

    C. Babiloni, R. Lizio, N. Marzano, P. Capotosto, A. Soricelli, A. I. Triggiani, S. Cordone, L. Gesualdo, C. Del Percio, Brain neural synchronization and functional coupling in Alzheimer’s disease as revealed by resting state EEG rhythms, Int. J. Psychophysiol. 103 (2016) 88–10...

    2016 doi

  40. [49]

    TransformEEG: Towards Im- proving Model Generalizability in Deep Learning-based EEG Parkinson’s Disease Detection

    R. Nardone, L. Sebastianelli, V. Versace, L. Saltuari, P. Lochner, V. Frey, S. Golaszewski, F. Brigo, E. Trinka, Y. Höller, Usefulness of EEG techniques in distinguishing frontotemporal dementia from Alzheimer’s disease and other dementias, Dis. Markers 2018 (1) (2018) 6581490...

    2018 doi

  41. [85]

    doi:10.1007/978-3-030-71278-5\_6

    doi

  42. [4186]

    doi:10.18653/v1/N19-1423

    doi

Pith tools

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