{"paper_id":"5f6f32ce-a5ac-41d0-9c18-657c5ab2feed","body_text":"Surfing beta burst waveforms to improve motor imagery-based BCI\nS. Papadopoulos1,2,3,*, L. Darmet1,2,3, M.J. Szul1,3, M. Congedo4, J.J. Bonaiuto1,3,†, J. Mattout1,2,†\n1 University Lyon 1, Lyon, France\n2 Lyon Neuroscience Research Center, CRNL, INSERM, U1028, CNRS, UMR 5292, Lyon, France\n3 Institut des Sciences Cognitives Marc Jeannerod, CNRS, UMR 5229, Lyon, France\n4 GIPSA-lab, University Grenoble Alpes, CNRS, Grenoble-INP , Grenoble, France\n† These authors contributed equally\n* Corresponding author: Sotirios Papadopoulos\nE-mail: sotirios.papadopoulos@univ-lyon1.fr\nKeywords: beta bursts, brain–computer interface (BCI), decoding, electroencephalography (EEG),\nmotor imagery (MI)\nAbstract\nOur understanding of motor-related, macroscale brain processes has been significantly shaped by\nthe  description  of  the  event-related  desynchronization  (ERD)  and  synchronization  (ERS)\nphenomena in the mu and beta frequency bands prior to, during and following movement. The\ndemonstration  of  reproducible,  spatially-  and  band-limited  signal  power  changes  has,\nconsequently, attracted the interest of non invasive brain-computer interface (BCI) research for a\nlong time. BCIs often rely on motor imagery (MI) experimental paradigms that are expected to\ngenerate brain signal modulations analogous to movement-related ERD and ERS. However, a\nnumber of recent neuroscience studies has questioned the nature of these phenomena. Beta band\nactivity has been shown to occur, on a single-trial level, in short, transient and heterogeneous\nevents termed bursts rather than sustained oscillations. In a previous study, we established that an\nanalysis of hand MI binary classification tasks based on beta bursts can be superior to beta power\nin  terms  of  classification  score.  In  this  article  we  elaborate  on  this  idea,  proposing  a  signal\nprocessing algorithm that is comparable to- and compatible with state-of-the-art techniques. Our\npipeline filters brain recordings by convolving them with kernels extracted from beta bursts and\nthen applies spatial filtering before classification. This data-driven filtering allowed for a simple and\nefficient analysis of signals from multiple sensors thus being suitable for online applications. By\nadopting a time-resolved decoding approach we explored MI dynamics and showed the specificity\nof  the  new  classification  features.  In  accordance  with  previous  results,  beta  bursts  improved\nclassification  performance  compared  to  beta  band  power,  while  often  increasing  information\ntransfer rate compared to state-of-the-art approaches.\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 July 19, 2024. ; https://doi.org/10.1101/2024.07.18.604064doi: bioRxiv preprint \n\nSignificance statement\nPatterns of waveform-specific burst rate comprise an alternative, neurophysiology-informed way of\nanalyzing beta band activity during motor imagery (MI) tasks. By testing this method on multiple\nelectroencephalography datasets and comparing its corresponding classification scores against\nthose of conventional power-based features, this work demonstrates that brain-computer interface\napplications could benefit from utilizing beta burst activity. This activity gives access to a reliable\ndecoding performance often requiring short recordings. As such, this study shows that waveform-\nspecific beta burst rates encode information related to imagined (and presumably real) movements\nand serves as the first step for a real-time implementation of the proposed methodology.\nIntroduction\nTime-locked changes in induced power within specific frequency bands, originally described in a\nnumber of seminal studies in motor neuroscience (Pfurtscheller, 1981; Pfurtscheller and Berghold,\n1989; Pfurtscheller and Lopes da Silva, 1999) , have long influenced the way in which we interpret\nmacroscale recordings of brain activity such as those provided by electroencephalography (EEG).\nThese  studies  have  revealed  a  gradual  reduction  in  brain  signal  power  during  an  ongoing\nmovement or motor imagery (MI) task in the mu (~8-12 Hz) (Neuper et al., 2006; Pfurtscheller et\nal., 2006, 1997; Pfurtscheller and Lopes da Silva, 1999)  and beta (~13-30 Hz) (Pfurtscheller et al.,\n1997; Pfurtscheller and Lopes da Silva, 1999)  frequency bands relative to baseline activity. This\nphenomenon is termed event-related desynchronization (ERD). The same studies have, moreover,\ndescribed a relative-to-baseline increase in power in the beta band shortly following the end of the\nmovement or MI (Alayrangues et al., 2019; Neuper et al., 2006; Pfurtscheller et al., 1996) , known\nas event-related synchronization (ERS). These phenomena are especially marked over cortical\nareas contralateral to the real or imagined movement  (Kobler et al., 2020; Little et al., 2019;\nMakeig et al., 2000; Pfurtscheller and Berghold, 1989; Pfurtscheller and Neuper, 1997; Seeber et\nal.,  2016;  Zich  et  al.,  2023)  and  their  topographies  approximately  match  the  somatotopic\norganization of the sensorimotor cortices  (Gordon et al., 2023; Natraj et al., 2022; Penfield and\nRasmussen, 1950). Taken together, these observations have given rise to the hypothesis that the\nERD is an indication of brain processes pertaining to movement preparation and execution while\nthe ERS is an indication of processes related to movement completion (Kilavik et al., 2013).\nGiven the reproducibility of the spatial and frequency specificity of the ERD and ERS, these neural\nmarkers are often exploited by non-invasive BCI applications, especially those that are based on\nMI  paradigms  (Jayaram  and  Barachant,  2018;  Tangermann  et  al.,  2012) .  Such  paradigms,\ndesigned to reproduce consistent time-locked signal modulations, normally rely on transforming the\nrecordings in the time-frequency domain (TF)  (Brodu et al., 2011; Bruns, 2004; Herman et al.,\n2008) and  then  applying  spatial  filtering,  most  commonly  using  the  common  spatial  pattern\nalgorithm (CSP)  (Blankertz et al., 2008; Koles, 1991; Müller-Gerking et al., 1999) . This chain of\nsignal transformations is expected to increase signal-to-noise ratio by extracting signal power in\nspecific time windows and frequency bands of interest, and also to maximize the spatial disparity\namong different MI classes (e.g. ”left” or “right” hand, or “feet”), thus improving classification results\nand/or allowing for decoding of multiple commands with distinct signal features (Lotte, 2014; Lotte\net al., 2018).\nAlthough the ERD and ERS are consistently observed across subjects and recording modalities,\ntheir  nature  is  not  clear.  Based  on  the  assumption  of  amplitude  modulation  of  sustained\noscillations, these patterns are the result of signal power averaging in the TF domain over multiple\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 July 19, 2024. ; https://doi.org/10.1101/2024.07.18.604064doi: bioRxiv preprint \n\ntrials. However, converging evidence suggests that, on the contrary, beta band activity occurs in\nshort events termed bursts (Coleman et al., 2024; Jones, 2016; Little et al., 2019; Lundqvist et al.,\n2024, 2016; Shin et al., 2017; Torrecillos et al., 2018; Wessel, 2020)  on the single trial level,\ntherefore questioning the functional role of ERD and ERS altogether, at least within the beta band.\nBeta burst rate has been shown to be more behaviorally relevant in motor processes (Enz et al.,\n2021; Hannah et al., 2020; Little et al., 2019; Rayson et al., 2023; Szul et al., 2023; Wessel,\n2020) than averaged beta band power. Additionally, recent studies have shown that beta bursts are\nnot a unitary phenomenon but rather constitute heterogeneous events  (Szul et al., 2023)  with\ndifferent functions, alluded to by the differential modulation of their rate and shape depending on\ntask conditions (Langford et al., 2023; Papadopoulos et al., 2024)  or movement phase (Rayson et\nal., 2023; Szul et al., 2023) . As such, beta bursts have the potential to be a more sensitive marker\nof brain processes during real or imagined movements on the single trial level.\nTo test this hypothesis, in a previous study we examined multiple open MI EEG datasets. We\nanalyzed the activity of channels C3 and C4 during binary classification tasks of hand MI (“left” vs\n“right” hand) and demonstrated that the waveform-resolved beta burst rate is superior to beta band\npower changes in terms of classification (Papadopoulos et al., 2024) . In this article we streamline\nour approach.  We  develop  an algorithm  that  is  computationally  efficient  and  can  analyze  an\narbitrary number of recorded signals thus being comparable to state-of-the-art techniques. Beta\nburst  waveforms,  whose  rate  is  expected  to  be  maximally  modulated  during  the  trial  period\ncompared to baseline, are identified in calibration data. These bursts are used as data-driven\nkernels that filter the signals from all recording channels in the time domain. The convolved signals\nare then spatially filtered with CSP and the spatial features are used as classification features. We\nre-analyze the activity during “left” and “right” hand MI of the same open EEG datasets and also a\nrecently-published  composite  EEG  dataset,  now  in  a  time-resolved  fashion.  We  show  that\nclassification  features  based  on  waveform-resolved  beta  burst  rate  offer  better  decoding\nperformance  and  improve  the  decoding  speed  versus  accuracy  trade-off  when  compared  to\nstandard band-limited power-based classification features.\nMaterials and Methods\nDatasets: We analyzed the recordings of six open EEG MI benchmark datasets available through\nthe MOABB  (Aristimunha et al., 2023; Jayaram and Barachant, 2018)  project: BNCI 2014-001\n(Tangermann et al., 2012) , BNCI 2014-004  (Leeb et al., 2007) , Cho 2017  (Cho et al., 2017) ,\nMunichMI (Grosse-Wentrup 2009) (Grosse-Wentrup et al., 2009), Weibo 2014 (Yi et al., 2014) and\nZhou 2016 (Zhou et al., 2016) , and the recordings of a composite open EEG dataset that became\nrecently available (Dreyer et al., 2023)  referred to hereafter as Dreyer 2023. All datasets comprise\na number of subjects with recordings corresponding to multiple trials of two or more randomly\nchosen, sustained kinesthetic MI commands, each performed following the appearance of a visual\ncue on a screen (Table 1). For our analysis we only considered trials corresponding to the “left\nhand” or “right hand” classes.\nPre-processing: The epoched recordings of each subject were loaded using the MOABB python\npackage (v0.4.6, class LeftRightImagery; parameters: t min and t max as in indicated in Table 1), and\nwere filtered with a low pass cutoff of 120 Hz (parameters: f min = 0, f max = 120; default MNE\n(Gramfort et al., 2013)  zero-phase FIR filter designed with the windowed approach and transition\nbandwidth of 25% of the low pass frequency). Because the sampling frequency of the Weibo 2014\nrecordings is 200 Hz, we set the low pass cutoff to 95 Hz for this dataset. Finally, we used the\nautoreject python package (Jas et al., 2017) (v0.4.0, function get_rejection_threshold, default para-\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 July 19, 2024. ; https://doi.org/10.1101/2024.07.18.604064doi: bioRxiv preprint \n\nDataset # Subjects # Channels # Trials Baseline\nperiod (s)\nTask\nperiod (s)\nPost-task\nperiod\n(s)\nBNCI\n2014-001 9 22 288\n217 – 288 (243) -1.0 – 0.0 0.0 – 4.0 4.0 – 5.5\nBNCI\n2014-004 9 3 680 – 760 (720)\n269 – 621 (411) -1.0 – 0.0 0.0 – 4.5 4.5 – 6.5\nCho\n2017 49 64 200 – 240 (200)\n  38 – 237 (159) -1.0 – 0.0 0.0 – 3.0 3.0 – 5.0\nDreyer\n2023 87 27 160 – 240 (240)\n  23 – 240 (192) -1.0 – 0.0 0.0 – 5.0 5.0 – 6.0\nMunichMI\n(Grosse-\nWentrup 2009)\n10 128 300\n109 – 278 (200) -1.0 – 0.0 0.0 – 7.0 7.0 – 9.0\nWeibo\n2014 10 64 140 – 160 (160)\n  31 – 160 (132) -1.0 – 0.0 0.0 – 4.0 3.0 – 3.5\nZhou\n2016 4 14 290 – 319 (295)\n114 – 280 (150) -1.0 – 0.0 0.0 – 5.0 5.0 – 7.0\nTable 1. Dataset attributes. The lines in the fourth column indicate the original number of trials per subject (or\nthe range in case this number was different between subjects), and the range of remaining trials across all\nsubjects following trial rejection. Numbers in parentheses indicate the median number of trials.\nmeters) in order to remove noisy trials.\nBurst detection and kernel selection:  In order to select kernels for convolving the data from all\nchannels we first detected bursts from channels C3 and C4 (or equivalently channels 43 and 44 for\nthe MunichMI (Grosse-Wentrup 2009) dataset). To do so, we applied the  pre-processing steps\ndescribed  above  within  a  dataset-specific  cluster  of  channels  above  the  sensorimotor  cortex\n(Papadopoulos et al., 2024) . We applied a time-frequency (TF) decomposition in the 1 – 43Hz\nrange  on  each  selected  channel  separately,  using  the  superlets  algorithm  (Moca  et  al.,\n2021) (parameters: omin = 1, omax = 40, c = 4) with a frequency resolution of 0.5 Hz. We noted high-\npower artifacts at approximately 25 – 30 Hz when inspecting the TF of the Cho 2017, Dreyer 2023\nand MunichMI (Grosse-Wentrup 2009) datasets. This noise interferes with the burst detection step,\ntherefore, we included an extra pre-processing step, prior to trial rejection, based on a custom\nimplementation  of  the  ZapLine  algorithm  from  the  meegkit  python  package  (de  Cheveigné,\n2020) (v0.1.3, dss_line function) to remove these artifacts. Then, we detected bursts within the\nbeta frequency range (15 – 30 Hz) from each TF matrix channel and used their temporal location\nto extract their waveforms from the raw time series within a fixed time-window of 260ms. For more\ndetails regarding the burst detection step we refer the readers to previous work from our group\n(Szul et al., 2023).\nAs the number of detected bursts per subject is large, we randomly sampled 10% of the trials of\neach participant per dataset and created a matrix that contained the waveforms of all detected\nbursts regardless of the trial class (“left” or “right” hand) for a given dataset. Due to the large\nnumber of subjects of the Dreyer 2023 dataset we restricted the random sample to 5% of each\nsubject’s trials. Then, after robust scaling (scikit-learn package (Pedregosa et al., 2011) , v1.0.2),\nwe  reduced  the  time  dimension  of  the  waveforms  using  principal  component  analysis  (PCA)\n(Shlens, 2014) (scikit-learn package, v1.0.2).\nWe used the PCA score of each waveform detected from electrodes C3 and C4 (or equivalently\nchannels 43 and 44 for the MunichMI (Grosse-Wentrup 2009) dataset), which is a metric of the\ndifference between any waveform and the average shape of all bursts contained in the matrix\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 July 19, 2024. ; https://doi.org/10.1101/2024.07.18.604064doi: bioRxiv preprint \n\nprovided as input to PCA. We defined an index of lateralized modulation of the average-per-axis\nPCA score I m :\nI m=|( uipsi\nC 3\n−ucontra\nC 4\n) − (uipsi\nC 4\n−u contra\nC 3\n)|, m∈ {2 , ... , 9}\nu=|^scoretrial period − ^scorebaseline|\nwhere ipsi (contra) refers to bursts recorded from channels C3 / C4 during a left / right (right / left)\nhand MI (Figure 1 a). \nThis index measures the inter-hemispheric difference of the average waveform shape between the\nbaseline and trial periods. Its values span the range [0 , ∞ ) and higher values indicate greater\ndiscrepancies between hemispheres and the two recording periods.\nBased on observations from our previous study (Papadopoulos et al., 2024) , we computed I m\namong components 2 to 9 in order to find  three PCA axes that maximized this metric. We did not\ntake into account the first component because it likely describes the temporal skew of the bursts\n(Papadopoulos et al., 2024; Szul et al., 2023) . Finally, we divided the score range of each of the\nthree  selected  axes  in  seven  equally  spaced  groups,  each  group  corresponding  to  a  set  of\n“similarly shaped” bursts. We kept the two groups per axis that lie further away from the origin\n(score equal to 0) and, by computing the Euclidean-average waveform of bursts within each group,\nwe identified two kernels per axis. As such, we identified six kernels per dataset corresponding to\nburst waveforms whose rates were expected to be maximally modulated during the task, compared\nto baseline.\nFeature extraction:  For each subject we applied the pre-processing, burst detection and kernel\nselection steps described above (pre-processing was applied to all available recording channels).\nThen, we convolved the EEG recordings with the corresponding kernels thus computing a proxy of\nthe waveform-resolved burst rate per kernel. The temporally convolved, epoched data was then\nspatially  filtered  using  the  CSP  algorithm  (MNE  package,  v1.5.1,  function  CSP,  parameters:\nn_components = 4, transform_into = “average_power”). Finally, we concatenated all 24 spatial\nfeatures into a single vector for each trial (Figure 1 b).\nTo compare with, we also used standard approaches to compute spatial features of band-limited\npower modulations. After pre-processing, we independently filtered the epoched data in the mu (6\n– 15 Hz), beta (15 – 30 Hz) or both the mu and beta (6 – 30 Hz) bands, using either a single filter\nor a filter bank approach. Then, the filtered data served as inputs to the CSP algorithm (using the\nalready  described  parameters)  resulting  in  four  spatial  features  per  filter.  For  the  filter  bank\napproach, we split either frequency range in non-overlapping filter banks with a frequency span of\n3 Hz per filter. As such we defined three filters for the mu band (6 – 9 Hz, 9 – 12 Hz, 12 – 15 Hz),\nfive filters for the beta band (15 – 18 Hz, 18 – 21 Hz, 21 – 24 Hz, 24 – 27 Hz, 27 – 30 Hz) and eight\nfilters for the mu-beta band (6 – 9 Hz, 9 – 12 Hz, 12 – 15 Hz, 15 – 18 Hz, 18 – 21 Hz, 21 – 24 Hz,\n24 – 27 Hz, 27 – 30 Hz). Then, we again used CSP and concatenated all spatial features of each\nfilter bank, resulting in 12, 20 and 32 spatial features respectively per trial.\nClassification: We used a repeated ( n = 10), 5-fold cross validation procedure to estimate the\ndecoding score using linear discriminant analysis (LDA)  (Tharwat et al., 2017; Vidaurre et al.,\n2011)(scikit-learn, v1.0.2) as a classifier. We adopted a time-resolved decoding paradigm, using\nboth an incremental and a sliding time window. In the first case we started with a 100 ms time\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 July 19, 2024. ; https://doi.org/10.1101/2024.07.18.604064doi: bioRxiv preprint \n\nFigure 1: Illustration of methodology for computing classification features based on the convolution of raw\nsignals with beta burst waveform kernels.  (a) After randomly sampling the recording trials of all subjects\nwithin any dataset the beta bursts waveforms are analyzed using PCA. This constructs a high-dimensional\nspace whose origin corresponds to the shape of the average waveform or equivalently a score equal to 0,\nand each axis defines a different axis of waveform variation. By only considering the beta bursts of channels\nC3 and C4 that occur at any point in time, the lateralization modulation index Im dynamically identifies the\nexpected deviation of the average waveform shape from the overall average shape for each PCA axis. The\naxes that maximize  Im are identified, the bursts are projected on these axes, split in groups of similarly\nshaped waveforms and the average waveform shapes of the two extrema are computed. (b) The raw signals\nof all recording channels of each dataset are independently convolved with each selected waveform from a\nresulting in distinctly temporally filtered copies of the signals. Each copy is then spatially filtered using the\nCSP algorithm, and finally all spatial features are concatenated in a single matrix that is provided as input to\nthe classifier.\nwindow and repeated the classification procedure by incrementing this window by 100 ms at a\ntime. The baseline period was considered separately from the trial period. In the latter case we\nused 1 s long sliding time windows which moved in 50 ms increments. Decoding scores were\nbased on the area under the curve (AUC) of the receiver operating characteristic (scikit-learn,\nv1.0.2). All numeric computations were based on the numpy python package (v1.21.6; (Harris et\nal., 2020)) and an environment running python (v3.10).\nInformation transfer rate:  Information transfer rate (Arslan and Sinha, 2024; Sadeghi and Maleki,\n2019) was defined as:\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 July 19, 2024. ; https://doi.org/10.1101/2024.07.18.604064doi: bioRxiv preprint \n\nITR= 1−H\nT\nH=−p (t) log2 p (t) − (1−p (t)) log2(1− p(t ))\nwhere the binary entropy function H depends on the average accuracy probability at any time\nwindow p(t) ,  and T corresponds  to  the  maximum  recording  time  required  by  each  time\nwindow in seconds. T was shifted such that all time values are positive, i.e. using the absolute\ntime starting from the beginning of the baseline period (Table 1) when using a sliding window and\nwhen considering the baseline period using an incremental time window . The values of this metric\nspan the range [0 ,10] when using an incremental window decoding approach and [0 ,20]\nwhen using a sliding window. L arge values indicate a better trade-off between decoding accuracy\nand decoding speed.\nStatistical analysis: On the dataset level, we performed pairwise comparisons of the across-subject\naverage decoding score corresponding to the beta burst convolution spatial features and each of\nthe spatial features of the band-limited power modulations. These comparisons were based on\nthreshold-free  cluster-based  permutation  ( n =  2 13)  tests  (MNE  package,  v1.5.1,  function\npermutation_cluster_test, parameters: threshold = dict(start = 0, step = 0.2), tail = 1) that were\nsubsequently thresholded at significance level of a = 0.05 for visualization purposes.\nTo estimate, on the population level, any statistical differences between the maximum classification\nscores obtained using different feature extraction pipelines during the trial period, we compared the\nscores of the beta burst convolution pipeline against those based on classical filtering pipelines.\nWe used a linear mixed model with across-trials average classification score as the dependent\nvariable setting the number of trials as prior weights, the type of classification feature as a fixed\neffect, and subject nested within dataset as random intercepts. We implemented similar models to\ncompare the time required to achieve the maximum classification score per feature extraction\npipeline, and also the maximum ITR and time needed to reach it. In the latter two cases we first\ntransformed the values to logarithmic scale in order to ensure normality of the residuals. Statistical\nanalyses were conducted using R (v4.1.2) and lme4 (v1.1-31; (Bates et al., 2015) ). Fixed effects\nwere assessed using type II Wald X 2 tests using car (v3.1-1; (Fox and Weisberg, 2019)). Pairwise\nTukey-corrected  follow-up  tests  were  carried  out  using  estimated  marginal  means  from  the\nemmeans package (v.1,8,7; (Lenth, 2023)).\nResults\nIn summary, we have employed seven freely available datasets of EEG recordings from subjects\nperforming left and right hand MI. Within each dataset, we detected beta bursts for each subject\nwithin electrode clusters over left and right sensorimotor cortex and then randomly sampled 10% of\nthe trials containing these bursts (the sample size was limited to 5% for the Dreyer 2023 dataset).\nWe applied PCA to the matrix of all burst waveforms and defined a modulation index Im in order\nto find burst waveform shapes whose lateralized rates were expected to be maximally modulated\nbetween the baseline and trial periods. These waveforms were then employed as kernels for\nconvolving the EEG data in time domain before applying spatial filtering with CSP. Finally all spatial\nfeatures were combined and served as input for LDA, in order to classify “left” versus “right” hand\nMI. We also performed classification by applying standard temporal filtering techniques before\napplying spatial filtering with CSP, using either a single filter or a filter bank in the mu (6 – 15 Hz),\nbeta (15 – 30 Hz) and mu-beta (6 – 30 Hz) frequency bands. We estimated the time-resolved\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 July 19, 2024. ; https://doi.org/10.1101/2024.07.18.604064doi: bioRxiv preprint \n\ndecoding score per subject of each dataset for each classification feature employing both an\nincremental and a sliding decoding window (see Materials and Methods for details).\nAcross all datasets, the average decoding accuracy obtained using the proposed methodology\nbased on beta bursts outperformed the results based on standard beta band filtering irrespective of\nthe filtering (single filter or filter bank) or the windowing (incremental or sliding) technique during\nmost of the recording time (Figure 2 a, Sup. Figure 1 a). Within each dataset, the across-subjects\naverage score obtained by beta bursts was higher than that of any beta band filtering technique,\nusually shortly after the beginning of the trial or towards its end (Figure 2 b-h, Sup. Figure 1 b-h).\nExceptions to this finding when using an incremental window were the BNCI 2014-001 dataset, for\nwhich all features produced equivalent results (Figure 2 c, Sup. Figure 1 c), and the Weibo 2014\ndataset,  for  which  mu-beta  filtering  outperformed  beta  bursts  (Figure  2  g,  Sup.  Figure  1  g).\nThreshold-free cluster-based permutation tests (see Materials and Methods) revealed a significant\ncluster of increased accuracy for beta bursts compared to either beta filtering technique during\nmost of the recording time or following the trial onset for each windowing (incremental or sliding)\ntechnique respectively (Figure 2 a, Sup. Figure 1 a). Within each dataset, we found clusters similar\nto those of the population average for the Dreyer 2023 dataset (Figure 2 e, Sup. Figure 1 e). The\nacross-subjects average score obtained by beta bursts was higher than that of any beta band\nfiltering technique shortly after the beginning of the trial for the BNCI 2014-001 and Cho 2017\ndatasets (Figure 2 b, d, Sup. Figure 1 b, d). For the rest of the datasets (BNCI 2014-004, MunichMI\n(Grosse-Wentrup 2009), Weibo 2014, Zhou 2016) no differences were observed among the beta\nbursts  and  either  beta  band  filtering  method.  Overall,  in  terms  of  classification  accuracy,  we\nobserved an improvement with beta bursts over beta power on the population level and in 5/7 (4/7)\ndatasets, with clusters of statistically significant differences arising on the population level and in\n3/7 (2/7) datasets when using an incremental (sliding) window.\nWe did not observe such clear differences when comparing the beta bursts and mu-beta filtering\ndecoding scores. On the population level, when using an incremental time window, average beta\nburst convolution results outperformed the single filter and filter bank techniques early after the\nbeginning of a trial (Figure 2 a). On the dataset level, this was true for the BNCI 2014-001 and\nDreyer datasets (Figure 2 b, e). For the rest of the datasets, differences varied depending on the\nfiltering  technique.  Notably,  in  the  mu-beta  band,  both  filtering  techniques  produced  higher\ndecoding scores than beta bursts for the Weibo 2014 dataset (Figure 2 g). A similar pattern was\nalso observed when using a sliding window (Sup. Figure 1 a-h). The permutation tests between the\nbeta  burst  features  and  the  mu-beta  filtering  techniques  revealed  only  small  clusters  on  the\npopulation  level,  as  well  for  the  Dreyer  2023  dataset.  We  found  that  beta  bursts  improve\nclassification scores over the mu-band filtering techniques on the population level and 4/7 (2/7)\ndatasets when using an incremental (sliding) window, with small clusters of statistically significant\ndifferences on the population level and one dataset only when using an incremental window.\nFinally,  comparisons  between  beta  burst  convolution  and  mu  filtering  results  showed  an\nimprovement when using a sliding window. Clusters of statistically significant differences revealed\nthat  beta bursts  are better than  single-filter mu  band  power  on  the population  level  and  2/7\ndatasets (Sup. Figure 2 a-c). No differences were found on the population level when comparing\nbeta bursts to the filter bank technique, but on the dataset level the beta bursts score was better for\nthe BNCI 2014-001 dataset (Sup. Figure 2 b) and conversely worse for the Weibo 2014 dataset\n(Sup. Figure 2 g) based on cluster permutation tests. Comparisons of results when using a sliding\nwindow approach did not reveal any differences (Sup. Figure 3).\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 July 19, 2024. ; https://doi.org/10.1101/2024.07.18.604064doi: bioRxiv preprint \n\nFigure  2: (a) Population  average  time-resolved  decoding  score  and  standard  error  for the  beta  burst\nconvolution (red), beta band (yellow) and mu-beta band (purple) filtering pipelines using an incremental\nwindow. Due to the different duration of the task per dataset we restricted the time to the minimum trial\nperiod corresponding to 3 seconds. (b – h)  Average, time-resolved decoding score and standard error per\ndataset of the same features using an incremental window. For each panel, the left subplot depicts the\ndecoding results obtained using a single filter, while the right subplot depicts the results based on a filter\nbank technique. The beta burst results are the same for the pair of each panel. The horizontal dashed line\ncorresponds to the expected chance level. Vertical dotted lines represent the onset and end of the trial period\nof each dataset. Horizontal lines on the top of each subplot show the results of pair-wise permutation cluster\ntests between the beta bursts and either filtering technique, with correction for multiple comparisons at\nsignificance level of 0.05. The color of the lines indicate which feature produces, on average, better results at\nany time point. A lack of color indicates no statistically significant differences between the compared features.\nFor all datasets we computed the information transfer rate (ITR) in order to quantify the difference\nbetween all classification features in terms of the decoding speed-accuracy trade-off (see Materials\nand Methods). On the population level, beta bursts provided a higher ITR than either beta band\nfiltering technique across the whole trial (Figure 3 a). On the dataset level, the same pattern was\nobserved for the Dreyer 2023 dataset (Figure 3 e). Moreover, beta bursts ITR was higher than any\nbeta band filtering early after trial onset for the BNCI 2014-001, Munich MI (Grosse-Wentrup 2009)\nand Zhou 2016 datasets, and later for the Cho 2017 dataset. No differences between the features\nwere observed in the case of BNCI 2014-004 dataset, whereas beta bursts resulted in the lowest\nITR for the Weibo 2014 dataset (Figure 3 b-h). Permutation cluster tests revealed a significant\ndifference between beta bursts and either filtering technique in the beta band on the population\nlevel, and for most of the time for the Dreyer 2023 dataset. A cluster after the trial onset was found\nfor the BNCI 2014-001 dataset, and no clusters were found for the rest of the datasets. When\nusing a sliding window approach, beta bursts provided a higher ITR than either beta band filtering\ntechnique  on  the  population  level  early  after  the  beginning  of  the  trial  and  towards  its  end\nhighlighted by the presence of a cluster (Sup. Figure 4 a). A significant difference was also found\nfor the Dreyer 2023 dataset after the trial onset (Sup. Figure 4 e). For datasets BNCI 2014-001,\nCho 2017, Dreyer 2023 and Zhou 2016 we observed a higher ITR for beta bursts compared to\nbeta band filtering mainly after the trial onset, whereas for the rest of the datasets either feature\nresulted in equivalent ITRs (Sup. Figure 4 b-h). No significant clusters were found for any of these\ndatasets. In summary, irrespective of the windowing method beta bursts resulted in higher ITR than\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 July 19, 2024. ; https://doi.org/10.1101/2024.07.18.604064doi: bioRxiv preprint \n\nFigure 3: (a) Population average time-resolved information transfer rate (ITR) and standard error for the beta\nburst convolution (red), beta band (yellow) and mu-beta band (purple) filtering pipelines using an incremental\nwindow. Due to the different duration of the task per dataset we restricted the time to the minimum trial\nperiod corresponding to 3 seconds.  (b – h)  Average, time-resolved information transfer rate (ITR) and\nstandard error per dataset of the same features using an incremental window. For each panel, the left\nsubplot depicts the ITR results obtained using a single filter, while the right subplot depicts the results based\non a filter bank technique. The beta burst results are the same for the pair of each panel. Vertical dotted lines\nrepresent the onset and end of the trial period of each dataset. Horizontal lines on the top of each subplot\nshow the results of pair-wise permutation cluster tests between the beta bursts and either filtering technique,\nwith correction for multiple comparisons at significance level of 0.05. The color of the lines indicate which\nfeature produces, on average, better results at any time point. A lack of color indicates no statistically\nsignificant differences between the compared features. Insets provide a zoomed-in view of the values within\nthe corresponding panels.\nbeta power on the population level and for 5/7 datasets with clusters arising on the population level\nand 2/7 datasets.\nRegarding the comparison between beta bursts and the mu-beta filtering techniques, ITR was\nhigher for the former on the population level as well as datasets BNCI 2014-001, Cho 2017, Dreyer\n2023, Munich MI (Grosse-Wentrup 2009) and Zhou 2016 shortly after the trial onset, especially\nwhen adopting the filter bank method. No differences were observed for the BNCI 2014-004\ndataset. The mu-band filter bank technique produced higher ITR than beta bursts period in the\ncase of the Weibo 2014 dataset (Figure 3 a-h). The results were similar when using a sliding\nwindow (Sup. Figure 4 a-h). On the population level, cluster-based permutations tests revealed a\nsignificant difference between the beta bursts features and the filter bank in the mu-beta band\nusing  either  an  incremental  or  a  sliding  window  approach,  but  no  significant  clusters  when\ncomparing the beta bursts to the single filter in the mu-beta band (Figure 3 a, Sup. Figure 4 a).\nSimilar clusters of significant differences between the features were found only for the Dreyer 2023\ndataset with an incremental window (Figure 3 e, Sup. Figure 4 e). Overall, beta bursts yielded\nhigher  ITR  than  mu-beta  power  on  the  population  level  and  4/7  datasets  regardless  of  the\nwindowing technique. Permutation cluster tests revealed improvements attributable to beta bursts\ncompared to the filter bank technique on the population level, and for one dataset when using an\nincremental time window.\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 July 19, 2024. ; https://doi.org/10.1101/2024.07.18.604064doi: bioRxiv preprint \n\nWe used linear mixed models to quantify the differences in maximum decoding score, latency to\nachieve the maximum score, maximum ITR and latency to the maximum ITR per feature (see\nMaterials and Methods) using both the incremental and sliding windows (Figure 4, Sup. Figure 5).\nOn the population level, the maximum classification scores for the beta bursts technique, filtering in\nthe beta band using a single filter, filtering in the mu band using a single filter, and filter bank in the\nbeta or mu and beta bands using an incremental time window were 0.832 ± 0.066, 0.755 ± 0.065,\n0.827 ± 0.066, 0.734 ± 0.066 and 0.819 ± 0.066 (Figure 4 a) . The time required to reach each of\nthese decoding scores (latency) was 2.26 ± 0.91, 3.04 ± 0.90, 2.96 ± 0.90, 2.50 ± 0.90 and 2.44 ±\n0.91 seconds (Figure 4 c), respectively.  The maxima of the ITR before applying the logarithmic\ntransformation (see Materials and Methods) per feature were 0.6443  ± 0.207, 0.0554 ± 0.207,\n0.3531  ±  0.0206,  0.0667  ±  0.206,  0.5241  ±  0.207  (Figure  4  b). The  corresponding  average\nlatencies before the logarithmic transformation were 0.827 ± 0.203, 1.316 ± 0.204, 1.182 ± 0.203,\n1.064 ± 0.201, 0.945 ± 0.203 seconds (Figure 4 d). Regarding the analysis based on a sliding\nwindow, the maximum classification scores were 0.816 ± 0.050, 0.753 ± 0.050, 0.810 ± 0.050,\n0.753 ± 0.050 and 0.818 ± 0.050 respectively (Sup. Figure 5 a). The latencies were 2.03 ± 0.23,\n2.45 ± 0.22, 2.19 ± 0.23, 2.49 ± 0.24 and 2.22 ± 0.23 seconds (Sup. Figure 5 c). Before applying\nthe logarithmic transformation, the ITR maxima were 0.694 ± 0.180, 0.170 ± 0.178, 0.413 ± 0.177,\n0.206 ± 0.182, 0.533 ± 0.179 (Sup. Figure 5 b), and the corresponding latencies 0.575 ± 0.162,\n0.840 ± 0.161, 0.796 ± 0.160, 0.743 ± 0.161, 0.701 ± 0.162 seconds (Sup. Figure 5 d).\nAcross all datasets the maximum classification accuracy of the beta bursts technique when using\nan incremental window was significantly higher than that of the beta band single filter and filter\nbank pipelines ( X2(4) = 326.81, t(24587436) = 10.697, p < 0.001 and t(24548972) = 13.705 , p <\n0.001 respectively), but did not differ significantly from either technique exploiting both the mu and\nbeta bands (t(24904394) = 0.637, p = 0.9691 and t(24049399) = 1.777, p = 0.3873). The latency to\nachieve  the  maximum  score  was  significantly  lower  for  the  beta  burst  convolution  pipeline\ncompared to both the single filtering techniques in the beta and mu-beta bands ( X2(4) = 62.508,\nt(24698263) =  -6.361,  p <  0.001  and  t(25032435) =  -5.769,  p <  0.001),  but  did  not  differ\nsignificantly compared to the corresponding filter bank techniques ( t(24649522) = -1.926,  p =\n0.3034 and t(24151454) = -1.488, p = 0.5703). The logarithmic transform of the maximum ITR for\nthe beta bursts technique was significantly higher than the single filtering and filter bank techniques\nin  both  the  beta  and  mu-beta  bands  ( X2(4) =  309.58,  t(24722157) =  14.0967,  p <  0.001,\nt(24671284) = 12.904, p < 0.001 and t(25060057) = 5.127, p < 0.001, t(24173608) = 2.909, p =\n0.0298). The logarithmic transform of the latency to achieve maximum ITR was significantly lower\nfor the beta bursts technique compared to either single filtering pipeline in the beta and mu-beta\nbands (X2(4) = 40.75, t(24777354) = -5.863, p < 0.001 and t(25130735) = -3.885, p = 0.001), but\ndid not significantly differ compared to the filter bank method in either band ( t(24721433) = -2.071,\np = 0.2329 and t(24227058) = -1.545, p = 0.5329).\nWhen using a sliding window, a cross all datasets the maximum classification accuracy of the beta\nbursts technique was significantly higher than that of the beta band single filter and filter bank\npipelines (X2(4) = 237.95, t(24582761) = 10.072, p < 0.001 and t(24544732) = 10.055 , p < 0.001\nrespectively), but did not differ significantly from either technique exploiting both the mu and beta\nbands ( t(24898964) = 0.956,  p = 0.8748 and  t(24045211) = -0.418,  p = 0.9936). Similarly, the\nlatency  to achieve  the maximum  score was  significantly  lower  for  the beta  burst convolution\npipeline compared to either filtering technique in the beta band ( X2(4) = 25.552, t(24736767) = -\n3.890, p < 0.001 and t(24684274) = -4.281, p < 0.001), but did not differ significantly compared to\nthe mu-beta band ( t(25086359) = -1.467, p = 0.5840 and t(24189739) = -1.762, p = 0.3961). The\nlogarithmic transform of the maximum ITR for the beta bursts technique was significantly higher\nthan both filtering techniques in the beta band and the single filtering in the mu-beta band (X2(4) = \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 July 19, 2024. ; https://doi.org/10.1101/2024.07.18.604064doi: bioRxiv preprint \n\nFigure 4:  Population-level statistical analysis based on linear mixed models when using an incremental\nwindow per feature. (a) Average maximum decoding score. (b) Average maximum ITR. (c) Average latency\nto reach the maximum decoding score. (d) Average latency to reach the maximum ITR. Error bars show 95%\nconfidence  intervals.  Hatches indicate the use of a filter bank technique.  Asterisks indicate  statistically\nsignificant differences among pairwise comparisons of the beta bursts and the rest of the features (* : p <\n0.05. ** : p < 0.01, *** : p < 0.001). A lack of asterisks implies no statistically significant differences. Note that\nthe log transform of the the maximum ITR and latency to maximum ITR were used for the statistical analysis\n(see Materials and Methods), but panels b and d depict results before applying the transformation for ease of\ncomparisons with panels a and c respectively.\n162.26,  t(24732475) = 10.236,  p < 0.001,  t(24680636) = 9.172,  p < 0.001 and  t(25072582) =\n4.499, p < 0.001) but did not significantly differ from the mu-beta filter bank (t(24183763) = 1.665, p\n= 0.4559). The logarithmic transform of the latency to achieve maximum ITR was significantly\nlower for the beta bursts technique compared to either single filtering pipeline in the beta and mu-\nbeta bands ( X2(4) = 13.151,  t(24808114) = -3.069,  p = 0.0183 and  t(25164334) = -3.029,  p =\n0.0207),  but  did  not  significantly  differ  compared  to  the  filter  bank  method  in  either  band\n(t(24749694) = -1.796, p = 0.3756 and t(24255336) = -1.296, p = 0.6937).\nDiscussion\nStandard techniques for analyzing meso- and macro-scale neural signals recorded during the\nexecution  or  imagination  of  movements  typically  rely  on  signal  power  metrics  assuming  that\nrelevant changes in brain signals are reflected in amplitude modulation (Alayrangues et al., 2019;\nKilavik et al., 2013; Pfurtscheller, 1981; Pfurtscheller et al., 1997, 1996; Pfurtscheller and Lopes da\nSilva, 1999). However, there has recently been a considerable paradigm shift towards considering\ntransient signal features on the single trial level (Chen et al., 2021; Coleman et al., 2024; Jones,\n2016; Little et al., 2019; Lundqvist et al., 2024, 2016; Rayson et al., 2023; Shin et al., 2017; Szul et\nal., 2023; Torrecillos et al., 2018; Vigué-Guix and Soto-Faraco, 2022; Wessel, 2020) . Therefore,\nconsidering  that  computational  models  describing  the  neuronal  generators  of  specific  burst\nwaveform shapes (Bonaiuto et al., 2021; Sherman et al., 2016; Szul et al., 2023)  offer an improved\ntheoretical interpretability of the observed signal modulations, applications leveraging such signal\ncharacteristics, like beta bursts, could potentially benefit from incorporating recent neuroscience\nfindings (Papadopoulos et al., 2022). \nIn this work we analyzed the activity of seven open EEG MI datasets and focused on developing a\nstreamlined process for incorporating beta bursts into a BCI pipeline. We defined a modulation\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 July 19, 2024. ; https://doi.org/10.1101/2024.07.18.604064doi: bioRxiv preprint \n\nindex I m which  allowed  us  to  identify  beta  burst  waveforms  whose  rate  is  expected  to  be\nmodulated due to task demands. This was based on the assumption that maximal modulations of\nthe average waveform shape along specific PCA components are the result of a net imbalance of\nthe rates of bursts with different shapes driven by task demands. PCA constitutes a mathematically\ntractable and interpretable way of analyzing the diversity of beta burst waveform shapes, but other\nsupervised  dimensionality  reduction  algorithms  could  better  disentangle  the  waveforms  by\nimposing different constraints, e.g. taking into account trial labels.\nWe used these waveforms as kernels to convolve the raw signals in the time domain. We chose to\nkeep the number of kernels and the method for extracting them fixed based on insights from a\nprevious  study  (Papadopoulos  et  al.,  2024) ,  but  future  work  could  employ  a  formal  hyper-\nparameter search so as to maximize classification accuracy. Similarly, we did not check for any\nredundancy in the convolved signals due to selection of similarly shaped kernels, a point which\ncould be addressed by another dimensionality reduction algorithm. The implementation of the\nconvolution is virtually as computationally efficient as any filtering technique. However, we note that\nthe proposed methodology assumes the existence of data that can be analyzed offline in order to\nfirst find the relevant beta burst waveforms to use as kernels. Moreover, this data needs to be\nclean of artifacts as the beta burst detection algorithm may fail to detect transient activity in the\npresence of high-power oscillations, instead detecting only the latter. For this reason, we believe\nthat  the  superlets  algorithm  is  the  only  time-frequency  decomposition  technique  providing\nadequate time and frequency resolution in order to determine if the data is clean and to extract\nburst waveforms.\nThis data-driven, neurophysiology-informed filtering of the signals resulted in a proxy of waveform-\nresolved burst rate per kernel. We also performed a standard filtering in the mu (6 – 15 Hz) band,\nbeta (15 – 30 Hz) band, or a wider frequency range encompassing both the mu and beta (6 – 30\nHz) bands. Finally, we used CSP to extract spatial features for classification. We showed that\nclassification scores can be improved compared to a standard power-based analysis of the beta\nband activity, requiring briefer recordings to do so and that, without explicitly considering the mu (6\n– 15 Hz) band activity in our beta bursts pipeline, the corresponding classification scores are\nequivalent to scores of state-of-the-art approaches again needing shorter recordings. Further, the\nfilter bank-based analysis allowed us to verify that the decoding improvements where not simply\nthe result of an increase in number of spatial features used for classification. Instead, beta burst\nwaveforms are more informative of the underlying MI task on the population level than beta band\npower, and equally informative to standard power-based techniques that take into account mu\nactivity modulations. A possible explanation for this is that the beta burst kernels also capture\nslower modulations of the underlying activity. Future work could explore the incorporation of novel\nmu band features or a combination of mu power and beta burst features.\nBy employing both an incremental and a sliding window strategy for classification we can speculate\non the observed differences of decoding scores across classification features and datasets. First, it\nappears that MI does not begin right after the go cue in all datasets but can be delayed possibly\ndue to differences in the task design or the instructions given to the participants. Second, relatively\nsustained decoding performances when using a sliding window were translated into slow increases\nof those performances when using an incremental window. In contrast, a drop in performance\nwhen using a sliding window was reflected in a plateau when using an incremental window. Taken\ntogether these observations imply that finding an optimal decoding time window, especially for\nonline paradigms, is not trivial and depends not only on the selected classification features or\nalgorithm but also on experimental design variables.\nIn order to assess the trade-off between decoding accuracy and speed, we used the classification\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 July 19, 2024. ; https://doi.org/10.1101/2024.07.18.604064doi: bioRxiv preprint \n\nscores obtained by implementing each pipeline and computed the corresponding ITR. We showed\na statistically significant increase of the maximum ITR achieved using the beta burst kernel filtering\ncompared to any other method, and a statistically significant decrease of the time needed to\nachieve this value compared to single filters. These results suggest that beta bursts could be\nparticularly relevant for BCI applications that aim to minimize the recording time required before\nissuing a command like in the case of real-time decoding of a switch control, although the non-\nstructured  nature  of  such  applications  still  poses  a  considerable  challenge  (Carrara  and\nPapadopoulo, 2024).\nThis study focused on incorporating recent neurophysiology insights, specifically task dependent\nwaveform-specific modulation of beta burst rates, in a pipeline for decoding EEG signals during\nimagined  movements.  We  proposed  a  simple  and  computationally  efficient  algorithm  that\nleverages beta burst waveforms and transforms brain recordings in a way that is compatible with\nother widely adopted algorithms. We demonstrated that classification results based on beta bursts\nare superior to results based beta-band power alone, and are on-par with power-based results that\ntake into account the mu band. By computing the information transfer rate we showed that, often,\nfeatures based on beta bursts significantly improve the decoding speed-accuracy trade-off. We\nalso verified this finding using a sliding window decoding technique, a fact which further suggests\nthe feasibility for online decoding with this approach. Taking everything into account, we believe\nthat these findings can serve as an important step in the direction of improving online BCI decoding\nparadigms.\nData and Code Availability\nAll data are freely available via the  MOABB project  and the open access repository Zenodo at\nhttps://zenodo.org/records/8089820. All scripts necessary for reproducing the results of this article\nare  available  at  the  following  public  repository:  https://gitlab.com/sotpapad/bebopbci/-/tree/\npreprint_version_202407.\nAuthor Contributions\nS P , J B and J M conceptualized the manuscript. S P drafted the manuscript and performed the\nanalysis. All authors contributed to manuscript revision, read, and approved the submitted version.\nFunding\nS P , L D, M C, J B and J M are supported by the French National Research Agency (ANR) project\nHiFi (2020–2024, ANR-20-CE17-0023). M S and J B are supported by the European Research\nCouncil (ERC) under the European Union’s Horizon 2020 Research and Innovation Programme\n(ERC consolidator Grant 864550 to J B). This work was performed within the framework of the\nLABEX CORTEX (ANR-11LABX-0042) of Université de Lyon, within the program ‘Investissements\nd’Avenir’ (decision n◦ 2019-ANR-LABX-02) operated by the French National Research Agency\n(ANR).\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 July 19, 2024. ; https://doi.org/10.1101/2024.07.18.604064doi: bioRxiv preprint \n\nDeclaration of Competing Interests\nAll authors declare no competing interests.\nReferences\nAlayrangues J, Torrecillos F, Jahani A, Malfait N. 2019. Error-related modulations of the \nsensorimotor post-movement and foreperiod beta-band activities arise from distinct neural \nsubstrates and do not reflect efferent signal processing. Neuroimage 184:10–24. \ndoi:10.1016/j.neuroimage.2018.09.013\nAristimunha B, Carrara I, Guetschel P , Sedlar S, Rodrigues P, Sosulski J, Narayanan D, Bjareholt \nE, Quentin B, Schirrmeister RT, Kalunga E, Darmet L, Gregoire C, Abdul Hussain A, Gatti R, \nGoncharenko V, Thielen J, Moreau T, Roy Y , Jayaram V, Barachant A, Chevallier S. 2023. \nMother of all BCI Benchmarks. doi:10.5281/zenodo.10034224\nArslan SS, Sinha P. 2024. Information transfer rate in BCIs: Towards tightly integrated symbiosis. \nBiomed Signal Process Control 87. doi:10.1016/j.bspc.2023.105466\nBates D, Mächler M, Bolker BM, Walker SC. 2015. Fitting linear mixed-effects models using lme4. \nJ Stat Softw 67. doi:10.18637/jss.v067.i01\nBlankertz B, Kawanabe M, Tomioka R, Hohlefeld FU, Nikulin V, Müller KR. 2008. Invariant common\nspatial patterns: Alleviating nonstationarities in Brain-Computer Interfacing. Adv Neural Inf \nProcess Syst 20 - Proc 2007 Conf 1–8.\nBonaiuto JJ, Little S, Neymotin SA, Jones SR, Barnes GR, Bestmann S. 2021. Laminar dynamics \nof high amplitude beta bursts in human motor cortex. Neuroimage 242:118479. \ndoi:10.1016/j.neuroimage.2021.118479\nBrodu N, Lotte F, Lécuyer A. 2011. Comparative study of band-power extraction techniques for \nMotor Imagery classification. IEEE SSCI 2011 - Symp Ser Comput Intell - CCMB 2011 2011 \nIEEE Symp Comput Intell Cogn Algorithms, Mind, Brain 95–100. \ndoi:10.1109/CCMB.2011.5952105\nBruns A. 2004. Fourier-, Hilbert- and wavelet-based signal analysis: Are they really different \napproaches? J Neurosci Methods 137:321–332. doi:10.1016/j.jneumeth.2004.03.002\nCarrara I, Papadopoulo T. 2024. Pseudo-online framework for BCI evaluation: a MOABB \nperspective using various MI  and SSVEP datasets. J Neural Eng 21. doi:10.1088/1741-\n2552/ad171a\nChen YY, Lambert KJM, Madan CR, Singhal A. 2021. Mu oscillations and motor imagery \nperformance: A reflection of intra-individual success, not inter-individual ability. Hum Mov Sci \n78:1–12. doi:10.1016/j.humov.2021.102819\nCho H, Ahn M, Ahn S, Kwon M, Jun SC. 2017. EEG datasets for motor imagery brain-computer \ninterface. Gigascience 6:1–8. doi:10.1093/gigascience/gix034\nColeman SC, Seedat ZA, Pakenham DO, Quinn AJ, Brookes MJ, Woolrich MW, Mullinger KJ. \n2024. Post-task responses following working memory and movement are driven by transient \nspectral bursts with similar characteristics 1–14. doi:10.1002/hbm.26700\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 July 19, 2024. ; https://doi.org/10.1101/2024.07.18.604064doi: bioRxiv preprint \n\nde Cheveigné A. 2020. ZapLine: A simple and effective method to remove power line artifacts. \nNeuroimage 207. doi:10.1016/j.neuroimage.2019.116356\nDreyer P, Roc A, Pillette L, Rimbert S, Lotte F. 2023. A large EEG database with users’ profile \ninformation for motor imagery brain-computer interface research. Sci Data 10:1–14. \ndoi:10.1038/s41597-023-02445-z\nEnz N, Ruddy KL, Rueda-Delgado LM, Whelan R. 2021. Volume of β-bursts, but not their rate, \npredicts successful response inhibition. J Neurosci 41:5069–5079. \ndoi:10.1523/JNEUROSCI.2231-20.2021\nFox J, Weisberg S. 2019. An R Companion to Applied Regression, Third. ed. Sage.\nGordon EM, Chauvin RJ, Van AN, Rajesh A, Nielsen A, Newbold DJ, Lynch CJ, Seider NA, \nKrimmel SR, Scheidter KM, Monk J, Miller RL, Metoki A, Montez DF, Zheng A, Elbau I, \nMadison T, Nishino T, Myers MJ, Kaplan S, Badke D’Andrea C, Demeter D V., Feigelis M, \nRamirez JSB, Xu T, Barch DM, Smyser CD, Rogers CE, Zimmermann J, Botteron KN, Pruett \nJR, Willie JT, Brunner P , Shimony JS, Kay BP, Marek S, Norris SA, Gratton C, Sylvester CM, \nPower JD, Liston C, Greene DJ, Roland JL, Petersen SE, Raichle ME, Laumann TO, Fair DA,\nDosenbach NUF. 2023. A somato-cognitive action network alternates with effector regions in \nmotor cortex. Nature 617:351–359. doi:10.1038/s41586-023-05964-2\nGramfort A, Luessi M, Larson E, Engemann DA, Strohmeier D, Brodbeck C, Goj R, Jas M, Brooks \nT, Parkkonen L, Hämäläinen M. 2013. MEG and EEG data analysis with MNE-Python. Front \nNeurosci. doi:10.3389/fnins.2013.00267\nGrosse-Wentrup M, Liefhold C, Gramann K, Buss M. 2009. Beamforming in Noninvasive Brain–\nComputer Interfaces. IEEE Trans Biomed Eng 56:1209–1219. \ndoi:10.1109/TBME.2008.2009768\nHannah R, Muralidharan V, Sundby KK, Aron AR. 2020. Temporally-precise disruption of prefrontal \ncortex informed by the timing of beta bursts impairs human action-stopping. Neuroimage 222. \ndoi:10.1016/j.neuroimage.2020.117222\nHarris CR, Millman KJ, van der Walt SJ, Gommers R, Virtanen P , Cournapeau D, Wieser E, Taylor \nJ, Berg S, Smith NJ, Kern R, Picus M, Hoyer S, van Kerkwijk MH, Brett M, Haldane A, del Río \nJF, Wiebe M, Peterson P, Gérard-Marchant P , Sheppard K, Reddy T, Weckesser W, Abbasi H,\nGohlke C, Oliphant TE. 2020. Array programming with NumPy. Nature 585:357–362. \ndoi:10.1038/s41586-020-2649-2\nHerman P , Prasad G, McGinnity TM, Coyle D. 2008. Comparative analysis of spectral approaches \nto feature extraction for EEG-based motor imagery classification. IEEE Trans Neural Syst \nRehabil Eng 16:317–326. doi:10.1109/TNSRE.2008.926694\nJas M, Engemann DA, Bekhti Y, Raimondo F, Gramfort A, Gramfort A, Automated A, Engemann \nDA. 2017. Autoreject : Automated artifact rejection for MEG and EEG data. Neuroimage \n159:417–129. doi:10.1016/j.neuroimage.2017.06.030\nJayaram V, Barachant A. 2018. MOABB: Trustworthy algorithm benchmarking for BCIs. J Neural \nEng 15. doi:10.1088/1741-2552/aadea0\nJones SR. 2016. When brain rhythms aren’t ‘rhythmic’: implication for their mechanisms and \nmeaning. Curr Opin Neurobiol 40:72–80. doi:10.1016/j.conb.2016.06.010\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 July 19, 2024. ; https://doi.org/10.1101/2024.07.18.604064doi: bioRxiv preprint \n\nKilavik BE, Zaepffel M, Brovelli A, MacKay WA, Riehle A. 2013. The ups and downs of beta \noscillations in sensorimotor cortex. Exp Neurol 245:15–26. \ndoi:10.1016/j.expneurol.2012.09.014\nKobler RJ, Almeida I, Sburlea AI, Müller-Putz GR. 2020. Using machine learning to reveal the \npopulation vector from EEG signals. J Neural Eng 17. doi:10.1088/1741-2552/ab7490\nKoles ZJ. 1991. The quantitative extraction and topographic mapping of the abnormal components \nin the clinical EEG. Electroencephalogr Clin Neurophysiol 79:440–447. doi:10.1016/0013-\n4694(91)90163-X\nLangford ZD, Procyk E, Wilson CRE. 2023. Frontal oscillatory beta bursts have rhythmically \ndistinct regimes with differing functional relevance. bioRxiv 1–19. \ndoi:/doi.org/10.1101/2023.11.13.566594\nLeeb R, Lee F, Keinrath C, Scherer R, Bischof H, Pfurtscheller G. 2007. Brain-computer \ncommunication: motivation, aim, and impact of exploring a virtual  apartment. IEEE Trans \nneural Syst Rehabil Eng  a Publ  IEEE Eng Med Biol Soc 15:473–482. \ndoi:10.1109/TNSRE.2007.906956\nLenth R V. 2023. emmeans: Estimated Marginal Means, aka Least-Squares Means.\nLittle S, Bonaiuto J, Barnes G, Bestmann S. 2019. Human motor cortical beta bursts relate to \nmovement planning and response errors. PLoS Biol 17:1–30. \ndoi:10.1371/journal.pbio.3000479\nLotte F. 2014. A Tutorial on EEG Signal Processing Techniques for Mental State Recognition in \nBrain-Computer Interfaces To cite this version : Fabien Lotte . A Tutorial on EEG Signal \nProcessing Techniques for Mental State Recognition in A Tutorial on EEG Signal Processi. \nGuid to Brain-Computer Music Interfacing 32.\nLotte F, Bougrain L, Cichocki A, Clerc M, Congedo M, Rakotomamonjy A, Yger F. 2018. A review of\nclassification algorithms for EEG-based brain-computer interfaces: A 10 year update. J Neural\nEng 15. doi:10.1088/1741-2552/aab2f2\nLundqvist M, Miller EK, Nordmark J, Liljefors J, Herman P. 2024. Beta : bursts of cognition. Trends \nCogn Sci xx:1–15. doi:10.1016/j.tics.2024.03.010\nLundqvist M, Rose J, Herman P, Brincat S, Buschman T, Miller E. 2016. Gamma and beta bursts \nunderlie working memory. Neuron 90:152–164. doi:10.1016/j.neuron.2016.02.028\nMakeig S, Enghoff S, Jung TP, Sejnowski TJ. 2000. A natural basis for efficient brain-actuated \ncontrol. IEEE Trans Rehabil Eng 8:208–211. doi:10.1109/86.847818\nMoca V V., Bârzan H, Nagy-Dăbâcan A, Mureșan RC. 2021. Time-frequency super-resolution with \nsuperlets. Nat Commun 12:1–18. doi:10.1038/s41467-020-20539-9\nMüller-Gerking J, Pfurtscheller G, Flyvbjerg H. 1999. Designing optimal spatial filters for single-trial \nEEG classification in a movement task. Clin Neurophysiol 110:787–798. doi:10.1016/S1388-\n2457(98)00038-8\nNatraj N, Silversmith DB, Chang EF, Ganguly K. 2022. Compartmentalized dynamics within a \ncommon multi-area mesoscale manifold represent a repertoire of human hand movements. \nNeuron 110:154-174.e12. doi:10.1016/j.neuron.2021.10.002\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 July 19, 2024. ; https://doi.org/10.1101/2024.07.18.604064doi: bioRxiv preprint \n\nNeuper C, Wörtz M, Pfurtscheller G. 2006. Chapter 14 ERD/ERS patterns reflecting sensorimotor \nactivation and deactivation. Prog Brain Res 159:211–222. doi:10.1016/S0079-6123(06)59014-\n4\nPapadopoulos S, Bonaiuto J, Mattout J. 2022. An Impending Paradigm Shift in Motor Imagery \nBased Brain-Computer Interfaces. Front Neurosci 15. doi:10.3389/fnins.2021.824759\nPapadopoulos S, Szul MJ, Congedo M, Bonaiuto JJ, Mattout J. 2024. Beta bursts question the \nruling power for brain–computer interfaces. J Neural Eng 21:16010. \ndoi:10.1088/1741-2552/ad19ea\nPedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, Blondel M, Prettenhofer P , \nWeiss R, Bubourg V, Vanderplas J, Passos A, Cournapeau D, Brucher M, Perrot M, \nDuchesnay E. 2011. Scikit-learn: Machine Learning in Python Fabian. J Mach Learn Res \n12:2825–2830.\nPenfield W, Rasmussen T. 1950. The Cerebral Cortex of Man: A Clinical Study of Localization of \nFunction, Macmillan.\nPfurtscheller G. 1981. Central beta rhythm during sensorimotor activities in man. \nElectroencephalogr Clin Neurophysiol 51:253–264. doi:10.1016/0013-4694(81)90139-5\nPfurtscheller G, Berghold A. 1989. Patterns of cortical activation during planning of voluntary \nmovement. Electroencephalogr Clin Neurophysiol 72:250–258. doi:10.1016/0013-\n4694(89)90250-2\nPfurtscheller G, Brunner C, Schlögl A, Lopes da Silva FH. 2006. Mu rhythm (de)synchronization \nand EEG single-trial classification of different motor imagery tasks. Neuroimage 31:153–159. \ndoi:10.1016/j.neuroimage.2005.12.003\nPfurtscheller G, Lopes da Silva FH. 1999. Event-relared EEG/MEG synchronization and \ndesynchronization: basic principles. Clin Neurophysiol 110:1842–1857. doi:10.1016/S1388-\n2457(99)00141-8\nPfurtscheller G, Neuper C. 1997. Motor imagery activates primary sensorimotor area in humans. \nNeurosci Lett 239:65–68. doi:10.1016/S0304-3940(97)00889-6\nPfurtscheller G, Neuper C, Flotzinger D, Pregenzer M. 1997. EEG-based discrimination between \nimagination of right and left hand movement. Electroencephalogr Clin Neurophysiol 103:642–\n651. doi:10.1016/S0013-4694(97)00080-1\nPfurtscheller G, Stancák A, Neuper C. 1996. Post-movement beta synchronization. A correlate of \nan idling motor area? Electroencephalogr Clin Neurophysiol 98:281–293. doi:10.1016/0013-\n4694(95)00258-8\nRayson H, Szul MJ, El-Khoueiry P, Debnath R, Gautier-Martins M, Ferrari PF, Fox N, Bonaiuto JJ. \n2023. Bursting with potential: How sensorimotor beta bursts develop from infancy to \nadulthood. bioRxiv. doi:10.1101/2023.05.09.539976\nSadeghi S, Maleki A. 2019. Accurate estimation of information transfer rate based on symbol \noccurrence probability in brain-computer interfaces. Biomed Signal Process Control \n54:101607. doi:10.1016/j.bspc.2019.101607\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 July 19, 2024. ; https://doi.org/10.1101/2024.07.18.604064doi: bioRxiv preprint \n\nSeeber M, Scherer R, Müller-Putz GR. 2016. EEG oscillations are modulated in different behavior-\nrelated networks during rhythmic finger movements. J Neurosci 36:11671–11681. \ndoi:10.1523/JNEUROSCI.1739-16.2016\nSherman MA, Lee S, Law R, Haegens S, Thorn CA, Hämäläinen MS. 2016. Neural mechanisms of\ntransient neocortical beta rhythms : Converging evidence from humans , computational \nmodeling , monkeys , and mice. doi:10.1073/pnas.1604135113\nShin H, Law R, Tsutsui S, Moore CI, Jones SR. 2017. The rate of transient beta frequency events \npredicts impaired function across tasks and species. Elife. doi:10.7554/eLife.29086\nShlens J. 2014. A Tutorial on Principal Component Analysis. arXiv.\nSzul MJ, Papadopoulos S, Alavizadeh S, Daligaut S, Schwartz D, Mattout J, Bonaiuto JJ. 2023. \nDiverse beta burst waveform motifs characterize movement-related cortical dynamics. Prog \nNeurobiol 165187. doi:10.1016/j.pneurobio.2023.102490\nTangermann M, Müller KR, Aertsen A, Birbaumer N, Braun C, Brunner C, Leeb R, Mehring C, \nMiller KJ, Müller-Putz GR, Nolte G, Pfurtscheller G, Preissl H, Schalk G, Schlögl A, Vidaurre \nC, Waldert S, Blankertz B. 2012. Review of the BCI competition IV. Front Neurosci 6:1–31. \ndoi:10.3389/fnins.2012.00055\nTharwat A, Gaber T, Ibrahim A, Hassanien AE. 2017. Linear discriminant analysis: A detailed \ntutorial. AI Commun 30:169–190. doi:10.3233/AIC-170729\nTorrecillos F, Tinkhauser G, Fischer P, Green AL, Aziz TZ, Foltynie T, Limousin P , Zrinzo L, Ashkan \nK, Brown P , Tan H. 2018. Modulation of beta bursts in the subthalamic nucleus predicts motor \nperformance. J Neurosci 38:8905–8917. doi:10.1523/JNEUROSCI.1314-18.2018\nVidaurre C, Kawanabe M, Von Bünau P, Blankertz B, Müller KR. 2011. Toward unsupervised \nadaptation of LDA for brain-computer interfaces. IEEE Trans Biomed Eng 58:587–597. \ndoi:10.1109/TBME.2010.2093133\nVigué-Guix I, Soto-Faraco S. 2022. Using occipital -bursts to modulate behaviour in real-time. ⍺\nbioRxiv. doi:10.1101/2022.09.21.50882\nWessel JR. 2020. Β-Bursts Reveal the Trial-To-Trial Dynamics of Movement Initiation and \nCancellation. J Neurosci 40:411–423. doi:10.1523/JNEUROSCI.1887-19.2019\nYi W, Qiu S, Wang K, Qi H, Zhang L, Zhou P , He F, Ming D. 2014. Evaluation of EEG oscillatory \npatterns and cognitive process during simple and compound limb motor imagery. PLoS One \n9:1–19. doi:10.1371/journal.pone.0114853\nZhou B, Wu X, Lv Z, Zhang L, Guo X. 2016. A fully automated trial selection method for \noptimization of motor imagery based Brain-Computer interface. PLoS One 11:1–20. \ndoi:10.1371/journal.pone.0162657\nZich C, Quinn AJ, Bonaiuto JJ, O’Neill G, Mardell LC, Ward NS, Bestmann S. 2023. \nSpatiotemporal organization of human sensorimotor beta burst activity. Elife 12:e80160. \ndoi:10. 7554/ eLife. 80160\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 July 19, 2024. ; https://doi.org/10.1101/2024.07.18.604064doi: bioRxiv preprint \n\nSupplementary Material\nSup. Figure 1:  (a) Population average time-resolved decoding score and standard error for the beta burst\nconvolution (red), beta band (yellow) and mu-beta band (purple) filtering pipelines using a sliding window.\nDue to the different duration of the task for each dataset we restricted the time to the minimum trial period\ncorresponding to 3 seconds. (b – h)  Average, time-resolved decoding score and standard error per dataset\nof the same features using a sliding window. For each panel, the left subplot depicts the decoding results\nobtained using a single filter, while the right subplot depicts the results based on a filter bank technique. The\nbeta burst results are the same for the pair of each panel. The horizontal dashed line corresponds to the\nexpected chance level. Vertical dotted lines represent the onset and end of the trial period of each dataset.\nHorizontal lines on the top of each subplot show the results of pair-wise permutation cluster tests between\nthe beta bursts and either filtering technique, with correction for multiple comparisons at significance level of\n0.05. The color of the lines indicate which feature produces, on average, better results at any time point. A\nlack of color indicates no statistically significant differences between the compared features.\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 July 19, 2024. ; https://doi.org/10.1101/2024.07.18.604064doi: bioRxiv preprint \n\nSup. Figure 2:  (a) Population average time-resolved decoding score and standard error for the beta burst\nconvolution (red) and mu  band (turquoise) filtering  pipelines using an  incremental window.  Due to  the\ndifferent  duration  of  the  task  for  each  dataset  we  restricted  the  time  to  the  minimum  trial  period\ncorresponding to 3 seconds. (b – h)  Average, time-resolved decoding score and standard error per dataset\nof the same features using an incremental window. For each panel, the left subplot depicts the decoding\nresults obtained using a single filter, while the right subplot depicts the results based on a filter bank\ntechnique. The beta burst results are the same for the pair of each panel. The horizontal dashed line\ncorresponds to the expected chance level. Vertical dotted lines represent the onset and end of the trial period\nof each dataset. Horizontal lines on the top of each subplot show the results of pair-wise permutation cluster\ntests between the beta bursts and either filtering technique, with correction for multiple comparisons at\nsignificance level of 0.05. The color of the lines indicate which feature produces, on average, better results at\nany time point. A lack of color indicates no statistically significant differences between the compared features.\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 July 19, 2024. ; https://doi.org/10.1101/2024.07.18.604064doi: bioRxiv preprint \n\nSup. Figure 3:  (a) Population average time-resolved decoding score and standard error for the beta burst\nconvolution (red) and mu band (turquoise) filtering pipelines using a sliding window. Due to the different\nduration of the task for each dataset we restricted the time to the minimum trial period corresponding to 3\nseconds. (b – h) Average, time-resolved decoding score and standard error per dataset of the same features\nusing a sliding window. For each panel, the left subplot depicts the decoding results obtained using a single\nfilter, while the right subplot depicts the results based on a filter bank technique. The beta burst results are\nthe same for the pair of each panel. The horizontal dashed line corresponds to the expected chance level.\nVertical dotted lines represent the onset and end of the trial period of each dataset. Horizontal lines on the\ntop of each subplot show the results of pair-wise permutation cluster tests between the beta bursts and either\nfiltering technique, with correction for multiple comparisons at significance level of 0.05. The color of the lines\nindicate which feature produces, on average, better results at any time point. A lack of color indicates no\nstatistically significant differences between the compared features.\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 July 19, 2024. ; https://doi.org/10.1101/2024.07.18.604064doi: bioRxiv preprint \n\nSup. Figure 4:  (a) Population average time-resolved information transfer rate (ITR) and standard error for\nthe beta burst convolution (red), beta band (yellow) and mu-beta band (purple) filtering pipelines using a\nsliding window. Due to the different duration of the task for each dataset we restricted the time to the\nminimum trial period corresponding to 3 seconds. (b – h)  Average, time-resolved information transfer rate\n(ITR) and standard error per dataset of the same features using a sliding window. For each panel, the left\nsubplot depicts the ITR results obtained using a single filter, while the right subplot depicts the results based\non a filter bank technique. The beta burst results are the same for the pair of each panel. Vertical dotted lines\nrepresent the onset and end of the trial period of each dataset. Horizontal lines on the top of each subplot\nshow the results of pair-wise permutation cluster tests between the beta bursts and either filtering technique,\nwith correction for multiple comparisons at significance level of 0.05. The color of the lines indicate which\nfeature produces, on average, better results at any time point. A lack of color indicates no statistically\nsignificant differences between the compared features.\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 July 19, 2024. ; https://doi.org/10.1101/2024.07.18.604064doi: bioRxiv preprint \n\nFigure 5: Population-level statistical analysis based on linear mixed models when using a sliding window per\nfeature. (a) Average maximum decoding score. (b) Average maximum ITR. (c) Average latency to reach the\nmaximum decoding score. (d) Average latency to reach the maximum ITR. Error bars show 95% confidence\nintervals.  Hatches  indicate  the  use  of  a  filter  bank  technique.  Asterisks  indicate  statistically  significant\ndifferences among pairwise comparisons of the beta bursts and the rest of the features (* : p < 0.05. ** : p <\n0.01, *** : p < 0.001). A lack of asterisks implies no statistically significant differences. Note that the log\ntransform of the the maximum ITR and latency to maximum ITR were used for the statistical analysis (see\nMaterials and Methods), but panels b and d  depict results before applying the transformation for ease of\ncomparisons with panels a and c respectively.\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 July 19, 2024. ; https://doi.org/10.1101/2024.07.18.604064doi: bioRxiv preprint","source_license":"CC-BY-4.0","license_restricted":false}