{"paper_id":"d74f736b-00c0-4d8e-a6e3-607495f840f0","body_text":"1 \nReturn of the GEDAI:  \nUnsupervised EEG Denoising based on Leadfield Filtering  \n \nTomas Ros 1,2 *, Victor Férat 3, Yingqi Huang 2, Cristina Colangelo 1,2, Seyed Mostafa Kia 4, \nThomas Wolfers 5,6, Serge Vulliemoz 1,2, Abele Michela 1,2 \n \n1   Center for Biomedical Imaging (CIBM), Geneva-Lausanne, Switzerland \n2  Department of Clinical Neuroscience, University of Geneva, Geneva, Switzerland \n3  M/EEG & NEUROMOD Platform, Fondation Campus Biotech Geneva, Geneva,    \nSwitzerland \n4 Department of Cognitive Science and Artificial Intelligence, Tilburg University, Tilburg, the    \nNetherlands. \n5  Department of Psychiatry and Psychotherapy, University Hospital Tübingen, Tübingen, \nGermany \n6  German Center for Mental Health (DZPG), partner site Jena, Germany \n \n* Corresponding author: tomas.ros@unige.ch \nAbstract \n \nCurrent electroencephalogram (EEG) denoising methods struggle to remove the complex \nphysiological and environmental artifacts typical of real-world settings, which both hinders the \nisolation of true neural activity and limits the technology’s translational potential. We present \nthe Generalized Eigenvalue De -Artifacting Instrument (GEDAI), a novel algorithm for \ndenoising highly contaminated EEG. GEDAI employs leadfield filtering to selectively remove \nnoise and artifacts that diverge from a theoretically defined EEG forward model. This approach \noffers unique advantages over existing solutions, including 1) denoising of highly corrupt \nrecordings without “clean” reference data, 2) single -step correction of artifactual epochs and \nbad channels, 3) unsupervised detection of brain and noise components based on the signal \nand noise subspace alignment index (SENSAI). In ground-truth simulations with synthetic and \nempirical EEG contaminated with realistic artifacts (EOG, EMG, n oise), GEDAI globally \noutperformed leading denoising techniques based on principal component analysis (ASR) and \nindependent component analysis (IClabel, MARA), revealing large effect sizes in challenging \nscenarios with simultaneous artifact mixtures, low s ignal-to-noise ratio ( -9 dB), and high \ntemporal contamination (up to 100%). Its superior denoising also enhanced neurobehavioral \npredictions, yielding highest accuracies in ERP classification and brain fingerprinting. GEDAI’s \nautonomy, computational speed and noise-resilience could find future applications in 1) real-\nworld medical, mobile and dry electrode EEG recordings 2) magnetoenecephalography \n(MEG) denoising  (given the shared M/EEG forward model), and 3) real-time brain-computer \ninterfaces (BCIs). The Matlab code for GEDAI is available as an open-source EEGLAB plugin \nat https://github.com/neurotuning/GEDAI-master \n \n \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n2 \nIntroduction  \n \nDespite being well-researched and affordable, the multi-channel electroencephalogram (EEG) \nstruggles to expand beyond controlled research and clinical settings (Sawangjai et al., 2020). \nA major hurdle to broader EEG use is contamination by non -cerebral signals, known as \nartifacts (Mumtaz et al., 2021). EEG artifact removal often requires the supervision of a trained \noperator, leading to reduced scalability and rel iability. Hence, tackling artifacts has recently \nbeen rated by experts as the most important development needed in EEG research (Mushtaq \net al., 2024).  \nIn the realm of medical applications (Fratangelo et al., 2025) , EEG artifacts can \nobscure or mimic genuine brain activity, leading to inaccurate diagnoses and compromised \npatient care (e.g. in epilepsy, intensive care or sleep monitori ng) (Amin et al., 2023) . \nFurthermore, artifacts significantly hinder the development and reliability of brain -computer \ninterfaces (BCIs) (Mak & Wolpaw, 2009), which rely on consistent and accurate decoding of \nneural signals to translate brain activity into commands for external devices or neurofeedback \n(Ros et al., 2014). Artifacts introduce noise that can disrupt this decoding process, leading to \nreduced accuracy, slower response times, and ultimately, a less effective BCI system. \nTherefore, there remains a critical need for a fast, fully automated, and noise -agnostic EEG \nmethod to reliably remove all types of artifacts, particularly in highly contaminated recordings \nwhere clean reference data is unavailable. \nSeveral previous strategies have been developed to clean multi-channel EEG signals. \nThe most popular of which is indep endent component analysis (ICA), which performs blind \nsource separation and can be used to remove temporally-independent components from the \nsignal. While effective at removing ocular (Jung et al., 2000) and muscular artifacts (Olbrich \net al., 2011) , ICA is computationally demanding, requires a case -by-case inspection of \ncomponents to reject that calls for expert human input or that from machine learning classifiers \n(Frølich et al., 2015; Radüntz et al., 2017). Moreover, ICA cannot inherently identify Gaussian \ndistributed noise,  which might be spread across the components or  remain as unexplained \nresidual variance . ICA is therefore not an ideal choice for inexperienced users, large datasets \nor online analysis (despite promising attempts e.g. Hsu et al., 2014).   \nAnother algorithm, called artifact subspace reconstruction (ASR), directly uses \nprincipal component analysis (PCA) to exclude EEG components exceeding a variance \nthreshold, but requires a portion of clean EEG data as a reference (Kothe & Jung, 2015). While \nfaster, relying solely on a variance threshold is not per se sufficient to distinguish between \nsignal and noise (de Cheveigné & Parra, 2014), as in the case when noise amplitude is below \nthat of the signal. Moreover, since ASR is based on PCA, it can only separate orthogonal \ncomponents (Cohen, 2022a), which may limit its effectiveness when trying to resolve complex \nartifact mixtures.   \nTo overcome the limitations of ICA and PCA, Generalized Eigenvalue Decomposition \n(GEVD) offers a promising alternative (Koles, 1991; Y. Wang et al., 1999) . GEVD works by \ntaking a pair of EEG covariance matrices and decomposing them jointly – a process known \nas \"joint diagonalization\" – to find common underlying components (de Cheveigné & Parra, \n2014). The core objective of this joint diagonalization is to find a set of components that \nmaximizes the variance ratio between the signal and reference matrices (Cohen, 2 022a). \nAnother key advantage is that the resulting GEVD components do not need to be orthogonal \n(unlike PCA) or conform to Gaussian assumptions (unlike ICA) (Cohen, 2022a; Parra & Sajda, \n2003). \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n3 \nGEVD is widely employed in the BCI field for mental classification tasks, where it is \nknown as the common spatial patterns (CSP) algorithm (Blankertz et al., 2008; Lotte & Guan, \n2011). It has also been proposed for brain source separation (Cohen, 2022a; Parra & Sajda, \n2003), correction of ocular artifacts (Gouy-Pailler et al., 2009), removal of stimulation artifacts \n(Haslacher et al., 2021) and generic M/EEG data cleaning (Boudet et al., 2012; Somers et al., \n2018; F. Wang et al., 2025).  \nWe refer to this work as the “return” of the GEDAI (Generalized Eigenvalue De -\nArtifacting Instrument) because, while GEVD has been used in EEG denoising  before, our \napproach revisits it by addressing key limitations. An existing challenge in applying GEVD is \nthe need to specify the reference covariance matrix (refCOV) for decomposition. This typically \ninvolves selecting empirical EEG segments deemed \"clean\" (Haslacher et al., 2021; F. Wang \net al., 2025)  or \"artifactual\" (Gouy-Pailler et al., 2009; Somers et al., 2018) . However, this \nprocess is often subjective and circular since identifying representative segments requires \nprior knowledge of their typical characteristics. GEDAI addresses this by constructing the \nrefCOV theoretically, deriving it from EEG signal generation principles using the forward model \nvia the leadfield matrix (Weinstein et al., 2000) . This leadfield-based refCOV represents the \nexpected EEG signature of brain activity generated by internal sources, offeri ng a principled \nway to define the \"clean signal\" subspace for decomposition. We refer to this forward-modeling \napproach to separate brain signals from noise as leadfield filtering (LFF). An additional \nchallenge with GEVD is determining the cutoff that sepa rates 'signal' (brain activity) from \n'noise' (artifacts) components. GEDAI addresses this by using the theoretical refCOV as a \nbenchmark, selecting the optimal threshold based on subspace similarity between the cleaned \nEEG covariance and the refCOV (see Methods section).  \nTo evaluate GEDAI's performance, we tested it against prominent artifact correction \ntechniques using both synthetic and empirical EEG data, following recommendations for \nrobust testing (Mumtaz et al., 2021). This comparison includes the fa st PCA-based method \nASR (Kothe & Jung, 2015) and two widely used ICA -based methods known for effective \nresults: ICLabel (Pion-Tonachini et al., 2019) and Multiple Artifact Rejection Algorithm (MARA) \n(Winkler et al., 2014). This comparative analysis aims to  position GEDAI relative to current \nstate-of-the-art methods. \nMethods \nSection I of the methods describes the mathematical details of the GEDAI denoising \nalgorithm, whose code we also release as an open plugin for the EEGlab toolbox in MATLAB \n(Delorme & Makeig, 2004).  \nSection II  of the methods describes benchmarking using ground -truth simulations, \nwhere GEDAI’s performance is compared to current state -of-the-art algorithms for EEG de -\nartifacting, including ASR (Kothe & Jung, 2015) , ICLabel (Pion-Tonachini et al., 2019)  and \nMARA (Winkler et al., 2014) . The simulations were performed using both synthetic and \nempirical EEG datasets to which noise and/or artifacts were added.  \nSection III of the methods describes each algorithm’s denoising performance in the \ncontext of neurobehavioral prediction. Here, two publicly -available EEG datasets were used \nto compare how the denoising algorithms influence the prediction accuracy of sensory stimuli \n(“visual oddball”) and individual identity (“brain fingerprinting”). \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n4 \nSection I: The GEDAI algorithm \n \nFigure 1. An overview of the GEDAI algorithm \nA) Schematics illustrating the GEDAI pipeline; dataCOV: empirical data covariance matrix; refCOV: \ntheoretical covariance matrix from a leadfield model. B) An example dataCOV  of clean EEG data. C) \nAn example refCOV of a leadfield matrix. D) Graph illustrating similarity between the denoised data and \nthe refCOV across a range of artifact thresholds. SSSI: Signal Subspace Similarity Index, blue curve; \nNSSI: Noise Subspace Similarity Index, red curve; SENSAI: SSSI - NSSI, black curve; vertical dashed \nline: optimal artifacting threshold. \nAs illustrated in Figure 1A  above, multi -channel EEG may be considered to be a linear \nsummation of electrical activities from a brain “signal” subspace with one containing different \ntypes of non -cerebral noise or “artifact”. This mixture may be “unmixed” by linear \ndecomposition techniques (e.g. PCA or ICA) into separate components with individual source \nlocations (“topographies”) and respective time -courses (“waveforms”). However, as \"blind\" \nsource separation methods, PCA and ICA leverage statistical properties within mixed data to \nrecover underlying sources, functioning without a priori knowledge of the original signals or \ntheir mixing process. GEDAI combines prior knowledge of the brain’s “signal” subspace (i.e. \nits spatial covariance) with generalized eigenvalue decomposition (GEVD) to more effectively \nseparate source components belonging to the artifact -subspace from those of the brain -\nsubspace. \nFirst, refCOV is derived from a precomputed leadfield matrix for standard 10-5 system \nelectrode locations or, for non -standard electrode locations, from an interpolated leadfield \nmatrix calculated ‘on -the-fly’ via spherical interpolation. Next, the whole in put EEG signal \n(including artifacts) is epoched in circa 1-second windows, and an individual data covariance \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n5 \nmatrix (dataCOV) is generated for each epoch. Then, each data covariance matrix is \ndecomposed with GEVD, using refCOV as a fixed reference matrix a cross all epochs. After, \nthe output EEG data is reconstructed and evaluated using the Signal & Noise Subspace \nAlignment Index (SENSAI), in order to determine the optimal cutoff that separates brain from \nartifactual components, sweeping across a range of ar tifacting strengths. By respectively \nmaximising and minimizing the subspace similarities of the retained and removed data with \nthe refCOV, across different thresholding strengths, GEDAI estimates the optimal cutoff for \ncomponent removal. For the final step , the denoised time -series of each epoch is then \nreconstructed by using only the GEVD components belonging to the brain-subspace. \nEstimation of the refCOV \nThe GEDAI EEGlab plugin offers two options for refCOV estimation. The first method uses a \npre-computed covariance matrix of 343 standard EEG electrode locations (10-5 system), from \nwhich the plugin automatically matches the electrode labels present in the EEG recording (e.g. \nFp1, Pz, etc.). This leadfield matrix was generated with the Brainstorm Toolbox (Tadel et al., \n2011) using the OpenMEEG algorithm (Gramfort et al., 2010) and the ‘fsaverage’ adult head \nmodel (FreeSurfer’s default template based on 40 normative brains), employing the Boundary \nElement Method (BEM) with 3630 unconstrained brain dipolar sources (1210 vertices × 3 \norientations).  \nFor non -standard EEG recording montages, there is a second option: spherically \ninterpolating the precomputed leadfield to custom electrode locations. This method requires \nthat the electrodes’ spatial coordinates are provided within EEGlab. Although slightly less \naccurate, this approach is computationally much faster than estimating a custom BEM \nleadfield, taking only a couple of seconds compared to the several minutes required to \nrecompute with OpenMEEG. \nThe leadfield matrix parametrizes the “forward model” of how the EEG is generated by \nsources of neuronal activity in the brain, i.e., the EEG_ac tivity with dimensions [channels x \ntime], the leadfield_matrix with dimensions [channels x sources], and the \nbrain_source_activities with dimensions [sources x time]: \n \nFrom this, the spatial covariance matrix, refCOV, with dimensions [channels x channels] can \nbe simply computed as:  \n \nGEDAI's source -space to electrode -space projection relies on two main theore tical \nassumptions (de Munck et al., 1988, 1992): electrode potential is a weighted sum of underlying \ndipolar sources, and each source's strength is independent of other source parameters. Under \nthese conditions, de Munck et al. demonstrated a linear relationship between electrode signal \ncovariance and the summed, weighted covariance of individual dipolar sources. \nWe provide a simple empirical example supporting the aforementioned assumptions \nin Figure 1 where a leadfield -generated covariance matrix (panel C)  may be qualitatively \ncompared to the covariance matrix calculated from clean EEG data (panel B)  (Subject #2 \nfrom the Ehrlich dataset, see Methods). Both matrices were normalised [channel x c hannel] \n(i.e. zero mean with a standard deviation of one) for visualisation. \n \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n6 \nGeneralized Eigenvalue Decomposition  \nThreshold-based artifact rejection methods rely on the principle that artifacts are usually large \nand relatively rare events occurring in th e signal. Artifacts can therefore be identified as \nappearing in the upper tail of the eigenvalue distribution of EEG signals, and be excluded \nbased on a defined cutoff (Lazarevic & Kumar, 2005; Li et al., 2023). The large magnitude of \nartifacts is exploited by PCA -based procedures such as ASR, but these are constrained by \northogonality and blind to the spatial origin of the signal components.  \nIn contrast, GEDAI employs GEVD to isolate and remove artifactual components from \nthe signal. However, GEVD alters the significance of the largest eigenvalues. Instead of \nrepresenting \"components with high variance,\" they now signify \"components with high \nvariance that maximally deviate from the reference co variance matrix (refCOV).\" Given that \nthe refCOV encodes components that spatially originate only within the brain, this a priori \nprovides GEVD with extra “supervision” for separating artifactual from neural sources. \nThe GEVD may be summarised in one line of MATLAB/Python code, where e is the epoch \nnumber: \n \nMathematically, the GEVD decomposition follows the linear algebra equation: \n \nWhere dataCOV is the covariance matrix of a single epoch extracted from the EEG data, \nrefCOV is the predetermined reference covariance matrix, Λ is the diagonal matrix containing \nthe generalized eigenvalues and V is the matr ix containing the generalized eigenvectors. In \nGEDAI, a regularization technique is applied to refCOV before GEVD to enhance numerical \nstability, manage singular matrices, and reduce overfitting (Cohen, 2022a). \n \nThe SENSAI algorithm: separating neural from artifactual components \nAkin to finding the boundary that separates oil from water, differentiating artifacts from neural \nsignals depends on the exact magnitude of the eigenvalues, represented by the threshold T. \nSENSAI (Signal & Noise Subspace Alignment Index) essentially estimates the correct \nthreshold T by evaluating the “improvement” in the output EEG signal quality by sweeping \nover multiple threshold values of T. Here, the output EEG signal quality is represented by the \nSENSAI score, which is proportional to the cosine similarity between the refCOV matrix and \nthe empirical covariance of the retained (i.e. Signal) and the removed (i.e. Noise) data.  \nAs can be seen in Figure 1D, the SENSAI function computes a similarity index of the \nretained and removed EEG data for each threshold tested. The similarity index is calculated \nfirst by taking the covariance matrix of the cleaned EEG data and the refCOV, and performing \na classical ei gendecomposition separately on both (i.e. equivalent to PCA). The principal \nangles between the top components (i.e. eigenvectors) of the refCOV and cleaned data \nsubspaces are then estimated (Knyazev & Argentati, 2002). Here, given two subspaces with \ntheir orthonormal bases A (from refCOV) and B (from cleaned data), the principal angles θ i \nbetween these subspaces can be computed as follows: \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n7 \nCompute the projection matrix C:   \nPerform Singular Value Decomposition on the projection matrix :  \nwhere W is related to the orthonormal basis of the subspace A, Q is related to the orthonormal \nbasis of the  subspace B, and Σ is a diagonal matrix containing the cosines of the principal \nangles between the two subspaces. The principal angles θ i are obtained from the singular \nvalues σi in the diagonal matrix Σ: \n \nTo estimate the Signal Subspace Similarity Index (SSSI), we take the cosine of each principal \nangle and multiply them together: \n \nWhere 𝜃1, 𝜃2 and 𝜃3 are the principal angles between the top 3 eigenvectors of the cleaned \ndata and the top 3 eigenvectors of the refCOV. Similarly, the Noise Subspace Similarity Index \n(NSSI) is also estimated between the top 3 eigenvectors of the residual noise data (removed \nfrom the EEG) and the refCOV. By estimating the SSSI and NSSI through a broad range of \nartifacting thresholds, it becomes possible to optimize the tradeoff between either removing or \nretaining too many artifactual components. For example, increasing the threshold strength at \nfirst increases the similarity of the cleaned EEG covariance with refCOV, as non -cerebral \ncomponents are initially removed. Past a certain point, however, the threshold becomes too \naggressive and cerebral activity starts to be removed from the EEG recording, thus reducing \nthe similarity of the cleaned EEG covariance with refCO V. Similarly, by calculating the NSSI \nfor the data being removed from the EEG in the cleaning process, we can quantify how much \npotentially cerebral EEG activity gets included as noise.  \nSENSAI calculates a tradeoff score, called the SENSAI score, by subtr acting the subspace \nsimilarity index of the noise from the similarity index of the cleaned EEG, or mathematically: \n \nThe optimal threshold for cleaning the dat a is simply the eigenvalue that corresponds to the \nmaximum of this score. \n \nDenoised signal reconstruction \nCleaned EEG signal reconstruction occurs epoch -by-epoch. Within each epoch, artifactual \ncomponents are first identified (eigenvalues > T), and a spati al filter targeting these artifacts \nis formed using the corresponding eigenvectors. Concurrently, the activation patterns (or \ninverse spatial filters) for these artifactual components are derived from the pseudo-inverse of \nthe eigenvector matrix. The artifactual activity is then estimated in sensor space by combining \nthese activation patterns with their respective time courses (calculated using the spatial filter \nand the original EEG). Finally, subtracting this reconstructed artifact activity from the origi nal \nepoch yields the denoised EEG signal. To prevent signal discontinuities that can arise from \nconcatenating independently processed EEG epochs, GEDAI employs a 50% overlapping \nwindow approach during its epoch-by-epoch reconstruction. These overlapping segments are \ncombined using cosine weighting to ensure smooth transitions at epoch boundaries, mitigating \nthe edge effects inherent in segmented processing. \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n8 \n   \nMultiresolution GEDAI \nAlthough GEDAI can be applied end-to-end solely to broadband data (e.g. 1-45 Hz), we have \nfound that multiresolution -based denoising performs significantly better on pilot data. \nMultiresolution analysis decomposes the signal into frequency sub -bands across multiple \nscales, which permits perfect reconstruction, ensuring lossless re covery of the original EEG \ndata from its components (Clark et al., 1995). Hence, following an initial GEDAI denoising step \non broadband EEG to remove the largest artifacts, the GEDAI cleaning al gorithm is \nsubsequently applied to spectrally decomposed EEG data using a maximum overlap discrete \nwavelet transform (MODWT). The benefit of this extra step comes from the fact that artifact \ndistributions clearly differ between EEG bands (ranging from delt a to gamma). Thus, \nfrequency-specific thresholds are needed to ensure that high -frequency artifacts (e.g. in the \ngamma band) are not rejected using the same eigenvalue threshold estimated for lower \nfrequency bands (e.g. the delta band). Hence, the GEDAI al gorithm consists of a secondary \nstep on MODWT decomposed signals filtered in 10 dyadic frequency bands, using the ‘Haar’ \nwavelet. Among the wavelet families, the Haar wavelet has the best temporal localization, \nwhich is an important property for pinpointing abrupt changes, discontinuities, or spikes typical \nof artifactual signals. Compared to other types of wavelet analysis, the MODWT offers \nadvantages like conservation of signal energy (perfect reconstruction) and time alignment \n(zero-phase filtering, shif t invariance) across decomposition levels. For an EEG recording \nsampled at 256 Hz, the wavelet bands produced by the MODWT are as follows: 64 -128 Hz \n(band 1), 32-64 Hz (band 2), 16 -32 Hz (band 3), 8 -16 Hz (band 4), 4 -8 Hz (band 5), 2 -4 Hz \n(band 6), 1-2 Hz (band 7), 0.5-1 Hz (band 8), 0.25-0.5 Hz (band 9), 0.125-0.25 Hz (band 10). \nThen, for each narrow -band the optimal artifacting threshold is estimated separately via the \nSENSAI algorithm, and the final clean signal is reconstructed by adding up all the deno ised \nwavelet bands. \nSection II: Benchmark EEG simulations \nUsing EEGlab v2025.0.0 we compared the GEDAI algorithm to 3 state -of-the-art (SOTA) \ndenoising algorithms (ASR, MARA and ICLabel) in terms of ground-truth signal reconstruction \non both synthetic and empirical EEG data.  \nDenoising algorithms: ASR, IClabel, MARA, GEDAI \nThree established automated denoising algorithms were used to benchmark GEDAI, namely \nASR, IClabel and MARA. Like GEDAI, these are all projection algorithms that “correct” artifacts \nby subtracting them from EEG data, and which technically incur no temporal data loss. A \npanoply of pipelines also exists for denoising EEG data that involve wholesale “rejection” of \nEEG segments, but these are designed to solve a different problem, where full recovery of the \noriginal EEG signal is not the objective (Bailey et al., 2023; Gabard -Durnam et al., 2018; \nHajhassani et al., 2024). We have also not included deep learning based denoising algorithms \n(X. Zhang, 2024) , as these techniques rely on extensive training datasets and can exhibit \nlimited generalizability when applied to out-of-sample data. This could be investigated in future \nwork and is beyond the scope of this paper. \n \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n9 \nRaw \nAs a sanity-check, we provide the “raw” unprocessed data for comparison, which was used \nas the direct input to each denoising algorithm. Hence, the raw condition did not contain any \ndenoising and/or preprocessing, and consisted of the original EEG contaminated with artifacts. \nASR \nAs the first denoising algorithm, we included ASR (i.e. EEGlab’s clean_raw_data v2.91), as it \nis a fast, real -time capable method for EEG cleaning (Kothe, 2013) . The ASR algorithm \nperforms a standard eigendecomposition using one covariance matrix as input (i.e. PCA), \ninstead of a generalized eigendecomposition that utilizes two covariance matrices as input \n(i.e. GEVD). We used EEGlab’s ASR plugin default burst criterion value of 20 SD (Chang et \nal., 2020) . Given that bad (i.e. noisy) channels negatively affect ASR performance, we \nactivated clean_raw_data’s bad chan nel detection only for artifact categories where noisy \nchannels were present, i.e. NOISE or NOISE + Electromyography (EMG) + \nElectrooculography (EOG). Since bad channel detection was not universally applied over all \nartifact categories, the benchmark was not fully \"noise-agnostic” and slightly advantaged ASR. \nIClabel \nThe second denoising algorithm was ICLabel v1.6 (Pion-Tonachini et al., 2019) . This ICA -\nbased technique can automate artifact rejec tion by evaluating ICA topographies with a \nmachine learning based classifier trained on ICA topographies from over 6,000 datasets. Here, \nthe Infomax -extended ICA algorithm (Lee et al., 1999)  was used based on its superior \nperformance compared to other ICA algorithms (Delorme et al., 2012). Firstly, given that bad \nchannels may impair the quality of ICA decomposition, the ICA was pre ceded by EEGlab’s \nclean_raw_data bad channel rejection step only for artifact categories where noisy channels \nwere present (i.e. NOISE or NOISE + EMG + EOG). Moreover, if the artifact category \ncontained only EMG or EOG artifacts, IClabel was configured to remove only those ICA \ncomponents identified as EMG or EOG artifacts (all other components were retained). On the \nother hand, if the artifact category contained noise (i.e. NOISE or NOISE + EMG + EOG), \nIClabel was configured to only retain components that w ere classified as ‘brain’ or ‘other’. \nSince the above settings were not universally applied over all artifact categories, the \nbenchmark was not fully \"noise-agnostic” and provided a marginal advantage to IClabel. \n \nMARA \nFor the third denoising  algorithm, we selected MARA v1.2, the Multiple Artifact Rejection \nAlgorithm (Winkler et al., 2014). This ICA-based technique also involves a machine learning \nclassifier trained on a large datab ase of expert -labeled artifactual topographies. MARA is \nintegrated within popular denoising pipelines, such as the Harvard Automated Processing \nPipeline for Electroencephalography (HAPPE) (Gabard-Durnam et al., 2018). Unlike IClabel, \nMARA directly classifies ICA components as artifactual (or not), and hence only those \ncomponents were removed. To facilitate comparisons to IClabel, the Infomax -extended ICA \nalgorithm was used (Lee et al., 1999), which was preceded by EEGlab’s clean_raw_data bad \nchannel rejection step only for artifact categories where noisy channels were present (i.e. \nNOISE or NOISE + EMG + EOG). Since bad channel detectio n was not universally applied \nover all artifact categories, the benchmark was not fully \"noise -agnostic”, which slightly \nfavoured MARA’s performance. \n \n \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n10 \nGEDAI \nWe executed EEGlab’s GEDAI v1.0 plugin using its default settings: with denoising strength \nset to ‘auto’, an epoch size of 1.0 seconds, and a ‘precomputed’ BEM lead field matrix. No \nbad channel rejection step was applied to any of the artifact categories. The above settings \nwere universally applied across all datasets, hence the GEDAI benchmark results  can be \nconsidered as effectively \"noise-agnostic\". \n \nDenoising scenarios: signal-to-noise ratio, temporal contamination, artifact type  \nEach denoising algorithm was run with its default parameter settings across the complete \nrange of noise scenarios descri bed below. Fixing the parameter settings was necessary to \ntest each algorithm’s automation and/or generalization to a range of real -world “noise” \nscenarios, which were varied across 3 key axes:  \n \n● Signal to Noise Ratio (SNRbefore ) of clean EEG power relative to artifact power (-9, -\n6, -3, 0 dB) \nwhere SNRbefore = 10 * log10(original_clean_signal_power / original_artifact_power) \nwas estimated across all epochs (i.e. the whole EEG recording) \n● Temporal contamination of EEG contaminated by artifacts (25, 50, 75,100%) \nwhere % reflects the proportion of samples containing artifacts. For example, for a \ndataset of 60 seconds and sampled at 200 Hz, we added artifact segments with \nrandom offsets into a total of 4000, 6000, 8000, or all 12000 samples. The artifact \nsegment duration was randomized, ranging from 1 sample to 1 second.. \n● Artifact type (EOG, EMG, NOISE, or their combination), where different artifact types \nwere linearly superimposed within each epoch. \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n11 \nDenoising Benchmark: Synthetic and Empirical EEG Simulations \n \nFigure 2. Simulated EEG data generation \nBackground “ground -truth” EEG ( signal) was obtained either from clean empirical recordings or \nsynthetic data generated by a forward model with 100 -300 sources. Contaminated ( simulated EEG) \ndata were generated by a sensor -space summation of the signal (blue trace) with different types of \nartifacts (red trace) (Ille et al., 2002). In this example, muscle artifact was generated by multiplying the \ntime-series (“waveforms”) of EMG electrode activity with source locations (“topographies”) consistent \nwith EMG artifact. Artifact waveforms were derived from the EEGdenoiseNet dataset, which consisted \nof direct recordings from individual EO G and EMG electrodes, respectively placed over the eyes and \nfacial muscles (H. Zhang et al., 2021). \n \nSynthetic EEG simulations \nSynthetic data for these simulations were generated using a combination of the EEG forward \nmodelling toolbox SEREEGA (Krol et al., 2018)  and the EEGdenoiseNet dataset based on \nreal EOG and EMG electrode recordings (H. Zhang et al., 2021). \nClean EEG signal generation \nHere, we exclusively used the SEREEGA toolbox to generate clean “background” EEG data. \nA total of 10 independent datasets were simulated with distinct spatio -temporal dynamics, \nincluding: 2 datasets with autoregressive noise, 6 datasets with varying combinations of \ncolored noise (white, blue, brown) plus amplitude-modulated oscillations (in delta, theta, alpha \nand beta bands), and 2 datasets with simulated epileptic spiking activity (small or large spikes). \nThe forward model consisted of the New York head model (i.e. ICBM152 average template) \nwith a precomputed Finite Element Method (FEM) leadfield (Huang et al., 2016). Here, 100-\n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n12 \n300 dipolar sources, randomly distributed within the brain volume, were used to generate a \nmulti-channel time series across 100 independent 1-second epochs. Each output dataset had \na duration of 100 seconds and consisted of 6 4 channels (standard Biosemi montage) with a \nsampling frequency of 256 Hz. See Fig S1 for examples of clean synthetic EEG that was used. \nEOG and EMG artifacts \nIn order to make the synthetic artifacts as realistic as possible, we used the real -world \nrecordings of sensor data directly overlying the eyes (EOG, electro -oculogram) and facial \nmuscles (EMG, electromyogram) from the EEGdenoiseNet dataset \nhttps://github.com/ncclabsustech/EEGdenoiseNet (H. Zhang et al., 2021).  Then, as shown in \nFigure 2, these EOG/EMG single time-series (“waveforms”) were multiplied by dipolar source \npatterns (“topographies”) to generate multi -channel EEG in sensor-space. In order for these \ntopographies to be consistent with EOG/EMG generation, we utilized the HArtMuT head model \n(Harmening et al., 2022), which enables simulating EEG activity originating from extra-cerebral \nsources located in the eyes and muscles. Here, we extracted a random sample of 16 eye and \n10 scalp muscle source locations (as illustrated in Figure S3). Then, for each dataset, a \nminimum of 2 sources of EOG and 1 -3 sources of EMG  were randomly selected and their \ntopographic information (i.e. source location and orientation) was retained. Finally, these \nEOG/EMG topographies were respectively multiplied by the EOG/EMG single time -series \nfrom EEGdenoiseNet (with 50 randomly sampled e pochs of 2 seconds) to construct multi -\nchannel datasets containing only artifacts. All EOG signals were band -pass filtered between \n0.3 and 10 Hz. EMG signals were band-pass filtered between 1 to 120 Hz and notched at the \npowerline frequency of 50 Hz. See Fig S2 for examples of artifactual semi-synthetic EEG that \nwas used. \n \nNOISE artifacts \nDatasets with noise artifacts were identical in size to those with EOG/EMG, and were \ngenerated to contain the following types of synthetic (i.e. MATLAB simulated) “noise” across \n5 datasets: with 3 bad channels (white noise, 1 dataset), 5 bad channels (sawtooth wave, 1 \ndataset), 6 bad channels  (square wave, 1 dataset), spike artifact (synthetic, 1 dataset), \nimpulsive noise (1 dataset).  \n \nArtifact mixtures \nIn order to challen ge the denoising algorithms with more complex scenarios of artifact \nmixtures, the baseline EOG, EMG, and/or NOISE datasets were z -score normalised and \nlinearly combined with each other to produce a mixed artifact category: “ NOISE + EMG + \nEOG ”. To generate a total of 5 datasets in this category, only datasets with matching numbers \nwere added across categories (e.g. EOG dataset 1 + EMG dataset 1 + NOISE dataset 1). \nIn summary, 5 separate artifact datasets were generated for each of the following 4 artifact \ncategories: \n1. EMG  (muscle activity) \n2. EOG (eye blinks and/or movement) \n3. NOISE (empirical or synthetic noise) \n4. NOISE + EMG + EOG \n \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n13 \n \nBad channels \nOut of the above total of 20 artifactual datasets, there were 6 datasets containing bad channels \n(3 datasets in the NO ISE category, and 3 datasets in the NOISE + EMG + EOG category). \nThus, bad channels were present in 30% (6/20) of all synthetic simulations. \nMixed clean EEG + artifact signal generation \nIn line with the model shown in Figure 2, each of the 10 clean recordings was pairwise “mixed”, \nthrough linear summation, with each of 5 datasets from a specific artifact category. This \nresulted in 50 contaminated datasets within each one of the 4 artifact categories  (EMG; EOG; \nNOISE; and EMG + EOG + NOISE).  \n \nThus, the synthetic simulations consisted of 200 unique datasets (50 datasets x 4 artifact \ncategories), varied across 4 temporal contamination levels x 4 SNR levels, yielding a \n“synthetic” benchmark sample of 3,200 datasets .  Each dataset had 64 channels and a \nlength of 100 seconds, sampled at 256 Hz (i.e. each data matrix: 64 channels x 25,600 \nsamples).  \n \n \nEmpirical EEG simulations \nClean EEG signal generation \nHere, we used Stefan Ehrlich’s experimentally-acquired recordings of clean EEG  \nhttps://github.com/stefan-ehrlich/dataset-automaticArtifactRemoval (Ehrlich, 2019/2024) , \nwhere 10 separate subjects were selected from the Ehrlich dataset in the ‘eyes-closed’ resting \ncondition with a duration of 100 seconds. These recordings were visually inspected for the \npresence of any artifacts, and any samples that were suspected of contamination were \nwholesale rejected (i.e. removed across all channels). For every subject, clean EEG segments \nof at least 2 seconds were concatenated and low-pass filtered at 60 Hz.  Each output dataset \nhad a duration of 60 seconds  and consisted of 27 channels with a sampling frequency of 200 \nHz. See Fig S4 for examples of clean empirical EEG that was used. \nArtifact signal generation \nEOG and EMG artifacts \nEven dedicated scalp EEG recordings of extra -cerebral artifacts inevitably contain brain \nsignals. Consequently, ground-truth brain and artifact signals cannot be established, which is \nnecessary for reliably evaluating denoising performance. To address this problem, we used \nempirical recordings of sensor data directly overlying the eyes (EOG, electro-oculogram) and \nfacial muscles  (EMG, electromyogram) from the EEGdenoiseNet dataset \nhttps://github.com/ncclabsustech/EEGdenoiseNet (H. Zhang et al., 2021).  Then, as shown in \nFigure 2, the EOG/EMG single time-series (“waveforms”) can be multiplied by their relevant \nspatial patterns (“topographies”) to generate multi -channel EEG in sensor -space. For these \ntopographies to be consistent with EOG/EMG generation,  we separately performed (Infomax-\nextended) ICA on multichannel EOG/EMG artifact recordings from the Ehrlich dataset (Ehrlich, \n2019/2024). Here, for each dataset, a minimum of 2 components of EOG and 4 components \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n14 \nof EMG were manually selected and only their topographic information was retained. Finally, \nas described above, these EOG/EMG topographies were respectively multiplied by the \nEOG/EMG single time-series from EEGdenoiseNet (with 30 randomly sampled epochs of 2 \nseconds) to construct multi -channel datasets containing only artifacts. The use of ICA in \ngenerating the artifactual EEG data could raise concerns about circularity (i.e. \"double \ndipping”), potentially favouring ICA -based algorithms (IC label, MARA). Such a bias would \nindeed be possible if we had used the independent, non-Gaussian time courses generated by \nICA itself (i.e. the time -courses matching the Ehrlich dataset topographies). However, we \ncircumvented this issue by instead using dir ect EOG/EMG sensor recordings from the \nEEGdenoiseNet dataset as the artifact time courses. These physiological signals were \nobtained without imposing ICA's assumptions of independence or non -Gaussianity, thereby \nensuring a fairer evaluation of the ICA -based algorithms. All EOG signals were band -pass \nfiltered between 0.3 and 10 Hz, and then re-sampled to 200 Hz. EMG signals were band-pass \nfiltered between 1 to 120 Hz and notched at the powerline frequency of 50 Hz, and then \nresampled to 200 Hz.  See Fig S5 for examples of artifactual empirical EEG that was used. \n \nNOISE artifacts \nDatasets with noise artifacts were identical in size to those with EOG/EMG, and were \ngenerated to contain the following types of empirical (e.g. based on Stefan Ehrlich’s empirical \nrecordings) and synthetic (i.e. MATLAB simulated) “noise” across 10 datasets: with 3 -7 bad \nchannels (synthetic, 4 datasets), movement artifact (empirical, 1 dataset), electrode pop  \n(empirical, 1 dataset), step artifacts (synthetic, 1 dataset), temporally non-stationary noise \n(synthetic, 1 dataset), and spatially non-stationary noise (synthetic, 2 datasets). See Fig S6 \nfor examples of noisy empirical EEG that was used. \n \nArtifact mixtures \nIn order to challenge the denoising  algorithms with more complex scenarios of artifact \nmixtures, the baseline EOG, EMG, and/or NOISE datasets were z -score normalised and \nlinearly combined to produce a mixed artifact category: “NOISE + EMG + EOG ”. To generate \na total of 10 datasets in this category, only datasets with matching numbers were added across \ncategories (e.g. EOG dataset 1 + EMG dataset 1 + NOISE dataset 1). \nIn summary, 10 separate artifact datasets were generated for each of the following 4 artifact \ncategories: \n1. EMG  (muscle activity) \n2. EOG (eye blinks and/or movement) \n3. NOISE (empirical or synthetic noise) \n4. NOISE + EMG + EOG \nBad channels \nOut of the above total of 40 purely artifactual datasets, there were 8 datasets containing bad \nchannels (4 datasets in the NOISE category, and 4 datasets in the NOISE + EMG + EOG \ncategory). Thus, bad channels were present in 20% (8/40) of all empirical simulations. \nMixed clean EEG + artifact signal generation \nFor the final step, and according to the model shown in Figure 2 above, each of the 10 clean \nrecordings was pairwise “mixed”, through linear summation, with each of 10 datasets from a \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n15 \nspecific artifact category. This resulted in 100 contaminated datasets within each one of the 4 \nartifact categories  (EMG; EOG; NOISE; and EMG + EOG + NOISE).  \n \nThus, the empirical simulations consisted of 400 unique datasets (100 datasets x 4 artifact \ncategories), varied across 4 temporal contamination levels x 4 SNR levels, yielding an \n“empirical” benchmark sample of 6,400 datasets . Each dataset had 27 channels a nd a \nlength of 60 seconds, sampled at 200 Hz (i.e. each data matrix: 27 channels x 12,000 \nsamples).  \n \n \nMeasures of denoising performance:  SNR, relative root mean square error \n(RRMSE) and correlation coefficient \nHere, clean EEG recordings served as the ground truth. Denoising algorithms were evaluated \non their accuracy in recovering this signal from contaminated data using 3 standard metrics, \nallowing comparison with prior work (Mumtaz et al., 2021). \n● SNR: higher is better \n \n \nwhere:  original_clean_signal_power is the variance of the original clean signal elements: \n \nresidual_artifact_power is the variance of the error signal (difference between denoised and \noriginal clean signal elements): \n \n● Relative Root Mean Square Error (RRMSE): lower is better \n \n \n \nwhere: RMSE is the standard deviation of the error signal (difference between denoised and \noriginal clean signal elements) and RMS is the amplitude of the original clean signal \nelements. \n                       \n● Correlation Coefficient (R): higher is better  \n \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n16 \n \nwhere:   \n is the mean of the original clean signal elements \nand \nis the mean of the denoised signal elements \n \nStatistical Analysis  \nTo compare the benchmark performance of GEDAI against the other algorithm s (ASR, \nICLabel, MARA) paired, non -parametric, statistical tests were employed due to a) the same \ndataset(s) were denoised by all algorithms, and b) the potential non -normal distribution of \nperformance metrics. The primary performance metrics evaluated wer e the Correlation \nCoefficient R, SNR, and RRMSE) between the denoised signal and the ground -truth clean \nsignal. Statistical significance was set at an alpha level of 0.05. Here, we used a Friedman \ntest (i.e. a non-parametric equivalent of a repeated-measures ANOVA) to assess if there was \na statistically significant difference in the distributions of denoising performance scores \nbetween the algorithms (GEDAI, ASR, ICLabel, MARA) when applied to the same \ncontaminated datasets (varied by SNR before, temporal contamination, and artifact type). \nFollowing a significant two -sided Friedman test result (i.e. p < 0.05) on SNR after, post -hoc \npairwise two -tailed comparisons between algorithms were conducted using MATLAB's \nmultcompare function with a Bonferro ni multiple-comparison correction to control for family -\nwise error rate. Effect sizes for each pairwise comparison were reported using the \nstandardized Z-score effect size  (r = Z/√N ), where Z is the Z -score from the Wilcoxon test \nand N is the number of pairs (where r ≈ 0.1 is small, r ≈ 0.3 is medium, and r ≈ 0.5 is large). \nAll statistical analyses were performed using MATLAB (v2024b). \nSection III: Neurobehavioral Prediction \nThis section investigated GEDAI's denoising performance using real-world EEG data, moving \nbeyond simulated environments. While genuine ground-truth EEG data is unobtainable in such \ncircumstances, EEG can still be used to predict specific objective information from an \nexperiment. Our central hypothe sis was that superior denoising would preserve the crucial \nneural information necessary for more accurate predictions. We tested this hypothesis across \ntwo distinct datasets and scenarios below. \n \n \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n17 \nClassification of visual stimuli based on single -trial Event Related Potentials \n(ERP)  \nHere, we used publicly available data ( https://zenodo.org/records/7495536) from a visual \noddball study by Omejc and colleagues (Omejc et al., 2023) . A total of 70 healthy adult \nindividuals passively viewed frequent (white square) and rare (Einstein's face) stimuli while \nbeing recorded with a 32 -channel EEG sampled at 256 Hz. Each participant was presented \nwith 124 frequent (84%) and 23 rare (16%) stimuli. For full details of the experiment please \nsee (Omejc et al., 2023) . After 1 -40 Hz band -pass filtering and average referencing the \ncontinuous EEG data, we ran each denoising algorithm with its default settings (as described \nin the EEG simulations) without any bad epoch rejection. Accordingly, ASR’s bad channel \nrejection step was enabled for this algorithm, as well as prior to running IClabel or MARA. No \nexplicit bad channel rejection was utilized for GEDAI. For IClabel, only components flagged \nas \"brain\" or \"other\" were retained as clean EEG data.  The open MATLAB analysis code \n(https://github.com/NinaOmejc/VEP_classification_aging) was then utilized for the extraction \nof single-trial ERP features (Omejc et al., 2023) . Time-independent statistical ERP features \nwere extracted ba sed on four ERP components: P1, N170, P2, and P3. Each of these \ncomponents was parameterized by four metrics: peak amplitude, mean amplitude, peak \nlatency, and fractional 50% peak latency. After feature selection based on mutual information, \neight time-independent features were used for classification, specifically the peak amplitudes \nand fractional 50% peak latencies at four electrode clusters (occipital, parietal, central, and \nfrontal).  \nWe performed the binary classification task by concatenating all subjects’ trials together, and \nused a linear support vector machine (Matlab’s fitcsvm function) with tenfold cross-validation, \naveraging the validation scores over 100 runs. In order to assess each algorithm’s full potential \nfor signal recovery, we did not perform any “bad” ERP trial rejection before or after denoising. \nHence, all trials were classified based on whether they contained a frequent or infrequent \nstimulus. Given the strong class imbalance, classification performance was reported as the \narea under the curve (AUC) of the receiver operating characteristic (ROC), averaged across \nall subjects. DeLong’s test (DeLong et al., 1988)  was used to estimate the statistical \nsignificance of AUC between models, with Bonferroni multiple comparison correction. \n \n \nSubject identification based on brain fingerprinting  \nHere, we used each person's unique resting-state EEG activity to identify them among a group \nof 100 individuals, also known as “brain fingerprinting”. For this, we utilised the Dortmund Vital \nStudy dataset, freely accessible at OpenNeuro \n(https://doi.org/10.18112/openneuro.ds005385.v1.0.3). This dataset contained EEG \nrecordings of adult subjects (mean age: 44 years) in a test -retest design, where the retest of \nthe same individuals was rec orded at 5 years follow -up. From this dataset, we included the \nfirst n=100 subjects that contained recordings of both Session 1 (“test”) and Session 2 (“retest \n“). Each test and retest file consisted of 180 seconds of eyes-open resting state before a \ncognitive task, recorded with 62 channels and a sampling rate of 100 Hz. After 1-45 Hz band-\npass filtering and average referencing, all EEG recordings were then denoised with ASR, \nIClabel, MARA, or GEDAI using default settings (identical to the EEG simulations) . Hence,  \nbad channel rejection was enabled for ASR, IClabel and MARA, but not GEDAI. For IClabel, \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n18 \nonly components flagged as \"brain\" or \"other\" were retained as clean EEG data. In order to \nconduct fingerprinting, we first calculated the spatial covariance matrix (i.e. channel x channel) \nof the denoised data using MATLAB’s cov function. Then, the correlation distance (with range \n-1 to +1) between the Session 1 recordings’  and Session 2 recordings’ covariance matrices \nwas calculated using MATLAB’s pdist2 function. The values from this square distance matrix \nwere subtracted from 1, yielding a 100 x 100 similarity matrix between Sessions 1 (test) and \nSessions 2 (retest), also known as the Identifiability matrix (Amico & Goñi, 2018) . The \nIdentifiability matrix has subjects as rows and columns, and encodes the information about \nthe self-similarity (Iself, main diagonal elements) of each subject with themself, across the \ntest/retest sessions, and the similarity of each subject with the others (or Iothers, off-diagonal \nelements). The overall goal is to predict the identity of a single subject from Session 1 based \non the similarity of that recording with multi -subject data from Session 2. As a measure of \nfingerprinting performance, we used the success rate (Sorrentino et al., 2021)  from code \navailable at https://github.com/eamico/Clinical_fingerprinting/blob/master/FC_fingerprint.m. \nThe success rate (%) indicates how many times an Iself value is higher than the Iothers values \non the same row and column of the Identifiability matrix. In other words, th e success rate \nrepresents the percentage of the number of individual off-diagonal comparisons when the self-\ncorrelation “wins”, which is a more granular comparison rather than a binary correct/incorrect \nprediction of the subject. To estimate the variabilit y in fingerprinting performance for each \nalgorithm, we performed 50 random resamples of the denoised EEG data. Each resample \nconsisted of a total of 10 seconds (i.e. 1000 random data samples), from which the spatial \ncovariance matrix was calculated and the  success score estimated.  Following a significant \nKuskall-Wallis test result (p <0.05), post -hoc comparisons between algorithms were \nconducted using MATLAB's multcompare function with a Bonferroni correction to control the \nfamily-wise error rate. \n \nHardware and Software Platform \nAll computations were carried out using MATLAB software v2024b running on a Windows \ndesktop PC with an Intel-Core Ultra 9 285K CPU with 24 parallel cores and 64GB of RAM. All \ntests were performed with the CPU using parallel processing, with no GPU acceleration. \n \n \n \n \n \n \n \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n19 \nResults \nHere, we compared the GEDAI algorithm to three established denoising algorithms (ASR, \nMARA and ICLabel) in terms of ground -truth (i.e. clean background signal) recovery. We \nprincipally report the SNR after and RRMSE effect sizes. Matching figures for RRMSE can be \nfound in Fig. SX of the Supplementary Results. For the Correlation Coefficient R, we show its \nvalues in the global comparisons at the end of each section.  \nDenoising Benchmark of Synthetic EEG  \nWe compared the denoising of 3,200 contaminated datasets, each containing 100 seconds of \n64-channel EEG, which varied across 4 levels of temporal contamination, baseline SNRs, and \nartifact type. These datasets combined synthetic background EEG (simulated  in s ilico via \nforward modelling) with semi-synthetic EOG/EMG artifacts and synthetic noise (see Methods \nfor more details).  \nFor example videos  of synthetic ground-truth, noise -contaminated and GEDAI -denoised \nEEG, see Synthetic_Video_1 and Synthetic_Video_2. \n \n \nFigure 3. Denoising of Synthetic EEG \nA) By temporal contamination (%). Results are pooled across all SNRbefore and all artifact types.  For \nevery temporal contamination level, each coloured data point represents 1 of 800 datasets (50 datasets \nx 4 SNR levels x 4 artifact types). B) By SNR before (dB). Results are pooled across all tem poral \ncontamination levels and all artifact types. For every SNR before, each coloured data point represents 1 \nof 800 datasets (50 datasets x 4 artifact types x 4 temporal contamination levels). C) By artifact type. \nResults are pooled across all SNRbefore and temporal contamination levels. For each artifact type, each \ncoloured data point represents 1 of 800 datasets (50 datasets x 4 SNR levels x 4 temporal \ncontamination levels). D) Computational time (seconds).  Lower values indicate faster calculations. \nAn asterisk indicates the winning algorithm (p < 0.05 Bonferroni corrected). If no asterisk is present, it \nsignifies a statistical tie. \n \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n20 \nFigure 3  above shows benchmarking across different aggregation levels (temporal \ncontamination levels, SNRbefore , artifact type and computational time). Higher SNRafter and lower \nRRMSE (for RRMSE see Supplementary  Fig. S7 ) values indicate better denoising, \nrespectively.  Differences in SNRafter and RRMSE are reported below (effect size r = Z/√N). \n \n By temporal contamination (Fig 3A): \nFor 25% , GEDAI (median = 14.54 dB, IQR = 12.67 − 16.35) outperformed all algorithms \nexcept for MARA (median = 16.34 dB, IQR = 9.91−21.63), with non-significant differences in \nSNRafter (r = -0.13, p = 0.24) and RRMSE (r =  0.0082, p = 0.4 6); following a Friedman test: \nχ²(3) = 257, p = 2.6 x 10-55. \n For 50% , GEDAI (median = 13.00 dB, IQR = 10.71 −14.99) outperformed all algorithms, \nincluding the next-best MARA (median = 12.00 dB, IQR = 4.82−17.95) in SNRafter (r = 0.090, \np = 0.0036) and RRMS E (r = -0.22, p = 0.0015); following a Friedman test: χ²(3) = 519, p = \n2.9 x 10-112. \n For 75% , GEDAI (median = 12.43 dB, IQR = 10.12 −14.63) outperformed all algorithms, \nincluding the next-best MARA (median = 10.12 dB, IQR = 2.93−14.65) in SNRafter (r = 0.34, p \n= 6.4e-15) and RRMSE (r = -0.46, p = 1.9 x 10-18); following a Friedman test: χ²(3) = 899, p = \n1.50 x 10-194. \n For 100% , GEDAI (median = 12.38 dB, IQR = 9.96 −14.55) outperformed all algorithms, \nincluding the next-best MARA (median = 6.75 dB, IQR = -0.28−11.47), the closest competitor, \nin SNR after (r = 0.59, p = 4.7 x 10 -51) and RRMSE (r = -0.70, p = 1.8 x 10 -60); following a \nFriedman test: χ²(3) = 1028, p = 1.6 x 10-222. \n \nBy SNRbefore (Fig 3B) \nFor -9 dB, GEDAI (median = 12.68 dB, IQR = 10.28 −15.16) outperformed all algorithms, \nincluding the next-best MARA (median = 9.39 dB, IQR = 0.23−13.59), the closest competitor, \nin SNRafter (r = 0.46, p = 1.40 x 10 -30) and RRMSE (r =  -0.56, p = 3.7 x 10 -33); following a \nFriedman test: χ²(3) = 657, p = 4.3 x 10-142. \n For -6 dB, GEDAI (median = 13.00 dB, IQR = 10.59 −15.24) outperformed all algorithms, \nincluding the next -best MARA (median = 10.19 dB, IQR = 2.63 − 15.81), the closest \ncompetitor, in SNRafter (r = 0.33, p = 2.0 x 10-16) and RRMSE (r = -0.46, p = 1.3 x 10-19); following \na Friedman test: χ²(3) = 607, p = 2.4 x 10-131. \n For -3 dB, GEDAI (median = 13.23 dB, IQR = 10.72 −15.43) outperformed all algorithms, \nincluding the next-best MARA (median = 11.36 dB, IQR = 4.01−17.78), the closest competitor, \nin SNRafter (r = 0.14, p = 0.0025) and RRMSE (r = -0.30, p = 3.7 x 10-05 ); following a Friedman \ntest: χ²(3) = 565, p = 3.2 x 10-122. \n For 0 dB , GEDAI (median = 13.37 dB, IQR = 11.06 −15.52) outperformed all algorithms \nexcept for MARA (median = 13.04 dB, IQR = 6.97−19.05), with non-significant differences in \nSNRafter (r = -0.03, p = 1) and RRMSE (r = -0.14, p = 0.83); following a Friedman test: χ²(3) = \n493, p = 1.8 x 10-106. \n \nBy artifact type (Fig 3C) \nFor EOG , GEDAI (median = 13.12 dB, IQR = 11.41 −14.71) outperformed all algorithms \nexcept for MARA (median = 11.70 dB, IQR = 6.35−18.62), with non-significant differences in \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n21 \nSNRafter (r = 0.049, p = 0.79) and RRMSE (r = -0.16, p = 0.79); following a Friedman test: χ²(3) \n= 1127, p = 4.1 x 10-244. \nFor EMG , GEDAI (median = 13.33 dB, IQR = 10.87 −14.65) outperformed all algorithms, \nincluding the next-best MARA (median = 10.62 dB, IQR = 3.22−17.75) in SNRafter (r = 0.25, p \n= 1.9 x 10-10) and RRMSE  (r = -0.36, p = 1.6 x 10-10); following a Friedman test: χ²(3) = 614, \np = 1.0 x 10-132. \nFor NOISE, GEDAI (median = 17.54 dB, IQR = 14.35 - 18.88) outperformed all algorithms, \nincluding the next-best ASR (median = 14.75 dB, IQR = 9.28−19.51) in SNRafter (r = 0.073, p \n= 3.8 x 10-06) and RRMSE  (r = -0.33, p = 5.2 x 10-11); following a Friedman test: χ²(3) = 315, \np = 5.9 x 10-68. \nFor NOISE + EOG + EMG, GEDAI (median = 10.82 dB, IQR = 9.16−12.02) outperformed all \nalgorithms, including the next-best MARA (median = 8.10 dB, IQR = 1.43 −12.50) in SNRafter \n(r = 0.44, p = 3.4 x 10 -21) and RRMSE  (r = -0.52, p = 4.7 x 10 -24); following a Friedman test: \nχ²(3) = 942, p = 7.7 x 10-204. \nComputational time (Fig 3D) \nFor EOG, GEDAI (median = 4.62 s, IQR = 4.38−4.97) outperformed all algorithms, including \nthe next-best ASR (median = 5.30 seconds, IQR = 4.33 - 6.16) in Time (r = -0.44, p = 5.1 x \n10-05); following a Friedman test: χ²(3) = 2180, p < m.p. \nFor EMG, GEDAI (median = 4.61 s, IQR = 4.34−4.96) outperformed all algorithms except for \nASR (median = 3.77 seconds, IQR = 3.10−5.18), with a significant difference in Time (r = 0.40, \np = 2.1 x 10-9); following a Friedman test: χ²(3) = 2199, p < m.p. \nFor NOISE, GEDAI (median = 4.48 s, IQR = 4.26 −4.82) outperformed all algorithms except \nfor ASR (median = 4.00 s, IQR = 3.36−4.97), with a significant difference in Time (r = 0.24, p \n= 1.6 x 10-9); following a Friedman test: χ²(3) =2200, p < m.p. \nFor NOISE + EOG + EMG , GEDAI (median = 4.52 s, IQR = 4.27 −4.89) outperformed all \nalgorithms except for ASR (median = 5.04, IQR = 3.83−5.99), with a non-significant difference \nin Time (r = -0.29, p = 0.13); following a Friedman test: χ²(3) = 2165, p < m.p. \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n22 \n \nFigure 4. Global statistics for synthetic EEG benchmark \nGlobal means and 95% conf idence intervals for synthetic EEG denoising across all temporal \ncontamination levels, baseline SNRs and artifact types for: Correlation (top left panel, higher is better), \nSNRafter in dB (top right panel, higher is better), RRMSE (bottom left panel, lower is better), \nComputational time in seconds (bottom right panel, lower is better). Confidence intervals are plotted \nbut are very narrow due to the large sample size. \n \nGlobally: following a Friedman test  χ²(3) = 2252, p < m.p. across all datasets and conditions, \nGEDAI outperformed all algorithms in post -hoc pairwise comparisons, with absolute means \nand 95% confidence intervals illustrated in Fig. 4. GEDAI significantly differed from MARA, \nthe closest competitor in SNRafter (r = 0.23 ; GEDAI mean =  13.1 dB ; MARA mean = 10.8 dB), \nRRMSE (r = -0.37 ; GEDAI mean = 0.24 ; MARA mean = 0.51) and Correlation Coefficient (r \n= 0.37 ; GEDAI mean = 0.96 ; MARA mean = 0.87). GEDAI did not significantly differ from \nASR, the closest competitor in Time (r = -0.03 ; GEDAI mean = 4.6 s; ASR mean = 4.7 s). \nDenoising Benchmark of Empirical EEG  \nWe contrasted the denoising of 6,400 contaminated datasets, each containing 60 seconds of \n27-channel EEG, which varied across 4 levels of temporal contamination, baseline SNR \n(SNRbefore), and artifact type. These datasets combined genuine background EEG (serving as \nthe ground truth) with actual EOG/EMG/NOISE artifacts (see Methods for more details). \nFor example videos  of empirical ground-truth, noise -contaminated a nd GEDAI -denoised \nEEG, see Empirical_Video_1 and Empirical_Video_2. \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n23 \n \nFigure 5. Denoising of Empirical EEG \nA)  By temporal contamination (%). Results are pooled across all SNRbefore and all artifact types. For \nevery temporal contamination level, each coloured data point represents 1 of 1,600 datasets (100 \ndatasets x 4 SNR levels x 4 artifact types). B)  By SNRbefore (dB). Results are pooled across all temporal \ncontamination levels and all artifact types. For every SNR before, each coloured data point represents 1 \nof 1,600 datasets (100 datasets x 4 artifact types x 4 temporal contamination levels). C) By artifact \ntype. Results are pooled across all SNRbefore and temporal contamination levels. For each artifact type, \neach coloured data point represents 1 of 1,600 datasets (100 datasets x 4 SNR levels x 4 temporal \ncontamination levels). D) Computational time (seconds).  Lower values indicate faster calculations. \nAn asterisk indicates the winning algorithm (p < 0.05 Bonferroni corrected). If no asterisk is present, it \nsignifies a statistical tie . \nFigure 5  above shows benchmarking across different aggregation levels (temporal \ncontamination levels, SNRbefore , artifact type and computational time). Higher SNRafter and lower \nRRMSE (for RRMSE see Supplementary Fig. S7 ) values indicate better denoising, \nrespectively. Differences in SNRafter and RRMSE are reported below (effect size r = Z/√N). \nBy temporal contamination (Fig 5A)  \nFor  25%, GEDAI (median = 9.84 dB, IQR = 8.77-11.06) outperformed all algorithms, including \nthe next-best ASR (median = 9.79 dB, IQR = 7.97-11.97) in SNRafter (r = -0.07, p = 0.032) and \nRRMSE (r =  0.0046, p = 0.02); following a Friedman test: χ²(3) = 1855, p < m.p. \nFor 50%, GEDAI (median = 8.66 dB, IQR = 7.43-10.03) outperformed all algorithms, including \nthe next-best MARA (median = 6.95 dB, IQR = 4.34-9.33) in SNRafter (r = 0.55, p = 4.8 x 10-90) \nand RRMSE (r = -0.60, p = 2.2 x 10-91); following a Friedman test: χ²(3) = 1134, p = 1.6 x 10-\n245. \nFor 75%, GEDAI (median = 7.98 dB, IQR = 6.67-9.45) outperformed all algorithms, including \nthe next-best MARA (median = 6.06 dB, IQR = 2.67-8.92) in SNRafter (r = 0.53, p = 1.0 x 10-65) \nand RRMSE (r = -0.60, p = 1.1 x 10-73);  following a Friedman test: χ²(3) = 1356, p = 1.3 x 10-\n293. \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n24 \nFor 100%, GEDAI (median = 7.98 dB, IQR = 6.60-9.27 ) outperformed all algorithms, including \nthe next-best MARA (median = 5.01 dB, IQR = 0.76 -8.34) in SNRafter (r = 0.62, p = 6.2 x 10 -\n105) and RRMSE (r = -0.67, p = 2.3 x 10-112); following a Friedman test: χ²(3) = 1615, p < m.p. \nby SNRbefore (Fig 5B) \nFor -9 dB, GEDAI (median = 8.24 dB, IQR = 6.69−9.60) outperformed all algorithms, including \nthe next-best MARA (median = 5.73 dB, IQR = 0.97−8.74) in SNRafter (r = 0.61, p = 1.92 x 10-\n101) and RRMSE (r =  -0.66, p = 2.1 x 10-104); following a Friedman test: χ²(3) =1334, p = 8.3 x \n10-289. \nFor -6 dB, GEDAI (median = 8.46 dB, IQR = 7.07−9.83) outperformed all algorithms, including \nthe next-best MARA (median = 6.49 dB, IQR = 2.77−9.03) in SNRafter (r = 0.58, p = 3.0 x 10 -\n91) and RRMSE (r = -0.64, p = 1.1 x 10-97) ; following a Friedman test: χ²(3) =1194, p  = 1.5 x \n10-258. \nFor -3 dB , GEDAI (median = 8.79 dB, IQR = 7.28 −10.24) outperformed all algorithms, \nincluding the next-best MARA (median = 6.92 dB, IQR = 4.06−9.27) in SNRafter (r = 0.58, p = \n9.8 x 10-92) and RRMSE (r = -0.63, p = 1.5 x 10-98); following a Friedman test: χ²(3) = 1078, p  \n= 2.7 x 10-233. \nFor 0 dB , GEDAI (median = 9.15 dB, IQR = 7.85 −10.62) outperformed all algorithms, \nincluding the next-best ASR (median = 7.49 dB, IQR = 4.35 −10.34) in SNRafter (r = 0.47, p = \n9.2 x 10-85) and RRMSE (r = -0.55, p = 5.9e-88); following a Friedman test: χ²(3) =1056, p  = \n1.6  x 10-228. \nBy artifact type (Fig 5C) \nFor EOG, GEDAI (median = 8.37 dB, IQR = 7.44−9.32) outperformed all algorithms, including \nthe next-best IClabel (median = 8.15 dB, IQR = 6.50 −9.45) in SNR after (r = 0.26, p = 0.003) \nand RRMSE (r = -0.30, p = 0.003); following a Friedman test: χ²(3) = 1918, p < m.p. \nFor EMG, GEDAI (median = 9.97 dB, IQR = 9.04−10.99) outperformed all algorithms, except \nfor MARA (median = 10.39 dB, IQR = 8.04−11.76), with non-significant differences in SNRafter \n(r = -0.06, p = 0.70) and RRMSE  (r = -0.007, p = 0.70); following a Friedman test: χ²(3) = \n1574, p < m.p. \nFor NOISE , GEDAI (median = 9.08 dB, IQR = 7.47 −11.16) outperfor med all algorithms, \nincluding the next-best ASR (median = 8.08 dB, IQR = 3.76 −11.56) in SNRafter(r = 0.24, p = \n6.5e-17) and RRMSE  (r = -0.40, p = 7.9e-26); following a Friedman test: χ²(3) =2301, p < m.p. \nFor NOISE + EOG + EMG , GEDAI (median = 7.1 dB, IQR  = 5.93−8.29) outperformed all \nalgorithms, including the next-best MARA (median = 4.21 dB, IQR = 1.83−6.36) in SNRafter (r \n= 0.82, p = 8.4  x 10 -186) and RRMSE  (r = -0.83, p = 1.9 x 10 -192); following a Friedman test: \nχ²(3) = 2478, p < m.p. \nComputational time (Fig 5D) \nFor EOG, GEDAI (median = 1.06 s, IQR = 1.03−1.11) outperformed all algorithms, including \nthe next-best ASR (median = 2.13 s, IQR = 1.80 −2.58) in Time ( r = -0.87, p = 6.0 x 10 -141); \nfollowing a Friedman test: χ²(3) = 4493, p < m.p. \nFor EMG, GEDAI (median = 1.07 s, IQR = 1.03−1.11) outperformed all algorithms, including \nthe next-best ASR (median = 1.99 s, IQR = 1.67−2.47), the closest competitor, in Time (r = -\n0.87, p = 4.3 x 10-134); following a Friedman test: χ²(3) = 4562, p < m.p. \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n25 \nFor NOISE, GEDAI (median = 1.07 s, IQR = 1.03−1.12) outperformed all algorithms, including \nthe next-best ASR (median = 1.96 s, IQR = 1.58 −2.50) in Time (r = -0.85, p = 6.1 x 10 -161); \nfollowing a Friedman test: χ²(3) = 4098, p < m.p. \nFor NOISE + EOG + EMG , GEDAI (medi an = 1.08 s, IQR = 1.04 - 1.13) outperformed all \nalgorithms, including the next -best ASR (median = 2.22 s, IQR = 1.83 −2.77) in Time (r = -\n0.87, p = 1.53 x 10-223 ); following a Friedman test: χ²(3) = 4018, p < m.p. \n \nFigure 6. Global statistics for empirical EEG benchmark \nGlobal means and 95% confidence intervals for empirical EEG denoising across all temporal \ncontamination levels, baseline SNRs and artifact types for: Correlation (top left panel, higher is better), \nSNRafter in dB (top right panel, higher is better), RRMSE (bottom left panel, lower is better), \nComputational time in seconds (bottom right panel, lower is better).  Confidence intervals are plotted \nbut are very narrow due to the large sample size. \nGlobally: following a Friedman test  χ²(3) = 4704, p < m.p. across all datasets and conditions, \nGEDAI outperformed all algorithms in post-hoc pairwise comparisons (p < m.p.), with absolute \nmeans and 95% confidence intervals illustrated in Fig. 6. GEDAI significantly  differed from \nMARA, the closest competitor in SNR after (r = 0.59 ; GEDAI mean= 8.67 dB ; MARA mean = \n6.14 dB) and RRMSE (r = -0.64 ; GEDAI mean = 0.38 ; MARA mean = 0.57). GEDAI \nsignificantly differed from ASR, the closest competitor, in Correlation Coeff icient (r = 0.63 ; \nGEDAI mean= 0.92 ; ASR mean= 0.83) and Time (r = -0.86 ; GEDAI mean = 1.1 s; ASR mean \n= 2.2 s). \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n26 \nNeurobehavioral Prediction \n \nFigure 7. Single-trial ERP Classification and Brain Fingerprinting \nA) ROC curves for binary classification of ER Ps from a visual oddball task, after denoising \nwith each algorithm. Each curve is based on >10,000 trials (no bad trials were rejected). B) \nMean success rate (%) of individual identification from 100 subjects, after denoising with each \nalgorithm. Each poin t represents 1 of 50 resamples from each subject’s resting -state EEG \ndata.  An asterisk indicates the winning algorithm (p < 0.05 Bonferroni corrected). \nSingle-trial ERP Classification \nThe goal here was to predict whether a subject observed either a freque nt (‘white \nsquare’) or infrequent (‘Einstein’s face’) visual stimulus, based on a single -trial ERP. \nFigure 7A  shows the receiver -operating characteristics (ROCs) indicating binary \nclassification performance following denoising by each algorithm without any bad trial \nrejection. DeLong tests on the area -under-the-curve (AUC) confirmed that GEDAI \n(mean AUC = 0.80) statistically outperformed (p = 4.2 x 10-36) the next-best algorithm \nASR (mean AUC = 0.72), followed by MARA (mean AUC = 0.58) and IClabel (mean \nAUC = 0.58). Basic average referencing plus 1 -40 Hz band -pass filtering, i.e RAW \n(mean AUC = 0.52) performed close to chance level (AUC of 0.50, dotted-line). \nBrain Fingerprinting  \nIn the context of brain fingerprinting, and as shown in Figure 7B, following a significant \nKuskall-Wallis test χ²(3) = 187, p = 3.4 x 10 -40,  post-hoc comparisons indicated that \nGEDAI (mean success rate = 91%) demonstrated higher accuracy (r = 0.86, p = 9.4 x \n10-5) than the next-best algorithm ASR (mean success rate = 76%), followed by MARA \n(mean success rate = 63%), RAW (mean success rate = 55%) and IClabel (mean \nsuccess rate = 54%). \n \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n27 \nFeature Comparison between Denoising Algorithms  \nKey features of state-of-the-art denoising algorithms \n ASR ICLabel MARA GEDAI  \nRank preservation ✓ - - ✓ \nShort data length (<1 second) ✓ - - ✓ \nAutomatic thresholding / feature \nselection \n- ✓ ✓ ✓ \nNon-orthogonal artifacts - ✓ ✓ ✓ \nGaussian artifacts ✓ - - ✓ \nSpatially non-stationary artifacts ✓ - - ✓ \n100% temporally contaminated data - ✓ ✓ ✓ \nReal-time capable ✓ - - ✓ \n \nTable 1. Algorithmic feature comparison of the tested denoising algorithms \nNote: ✓ means that the performance of the function should not be affected in the listed \ncondition. However, edge cases in which a technique could perform to a satisfying level in a \ncondition for which it is not marked with an ✓ are possible. \nDiscussion \nThis paper introduces GEDAI, a novel algorithm designed to advance EEG artifact correction, \naddressing a key challenge in advancing EEG technology, as highlighted in a survey by \nMushtaq and colleagues (Mushtaq et al., 2024) . Their study, involving 500 experts across \nmore than 50 countries, identified improved artifact correction as the top priority in the field. \nDue to its theoretical underpinnings, GEDAI is fully automated, requiring no user expertise or \ninput, and may be used to recover brain signals from low signal -to-noise recordings (e.g. -9 \ndB) with up to 100% temporal contamination. Validated through rigorous simulations on \nthousands of synthetic and real -world datasets, GEDAI marks a significant a dvance in EEG \nartifact removal, delivering a unique blend of processing speed and denoising precision. As \nelaborated below, GEDAI consistently performed better than, or on par with, competing \nalgorithms in simulations containing a single type of artifact (EMG or EOG only). Interestingly, \nits most significant performance improvements were evident when handling multiple spatially \nnon-stationary and temporally overlapping artifacts, such as NOISE + EMG + EOG. This \nmakes it especially promising for use in real-world, noise-agnostic environments.  \n \n \n \n \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n28 \nBenchmark Simulations: synthetic and empirical EEG  \nAcross both synthetic and empirical EEG benchmarks, GEDAI proved to be a highly effective \nand robust method for artifact removal, demonstrating superior performance under diverse \nand challenging conditions. \nGEDAI's key strength was its superior performance on heavily contaminated data. \nWhile comparable to other algorithms like ASR and MARA in datasets with sparse artifacts \n(25% contamination), its superior ity grew significantly as contamination levels increased to \n50%, 75%, and 100%, making it optimal for processing pervasively corrupted recordings. \nFurthermore, GEDAI excelled at handling complex, mixed artifacts. While its \nperformance on isolated ocular or muscular artifacts was similar to that of ICA-based methods, \nit showed a large effect size of |r| > 0.5 when disentangling mixtures of EOG, EMG, and \nnoise—a crucial capability for real-world applications. \nFinally, the algorithm consistently outperformed competitors in low signal-to-noise ratio \n(SNR) environments (from -9 dB to -3 dB). This ability to reliably recover weak neural signals \nfrom substantial background noise confirms GEDAI's suitability for applications where data \nquality is a significant challenge. \nNeurobehavioral Prediction: ERP classification and brain fingerprinting \nThe neurobehavioral results demonstrate that the GEDAI algorithm significantly enhances the \nquality of EEG data for machine learning applications, outperforming all other tested methods \nin two distinct tasks. For single-trial ERP classification, GEDAI achieved a high mean AUC of \n0.80, substantially better than the next-best algorithm, ASR (0.72), and far exceeding the near-\nchance performance of raw data. This performance advantage was even more pronounced in \nthe brain fingerprinting paradigm, where GEDAI reached 91% accuracy in subject \nidentification, surpassing ASR (76%) and other methods that were only marginally better than \nbaseline. \nGEDAI Compared to Existing Denoising Methods \nThe GEDAI algorithm offers a combination of advantages that, to the best of our knowledge, \nis not currently exhibited by competing EEG denoising methods within a single package. As \nshown in Table 1, GEDAI appears to globally encompass the major strengths of ASR, ICLabel, \nand MARA, while their capabilities are limited to specific subsets of features.  GEDAI's \nfavorable outcomes may be firstly attributed to the adaptability of GEVD decomposition, \nconsistent with previous reports (Gouy-Pailler et al., 2009; Somers et al., 2018). Unlike PCA, \nGEVD is not limited by source orthogonality, nor is it restricted by Gaussianity, as is the case \nwith ICA (Cohen, 2022b).  \nWith parallel processing, GEDAI was between 3 to 15 times faster than ICA -based \ndenoising, and achieved speeds comparable to PCA-based ASR, a real-time capable method \n(Kothe & Jung, 2015). Hence, GEDAI is computationally light enough for BCI applications.  \nPerhaps GEDAI's biggest advantage is that it is fully automated and requires no \noperator expertise or input for component selection and/or hyperparameter tuning. While these \ncapacities are partially shared with the other algorithms, GEDAI can achieve automation in \nthe absence of any calibration or training data. Harnessing a \"noise -free\" EEG leadfield as a \ntheoretical reference model allows GEDAI to avoid shortcomings rela ted to the identification \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n29 \nof clean data, as is the case for PCA, or specific artifact types, as is the case for ICA. This \noverall combination of features makes GEDAI a great candidate for noise -agnostic settings, \nsuch as dry-electrode or mobile EEG recordings, where clean samples might not be available \nor the noise could be out-of-distribution. \nLeadfield Filtering: Theoretically Informed M/EEG Artifact Removal \nGEDAI’s key strength is the use of a theoretical M/EEG leadfield matrix for guiding the artifact \nremoval, a process we refer to as ‘leadfield filtering’. Existing denoising methods that use \ntheoretical reference models include Signal Space Projection (SSP) (Uusitalo & Ilmoniemi, \n1997), Signal Space Separation (SSS) (Taulu et al., 2004), PureEEG (Hartmann et al., 2014), \nSPHARA (Graichen et al., 2015)  and the SOUND algorithm (Mutanen et al., 2018). Of these, \nPureEEG and SOUND explicitly leverage leadfield modeling based on head anatomy. \nHowever, unlike SOUND, GEDAI’s algorithm does not require ill-posed inverse modeling used \nfor source activity estimation (Mutanen et al., 2018) . The PureEEG algorithm, in contrast to \nGEDAI’s, separately compu tes the artifact covariance in the frequency domain using a \nBayesian estimator, with the assumption that brain and artifact components are statistically \nuncorrelated (Hartmann et al., 2014). \nCutting through the Noise with SENSAI  \nIn GEDAI, a reference covariance matrix (refCOV) models the expected spatial correlations \nof brain activity based on a physical head model. Using joint diagonalization (de Cheveigné & \nParra, 2014), GEDAI decomposes recorded data into components. Those components with a \nspatial covariance inconsistent with the brain model are identified as artifacts (large \neigenvalues), while consistent components are treated as neural signals (small eigenvalues). \nDetermining the optimal cutoff between the artifact and brain components is neither \nobvious nor trivial. GEDAI solves this with its second key innovation: the Signal & Noise \nSubspace Alignment Index (SENSAI). SENSAI aut omatically finds the best separation \nthreshold by calculating an alignment score for various cutoffs and selecting the one that \nmaximizes the similarity between the denoised data and the theoretical brain subspace \ndefined by refCOV. The resulting SENSAI sc ore is a relative, not absolute, measure of \ndenoised data quality (0 -100%), dependent on the specific head model used.  Although \nbeyond the scope of this paper, it also has potential future use as a quantitative index for \ncomparing EEG recording quality or the performance of different denoising methods. \nTheoretical vs. Empirical Denoising Frameworks \nThis subject concerns the apparent dichotomy between the theoretical (leadfield-driven) \nreference matrix used by GEDAI, and the empirical (data-driven) reference matrix used by \nexisting GEVD implementations for M/EEG denoising (Haslacher et al., 2021; F. Wang et al., \n2025). The core trade -off between using a leadfield -based theoretical versus a n empirical \nrefCOV for M/EEG denoising centers on model accuracy versus data dependency. A data -\ndriven, empirical refCOV captures the actual spatial structure of noise or signal as it manifests \nin that specific recording and subject, but still requires that suitable segments of EEG can be \nreliably identified and/or denoised a priori. In contrast, using a theoretical refCOV bypasses \nthe often difficult challenge of identifying pure noise or clean signal EEG segments, but relies \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n30 \non a good fit between the theoretical model and the actual data. In this case, deviations from \nthe forward model that are specific to the subject (e.g. anatomical variations) or the recording \n(e.g. electrode misplacement) will likely lead to greater inaccuracies in the denoising process. \nBad channels and how to deal with them \nTraditional methods handle bad channels by rejecting them entirely (Bigdely-Shamlo et al., \n2015; Kumaravel et al., 2022) , which removes both artifactual  and potentially useful neural \nsignals. GEDAI offers a novel alternative by avoiding this binary rejection. It instead treats \nactivity from compromised channels as artifactual components, which are identified via GEVD \nand removed if their spatial character istics deviate from the theoretical brain signal model. \nAlthough our benchmark datasets included bad channels, this study did not explicitly evaluate \nGEDAI's effectiveness in correcting them at individual -channel level. Crucially, the other \nalgorithms bene fited from a bad -channel rejection pre -processing step only when bad \nchannels were present. This potential advantage was not extended to the fully noise-agnostic \nGEDAI pipeline. Future work should assess whether a hybrid approach, combining dedicated \nbad c hannel rejection with GEDAI, outperforms either method alone in specific noise \nscenarios. \nPotential Future Extensions of GEDAI \nThe simultaneous acquisition of EEG -fMRI or non -invasive brain stimulation (NIBS) \ntechniques like transcranial direct current sti mulation (tDCS), transcranial alternating current \nstimulation (tACS), and transcranial magnetic stimulation (TMS) present significant denoising \nchallenges due to large, complex artifacts often contaminating the entire recording. Given that \nGEDAI uses leadfield filtering to separate brain activity from extra -cerebral sources of noise, \nit may also prove effective for removing fMRI gradient and/or brain stimulation artifacts \noriginating outside the head. Although beyond the scope of this paper, our preliminary  tests \non concurrent fMRI-EEG and EEG-NIBS recordings indicated successful removal of very high-\namplitude electromagnetic noise components. Therefore, GEDAI’s automated, robust \ndenoising capabilities based on biophysical principles could make it a promising candidate for \nimproving data quality in challenging multimodal and stimulation experiments. Finally, as \nGEDAI utilizes a shared M/EEG forward model, its potential applicability could also extend to \nmagnetoencephalography (MEG) recordings, especially opti cally-pumped magnetometers \n(OPM-MEG) where sensors maintain a fixed position relative to the head (Brookes et al., \n2022). \nLimitations \nFirst, GEDAI’s performance is highly dependent on an accurate match between the theoretical \nleadfield model and the actual EEG electrode positions. Any spatial mismatch will \nproportionally degrade its efficacy, making precise electrode placement a prerequisite for \noptimal results. \nSecond, because GEDAI defines artifacts as signals originating outside the brain, it is \nnot designed to remove artifacts that originate from within the brain volume, such as those \nfrom deep brain stimulation, for example. \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n31 \nThird, the current implementation assumes uniform and uncorrelat ed brain sources. \nIncorporating a more neurophysiologically realistic source covariance model (e.g. one that \naccounts for functional networks and differing regional power) could improve performance, \nthough this requires further research. \nFinally, like PCA/ICA, the number of removable artifact components is limited by the \nnumber of channels, making GEDAI unsuitable for single-channel EEG.  \nConclusion \nTaken together, these findings position GEDAI as a state -of-the-art tool for EEG denoising, \ndistinguished by its robust performance in heavily contaminated, low-SNR conditions and its \nparticular strength in resolving complex mixtures of artifacts. \nData and code availability \nThe open-source GEDAI plugin for EEGLAB is available at: \nhttps://github.com/neurotuning/GEDAI-master \nAll simulated datasets and analysis codes used in this paper will be released as a standalone \nresource upon publication, known as “ BEAR: Benchmarking EEG Artifact Removal with \nSynthetic and Empirical Datasets”. The open EEGLAB code will allow users to consult and \nmodify the specific default parameters used in this study. \nAcknowledgements \nWe would like to thank Michael X Cohen for his helpful discussions as well as tutorials on \nEEG and GEVD, which have inspired this work. This study was supported by the Swiss \nNational Science Foundation (SNSF), grant number 215712. \nCompeting Interests Statement \nT.R. is an inventor on a patent application related to the GEDAI algorithm described in this \nmanuscript. The other authors declare they have no competing interests. \n \n \n \n \n \n \n \n \n \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n32 \nSupplementary Results \n \n \n \n \nFig S1: Clean synthetic data : example simulated ‘background’ EEG (i.e.noise-free)  \nsimulated with https://github.com/lrkrol/SEREEGA \n \n \n \n \nFig S2. Artifactual synthetic data : example of simulated data containing only EOG (left panel) or \nEMG artifacts (right panel) - using source locations from Fig S5 \nwith https://github.com/lrkrol/SEREEGA \n +     https://github.com/ncclabsustech/EEGdenoiseNet \n \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n33 \n \n \nFig S3. Synthetic artifact source locations for simulated EOG (left panel) or EMG artifacts (right \npanel) from https://github.com/lrkrol/SEREEGA \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n34 \n \nFig S4: Clean empirical data : example resting-state EEG recordings (i.e.noise-free)  \nfrom https://github.com/stefan-ehrlich/dataset-automaticArtifactRemoval \n \n \n \nFig S5. Artifactual empirical EEG data : example of simulated data containing only EOG (left panel) \nor EMG artifacts (right panel) \n \nfrom https://github.com/stefan-ehrlich/dataset-automaticArtifactRemoval \n +     https://github.com/ncclabsustech/EEGdenoiseNet \n \n \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n35 \n \nFig S6. Noisy empirical EEG data : example of simulated data containing step artifacts (left panel) or \nline noise \nfrom https://github.com/stefan-ehrlich/dataset-automaticArtifactRemoval \n +     https://github.com/ncclabsustech/EEGdenoiseNet \n+      https://github.com/vpKumaravel/NEAR/tree/main/SimulateArtifacts \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n36 \n \nFigure S7. Relative Root Mean Square Error (RRMSE) . Lower values indicate better denoising . \nAsterisks indicate the winning algorithm in each case. No asterisk signifies a statistical tie.  \nA) Synthetic denoising by temporal contamination.   For every temporal contamination level, each \ncoloured data point represents 1 of 800 datasets (50 datasets x 4 SNR levels x 4 artifact types).   \nB) Empirical denoising by temporal contamination. Results are pooled across all baseline SNRbefore \nand all artifact types.  For every temporal contamination level, each coloured data point represents 1 of \n1,600 datasets (100 datasets x 4 SNR levels x 4 artifact types).  \n \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n37 \nC) Synthetic denoising by SNRbefore. Results are pooled across all temporal contamination levels and \nall artifact types. For every SNRbefore, each coloured data point represents 1 of 800 datasets (50 datasets \nx 4 artifact types x 4 temporal contamination levels).  \nD) Empirical denoising by SNRbefore. Results are pooled across all temporal contamination levels and \nall artifact types. For every SNR before, each coloured data point rep resents 1 of 1,600 datasets (100 \ndatasets x 4 artifact types x 4 temporal contamination levels).  \nE) Synthetic denoising by artifact type.  Results are pooled across all SNR before and temporal \ncontamination levels. For every artifact type, each coloured data point represents 1 of 800 datasets (50 \ndatasets x 4 SNR levels x 4 temporal contamination levels).   \nF) Empirical denoising by artifact type. Results are pooled across all SNR before and temporal \ncontamination levels.  For every artifact type, each colour ed data point represents 1 of 1,600 datasets \n(100 datasets x 4 SNR levels x 4 temporal contamination levels).  \n \n \n \nBibliography \nAmico, E., & Goñi, J. (2018). The quest for identifiability in human functional connectomes. \nScientific Reports, 8(1), 8254. https://doi.org/10.1038/s41598-018-25089-1 \nAmin, U., Nascimento, F. A., Karakis, I., Schomer, D., & Benbadis, S. R. (2023). Normal \nvariants and artifacts : Importance in EEG interpretation. Epileptic Disorders, 25(5), \n591‑ 648. https://doi.org/10.1002/epd2.20040 \nBailey, N. W., Biabani, M., Hill, A. T., Miljevic, A., Rogasch, N. C., McQueen, B., Murphy, O. \nW., & Fitzgerald, P. B. (2023). Introducing RELAX : An automated pre-processing \npipeline for cleaning EEG data - Part 1: Algorithm and application to oscillations. \nClinical Neurophysiology, 149, 178‑ 201. https://doi.org/10.1016/j.clinph.2023.01.017 \nBigdely-Shamlo, N., Mullen, T., Kothe, C., Su, K.-M., & Robbins, K. A. (2015). The PREP \npipeline : Standardized preprocessing for large-scale EEG analysis. Frontiers in \nNeuroinformatics, 9, 16. https://doi.org/10.3389/fninf.2015.00016 \nBlankertz, B., Tomioka, R., Lemm, S., Kawanabe, M., & Muller, K. (2008). Optimizing Spatial \nfilters for Robust EEG Single-Trial Analysis. IEEE Signal Processing Magazine, \n25(1), 41‑ 56. https://doi.org/10.1109/MSP.2008.4408441 \nBoudet, S., Peyrodie, L., Forzy, G., Pinti, A., Toumi, H., & Gallois, P. (2012). Improvements \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n38 \nof Adaptive Filtering by Optimal Projection to filter different artifact types on long \nduration EEG recordings. Computer Methods and Programs in Biomedicine, 108(1), \n234‑ 249. https://doi.org/10.1016/j.cmpb.2012.04.005 \nBrookes, M. J., Leggett, J., Rea, M., Hill, R. M., Holmes, N., Boto, E., & Bowtell, R. (2022). \nMagnetoencephalography with optically pumped magnetometers (OPM-MEG) : The \nnext generation of functional neuroimaging. Trends in Neurosciences, 45(8), \n621‑ 634. https://doi.org/10.1016/j.tins.2022.05.008 \nChang, C.-Y., Hsu, S.-H., Pion-Tonachini, L., & Jung, T.-P. (2020). Evaluation of Artifact \nSubspace Reconstruction for Automatic Artifact Components Removal in Multi-\nChannel EEG Recordings. IEEE Transactions on Biomedical Engineering, 67(4), \n1114‑ 1121. https://doi.org/10.1109/TBME.2019.2930186 \nClark, I., Biscay, R., Echeverría, M., & Virués, T. (1995). Multiresolution decomposition of \nnon-stationary EEG signals : A preliminary study. Computers in Biology and \nMedicine, 25(4), 373‑ 382. https://doi.org/10.1016/0010-4825(95)00014-u \nCohen, M. X. (2022a). A tutorial on generalized eigendecomposition for denoising, contrast \nenhancement, and dimension reduction in multichannel electrophysiology. \nNeuroImage, 247, 118809. https://doi.org/10.1016/j.neuroimage.2021.118809 \nCohen, M. X. (2022b). A tutorial on generalized eigendecomposition for denoising, contrast \nenhancement, and dimension reduction in multichannel electrophysiology. \nNeuroImage, 247, 118809. https://doi.org/10.1016/j.neuroimage.2021.118809 \nde Cheveigné, A., & Parra, L. C. (2014). Joint decorrelation, a versatile tool for multichannel \ndata analysis. NeuroImage, 98, 487‑ 505. \nhttps://doi.org/10.1016/j.neuroimage.2014.05.068 \nDeLong, E. R., DeLong, D. M., & Clarke-Pearson, D. L. (1988). Comparing the areas under \ntwo or more correlated receiver operating characteristic curves : A nonparametric \napproach. Biometrics, 44(3), 837‑ 845. \nDelorme, A., & Makeig, S. (2004). EEGLAB : An open source toolbox for analysis of single-\ntrial EEG dynamics including independent component analysis. Journal of \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n39 \nNeuroscience Methods, 134(1), 9‑ 21. \nhttps://doi.org/10.1016/j.jneumeth.2003.10.009 \nDelorme, A., Palmer, J., Onton, J., Oostenveld, R., & Makeig, S. (2012). Independent EEG \nsources are dipolar. PLoS ONE, 7(2), e30135. \nhttps://doi.org/10.1371/journal.pone.0030135 \nde Munck, J. C., van Dijk, B. W., & Spekreijse, H. (1988). Mathematical dipoles are \nadequate to describe realistic generators of human brain activity. IEEE Transactions \non Biomedical Engineering, 35(11), 960‑ 966. https://doi.org/10.1109/10.8677 \nde Munck, J. C., Vijn, P. C., & da Silva, F. L. (1992). A random dipole model for spontaneous \nbrain activity. IEEE transactions on biomedical engineering, 39(8), 791‑ 804. \nEhrlich, S. (2024). Stefan-ehrlich/dataset-automaticArtifactRemoval [Logiciel]. \nhttps://github.com/stefan-ehrlich/dataset-automaticArtifactRemoval (Édition originale \n2019) \nFratangelo, R., Lolli, F., Scarpino, M., & Grippo, A. (2025). Point-of-Care \nElectroencephalography in Acute Neurological Care : A Narrative Review. Neurology \nInternational, 17(4), Article 4. https://doi.org/10.3390/neurolint17040048 \nFrølich, L., Andersen, T. S., & Mørup, M. (2015). Classification of independent components \nof EEG into multiple artifact classes. Psychophysiology, 52(1), 32‑ 45. \nhttps://doi.org/10.1111/psyp.12290 \nGabard-Durnam, L. J., Mendez Leal, A. S., Wilkinson, C. L., & Levin, A. R. (2018). The \nHarvard Automated Processing Pipeline for Electroencephalography (HAPPE) : \nStandardized Processing Software for Developmental and High-Artifact Data. \nFrontiers in Neuroscience, 12. https://doi.org/10.3389/fnins.2018.00097 \nGouy-Pailler, C., Sameni, R., Congedo, M., & Jutten, C. (2009). Iterative Subspace \nDecomposition for Ocular Artifact Removal from EEG Recordings. In T. Adali, C. \nJutten, J. M. T. Romano, & A. K. Barros (Éds.), Independent Component Analysis \nand Signal Separation (Vol. 5441, p. 419‑ 426). Springer Berlin Heidelberg. \nhttps://doi.org/10.1007/978-3-642-00599-2_53 \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n40 \nGraichen, U., Eichardt, R., Fiedler, P., Strohmeier, D., Zanow, F., & Haueisen, J. (2015). \nSPHARA - A Generalized Spatial Fourier Analysis for Multi-Sensor Systems with \nNon-Uniformly Arranged Sensors : Application to EEG. PLOS ONE, 10(4), e0121741. \nhttps://doi.org/10.1371/journal.pone.0121741 \nGramfort, A., Papadopoulo, T., Olivi, E., & Clerc, M. (2010). OpenMEEG: opensource \nsoftware for quasistatic bioelectromagnetics. BioMedical Engineering OnLine, 9(1), \n45. https://doi.org/10.1186/1475-925X-9-45 \nHajhassani, D., Mattout, J., & Congedo, M. (2024). An Automatic Riemannian Artifact \nRejection Method for P300-based BCIs. 2024 32nd European Signal Processing \nConference (EUSIPCO), 1616‑ 1620. \nhttps://doi.org/10.23919/EUSIPCO63174.2024.10715349 \nHarmening, N., Klug, M., Gramann, K., & Miklody, D. (2022). HArtMuT—modeling eye and \nmuscle contributors in neuroelectric imaging. Journal of Neural Engineering, 19(6), \n066041. https://doi.org/10.1088/1741-2552/aca8ce \nHartmann, M. M., Schindler, K., Gebbink, T. A., Gritsch, G., & Kluge, T. (2014). PureEEG: \nAutomatic EEG artifact removal for epilepsy monitoring. Neurophysiologie clinique = \nClinical neurophysiology, 44(5), 479‑ 490. \nhttps://doi.org/10.1016/j.neucli.2014.09.001 \nHaslacher, D., Nasr, K., Robinson, S. E., Braun, C., & Soekadar, S. R. (2021). Stimulation \nartifact source separation (SASS) for assessing electric brain oscillations during \ntranscranial alternating current stimulation (tACS). NeuroImage, 228, 117571. \nhttps://doi.org/10.1016/j.neuroimage.2020.117571 \nHsu, S.-H., Mullen, T., Jung, T.-P., & Cauwenberghs, G. (2014). Online recursive \nindependent component analysis for real-time source separation of high-density \nEEG. 2014 36th Annual International Conference of the IEEE Engineering in \nMedicine and Biology Society, 3845‑ 3848. \nhttps://doi.org/10.1109/EMBC.2014.6944462 \nHuang, Y., Parra, L. C., & Haufe, S. (2016). The New York Head-A precise standardized \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n41 \nvolume conductor model for EEG source localization and tES targeting. NeuroImage, \n140, 150‑ 162. https://doi.org/10.1016/j.neuroimage.2015.12.019 \nIlle, N., Berg, P., & Scherg, M. (2002). Artifact correction of the ongoing EEG using spatial \nfilters based on artifact and brain signal topographies. Journal of Clinical \nNeurophysiology: Official Publication of the American Electroencephalographic \nSociety, 19(2), 113‑ 124. https://doi.org/10.1097/00004691-200203000-00002 \nJung, T. P., Makeig, S., Humphries, C., Lee, T. W., McKeown, M. J., Iragui, V., & Sejnowski, \nT. J. (2000). Removing electroencephalographic artifacts by blind source separation. \nPsychophysiology, 37(2), 163‑ 178. \nKnyazev, A. V., & Argentati, M. E. (2002). Principal Angles between Subspaces in an A-\nBased Scalar Product : Algorithms and Perturbation Estimates. SIAM Journal on \nScientific Computing, 23(6), 2008‑ 2040. \nhttps://doi.org/10.1137/S1064827500377332 \nKoles, Z. J. (1991). The quantitative extraction and topographic mapping of the abnormal \ncomponents in the clinical EEG. Electroencephalography and Clinical \nNeurophysiology, 79(6), 440‑ 447. https://doi.org/10.1016/0013-4694(91)90163-x \nKothe, C. A. E. (2013). The artifact subspace reconstruction method. Accessed: Jul, 17, \n2017. \nKothe, C. A. E., & Jung, T.-P. (2015). Artifact removal techniques with signal reconstruction \n(World Intellectual Property Organization Brevet No. WO2015047462A9). \nhttps://patents.google.com/patent/WO2015047462A9/en \nKrol, L. R., Pawlitzki, J., Lotte, F., Gramann, K., & Zander, T. O. (2018). SEREEGA : \nSimulating event-related EEG activity. Journal of Neuroscience Methods, 309, \n13‑ 24. https://doi.org/10.1016/j.jneumeth.2018.08.001 \nKumaravel, V. P., Farella, E., Parise, E., & Buiatti, M. (2022). NEAR : An artifact removal \npipeline for human newborn EEG data. Developmental Cognitive Neuroscience, 54, \n101068. https://doi.org/10.1016/j.dcn.2022.101068 \nLazarevic, A., & Kumar, V. (2005). Feature bagging for outlier detection. Proceedings of the \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n42 \neleventh ACM SIGKDD international conference on Knowledge discovery in data \nmining, 157‑ 166. https://doi.org/10.1145/1081870.1081891 \nLee, T.-W., Girolami, M., & Sejnowski, T. J. (1999). Independent Component Analysis Using \nan Extended Infomax Algorithm for Mixed Subgaussian and Supergaussian Sources. \nNeural Computation, 11(2), 417‑ 441. https://doi.org/10.1162/089976699300016719 \nLi, Z., Zhao, Y., Hu, X., Botta, N., Ionescu, C., & Chen, G. H. (2023). ECOD : Unsupervised \nOutlier Detection Using Empirical Cumulative Distribution Functions. IEEE \nTransactions on Knowledge and Data Engineering, 35(12), 12181‑ 12193. \nhttps://doi.org/10.1109/TKDE.2022.3159580 \nLotte, F., & Guan, C. (2011). Regularizing common spatial patterns to improve BCI designs : \nUnified theory and new algorithms. IEEE Transactions on Bio-Medical Engineering, \n58(2), 355‑ 362. https://doi.org/10.1109/TBME.2010.2082539 \nMak, J. N., & Wolpaw, J. R. (2009). Clinical Applications of Brain-Computer Interfaces : \nCurrent State and Future Prospects. IEEE reviews in biomedical engineering, 2, \n187‑ 199. https://doi.org/10.1109/RBME.2009.2035356 \nMumtaz, W., Rasheed, S., & Irfan, A. (2021). Review of challenges associated with the EEG \nartifact removal methods. Biomedical Signal Processing and Control, 68, 102741. \nhttps://doi.org/10.1016/j.bspc.2021.102741 \nMushtaq, F., Welke, D., Gallagher, A., Pavlov, Y. G., Kouara, L., Bosch-Bayard, J., van \nden Bosch, J. J. F., Arvaneh, M., Bland, A. R., Chaumon, M., Borck, C., He, X., Luck, \nS. J., Machizawa, M. G., Pernet, C., Puce, A., Segalowitz, S. J., Rogers, C., Awais, \nM., … Valdes-Sosa, P. (2024). One hundred years of EEG for brain and behaviour \nresearch. Nature Human Behaviour, 8(8), 1437‑ 1443. \nhttps://doi.org/10.1038/s41562-024-01941-5 \nMutanen, T. P., Metsomaa, J., Liljander, S., & Ilmoniemi, R. J. (2018). Automatic and robust \nnoise suppression in EEG and MEG : The SOUND algorithm. NeuroImage, 166, \n135‑ 151. https://doi.org/10.1016/j.neuroimage.2017.10.021 \nOlbrich, S., Jödicke, J., Sander, C., Himmerich, H., & Hegerl, U. (2011). ICA-based muscle \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n43 \nartefact correction of EEG data : What is muscle and what is brain?: Comment on \nMcMenamin et al. NeuroImage, 54(1), 1‑ 3. \nhttps://doi.org/10.1016/j.neuroimage.2010.04.256 \nOmejc, N., Peskar, M., Miladinović, A., Kavcic, V., Džeroski, S., & Marusic, U. (2023). On the \nInfluence of Aging on Classification Performance in the Visual EEG Oddball \nParadigm Using Statistical and Temporal Features. Life, 13(2), Article 2. \nhttps://doi.org/10.3390/life13020391 \nParra, L., & Sajda, P. (2003). Blind Source Separation via Generalized Eigenvalue \nDecomposition. Journal of Machine Learning Research, 4, 1261‑ 1269. \nPion-Tonachini, L., Kreutz-Delgado, K., & Makeig, S. (2019). ICLabel : An automated \nelectroencephalographic independent component classifier, dataset, and website. \nNeuroImage, 198, 181‑ 197. https://doi.org/10.1016/j.neuroimage.2019.05.026 \nRadüntz, T., Scouten, J., Hochmuth, O., & Meffert, B. (2017). Automated EEG artifact \nelimination by applying machine learning algorithms to ICA-based features. Journal \nof Neural Engineering, 14(4), 046004. https://doi.org/10.1088/1741-2552/aa69d1 \nRos, T., Baars, B. J., Lanius, R. A., & Vuilleumier, P. (2014). Tuning pathological brain \noscillations with neurofeedback : A systems neuroscience framework. Frontiers in \nHuman Neuroscience, 8, 1008. https://doi.org/10.3389/fnhum.2014.01008 \nSawangjai, P., Hompoonsup, S., Leelaarporn, P., Kongwudhikunakorn, S., & Wilaiprasitporn, \nT. (2020). Consumer Grade EEG Measuring Sensors as Research Tools : A Review. \nIEEE Sensors Journal, 20(8), 3996‑ 4024. IEEE Sensors Journal. \nhttps://doi.org/10.1109/JSEN.2019.2962874 \nSomers, B., Francart, T., & Bertrand, A. (2018). A generic EEG artifact removal algorithm \nbased on the multi-channel Wiener filter. Journal of Neural Engineering, 15(3), \n036007. https://doi.org/10.1088/1741-2552/aaac92 \nSorrentino, P., Rucco, R., Lardone, A., Liparoti, M., Troisi Lopez, E., Cavaliere, C., Soricelli, \nA., Jirsa, V., Sorrentino, G., & Amico, E. (2021). Clinical connectome fingerprints of \ncognitive decline. NeuroImage, 238, 118253. \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n44 \nhttps://doi.org/10.1016/j.neuroimage.2021.118253 \nTadel, F., Baillet, S., Mosher, J. C., Pantazis, D., & Leahy, R. M. (2011). Brainstorm : A \nUser-Friendly Application for MEG/EEG Analysis. Computational Intelligence and \nNeuroscience, 2011, 879716. https://doi.org/10.1155/2011/879716 \nTaulu, S., Kajola, M., & Simola, J. (2004). Suppression of interference and artifacts by the \nSignal Space Separation Method. Brain Topography, 16(4), 269‑ 275. \nhttps://doi.org/10.1023/b:brat.0000032864.93890.f9 \nUusitalo, M. A., & Ilmoniemi, R. J. (1997). Signal-space projection method for separating \nMEG or EEG into components. Medical & Biological Engineering & Computing, \n35(2), 135‑ 140. https://doi.org/10.1007/BF02534144 \nWang, F., Ma, Y., Gao, T., Tao, Y., Wang, R., Zhao, R., Cao, F., Gao, Y., & Ning, X. (2025). \nRepairbads : An automatic and adaptive method to repair bad channels and \nsegments for OPM-MEG. NeuroImage, 306, 120996. \nhttps://doi.org/10.1016/j.neuroimage.2024.120996 \nWang, Y., Berg, P., & Scherg, M. (1999). Common spatial subspace decomposition applied \nto analysis of brain responses under multiple task conditions : A simulation study. \nClinical Neurophysiology: Official Journal of the International Federation of Clinical \nNeurophysiology, 110(4), 604‑ 614. https://doi.org/10.1016/s1388-2457(98)00056-x \nWeinstein, D., Zhukov, L., & Johnson, C. (2000). Lead-field bases for \nelectroencephalography source imaging. Annals of Biomedical Engineering, 28(9), \n1059‑ 1065. https://doi.org/10.1114/1.1310220 \nWinkler, I., Brandl, S., Horn, F., Waldburger, E., Allefeld, C., & Tangermann, M. (2014). \nRobust artifactual independent component classification for BCI practitioners. Journal \nof Neural Engineering, 11(3), 035013. https://doi.org/10.1088/1741-\n2560/11/3/035013 \nZhang, H., Zhao, M., Wei, C., Mantini, D., Li, Z., & Liu, Q. (2021). EEGdenoiseNet : A \nbenchmark dataset for deep learning solutions of EEG denoising. Journal of Neural \nEngineering, 18(5), 056057. https://doi.org/10.1088/1741-2552/ac2bf8 \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint \n\n \n45 \nZhang, X. (2024). Deep Learning-Based Techniques for Electroencephalogram (EEG) \nSignal Denoising. Transactions on Computer Science and Intelligent Systems \nResearch, 5, 922‑ 927. https://doi.org/10.62051/rrve8560 \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n \n.CC-BY 4.0 International licensemade available under a \n(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is \nThe copyright holder for this preprintthis version posted October 7, 2025. ; https://doi.org/10.1101/2025.10.04.680449doi: bioRxiv preprint","source_license":"CC-BY-4.0","license_restricted":false}