Skip to main content
NCBI home page
Human Brain Mapping logoLink to Human Brain Mapping
. 2024 Oct 6;45(14):e70034. doi: 10.1002/hbm.70034

RELAX‐Jr: An Automated Pre‐Processing Pipeline for Developmental EEG Recordings

Aron T Hill 1,✉, Peter G Enticott 1, Paul B Fitzgerald 2,3, Neil W Bailey 2,3
PMCID: PMC11456615  PMID: 39370644

ABSTRACT

Automated EEG pre‐processing pipelines provide several key advantages over traditional manual data cleaning approaches; primarily, they are less time‐intensive and remove potential experimenter error/bias. Automated pipelines also require fewer technical expertise as they remove the need for manual artefact identification. We recently developed the fully automated Reduction of Electroencephalographic Artefacts (RELAX) pipeline and demonstrated its performance in cleaning EEG data recorded from adult populations. Here, we introduce the RELAX‐Jr pipeline, which was adapted from RELAX and designed specifically for pre‐processing of data collected from children. RELAX‐Jr implements multi‐channel Wiener filtering (MWF) and/or wavelet‐enhanced independent component analysis (wICA) combined with the adjusted‐ADJUST automated independent component classification algorithm to identify and reduce all artefacts using algorithms adapted to optimally identify artefacts in EEG recordings taken from children. Using a dataset of resting‐state EEG recordings (N = 136) from children spanning early‐to‐middle childhood (4–12 years), we assessed the cleaning performance of RELAX‐Jr using a range of metrics including signal‐to‐error ratio, artefact‐to‐residue ratio, ability to reduce blink and muscle contamination, and differences in estimates of alpha power between eyes‐open and eyes‐closed recordings. We also compared the performance of RELAX‐Jr against four publicly available automated cleaning pipelines. We demonstrate that RELAX‐Jr provides strong cleaning performance across a range of metrics, supporting its use as an effective and fully automated cleaning pipeline for neurodevelopmental EEG data.

Keywords: artefact, automated, child, cleaning, development, EEG, pipeline, pre‐processing

RELAX‐Jr is a fully‐automated toolbox for cleaning EEG data recorded from children. It is freely available as a plugin for EEGLAB and includes a graphic user interface. It is expected to facilitate effective and unbiased artefact removal in neurodevelopmental datasets.

graphic file with name HBM-45-e70034-g006.jpg

Summary.

  • RELAX‐Jr is a fully automated pipeline for pre‐processing developmental EEG recordings.

  • We assessed performance of the RELAX‐Jr pipeline across a range of cleaning metrics.

  • RELAX‐Jr demonstrated strong performance for reducing common artefacts across EEG recordings from children.

1. Introduction

Almost a century on from Hans Berger's seminal non‐invasive recordings of electrical brain activity in humans (Berger 1929), electroencephalography (EEG) remains a highly popular method for investigating functional brain dynamics in both health and disease. Key advantages of EEG are its millisecond temporal precision and cost‐effectiveness. These advantages have facilitated its integration into broad‐ranging experimental paradigms and clinical practice, where it is frequently used to interrogate both spontaneous and task‐related neural activity. The wide availability and affordability of EEG also makes it highly conducive to large‐scale cross‐sectional and longitudinal investigations into typical neurodevelopment, as well as for studying aberrant neural activity patterns across a range of neurodevelopmental disorders (Buzzell et al. 2023; Cellier et al. 2021; Gabard‐Durnam et al. 2019; Marshall, Bar‐Haim, and Fox 2002).

EEG data, however, are frequently contaminated by a number of artefacts, both biological and environmental, which, if not adequately removed or suppressed, can obscure the underlying neural signals of interest. Furthermore, many of the biological artefacts, such as electromyographic (EMG) activity, movement‐related interference, and eye blinks/ocular artefacts, are often more common and pronounced in children (Brooker et al. 2020; Herve et al. 2022). Additional common artefacts (that are not specific to children) include electrical line noise at either 50 or 60 Hz, electrocardiographic (ECG) related signals, and noise related to poor electrode impedance (Daube 2009; Sazgar and Young 2019; Tandle and Jog 2016). Pre‐processing strategies for reducing these artefacts can vary widely, both between investigators, and across laboratories, with a growing array of EEG artefact removal algorithms available to investigators (Jiang, Bian, and Tian 2019; Roy et al. 2021). Manual detection (via visual inspection) and removal of artefacts are frequently performed; however, this process is time‐consuming, subjective, often imprecise, and not easily scalable to large datasets (Delorme 2023; Mumtaz, Rasheed, and Irfan 2021). Manual artefact reduction also requires considerable operator expertise in order to accurately interpret the EEG signal, making it vulnerable to potential bias and inconsistency resulting from human influence (Fitzgibbon et al. 2007; Jas et al. 2017).

Advances in signal processing methods have seen automated and semi‐automated cleaning pipelines become increasingly utilised in EEG pre‐processing. These automated approaches have the benefit of reducing experimenter error and/or bias, improving efficiency, and creating transparent and replicable workflows, thus advancing reproducible and open science (De Blasio and Barry 2023). To this end, we previously released the open‐source Reduction of Electroencephalographic Arteacts (RELAX) software, which enabled fully automated pre‐processing of EEG data. RELAX showed favourable results for cleaning resting‐state and task‐related datasets from adult populations when compared to several other commonly used automated pipelines (Bailey, Biabani, et al. 2023; Bailey, Hill, et al. 2023). Development of RELAX was motivated by the shortage of fully automated pipelines available to the EEG community, as well as our experience that some data are not effectively cleaned (or are over‐cleaned) using existing independent component analysis (ICA)‐based approaches (Bailey, Biabani, et al. 2023; Dimigen 2020). In addition to RELAX, several other semi‐ and fully‐automated pipelines for pre‐processing EEG data have been developed in recent years (Chang et al. 2020; Gabard‐Durnam et al. 2018). While these have been mostly targeted for use in adult populations, some recent pipelines have been developed for use with data collected in children (e.g., Debnath et al. 2020; Flo et al. 2022). However, many automated EEG pre‐processing pipelines (including RELAX) have been tested only in adult data. EEG is often also used to assess neural activity across developmental populations (both typically developing and clinical), which can pose additional challenges for data cleaning. For instance, movement and muscle‐related artefacts are often more apparent in children (Herve et al. 2022), while their limited attentional capabilities often necessitate shorter recording times and less cognitively demanding tasks (Bell and Cuevas 2012; DiStefano et al. 2019; Herve et al. 2022).

Here, we present the RELAX‐Jr (Reduction of Electroencephalographic Artefacts for Juvenile Recordings) EEG pre‐processing software pipeline, which we have specifically developed for cleaning of EEG data recorded from children. This pipeline is based on the original RELAX software (Bailey, Biabani, et al. 2023; Bailey, Hill, et al. 2023), which implements multi‐channel Wiener filters (MWF) (Borowicz 2018; Somers, Francart, and Bertrand 2018) and/or wavelet‐enhanced independent component analysis (wICA) (Castellanos and Makarov 2006) to reduce or remove artefactual signals. However, RELAX‐Jr also includes important modifications such as the inclusion of the ‘adjusted‐ADJUST’ independent component analysis (ICA) artefact component selection algorithm designed for Geodesic electrode nets, which are frequently used in EEG recordings in children (Leach et al. 2020), and the use of the Preconditioned ICA for Real Data (PICARD) algorithm for maximum likelihood ICA (Ablin, Cardoso, and Gramfort 2018a, 2018b). Adjusted‐ADJUST is more sensitive to the increased noise often present in data collected from children, and also considers a larger range of frequencies to identify alpha peaks in the neural signal, as alpha peak frequencies are often lower in children (Leach et al. 2020; Marshall, Bar‐Haim, and Fox 2002). PICARD has been shown to perform similarly to the popular Infomax ICA algorithm (Bell and Sejnowski 1995), but with a much faster speed of convergence (Frank, Makeig, and Delorme 2022). Like RELAX, RELAX‐Jr is fully automated, meaning that no user input is required after initial cleaning settings are determined, thus promoting streamlined and consistent/unbiased batch processing of data files. By default, RELAX‐Jr receives as inputs raw datafiles in EEGLAB format (Delorme and Makeig 2004), and outputs cleaned continuous data referenced to the robust average reference ready for further segmentation (optional) and analysis (Bigdely‐Shamlo et al. 2015).

In this article, we apply and rigorously test four separate versions of the RELAX‐Jr pipeline, testing four specific parameter variations. These tests were implemented using a large heterogeneous test dataset of resting‐state EEG (both eyes‐open and eyes‐closed recordings) taken from typically developing children (4–12 years of age). We compare the results obtained from RELAX‐Jr against four popular automated pre‐processing pipelines: the Maryland analysis of developmental EEG (MADE) (Debnath et al. 2020), the Automated Pipeline for Infants Continuous EEG (APICE) (Flo et al. 2022), the Harvard Automated Processing Pipeline for Electroencephalography (HAPPE) (Gabard‐Durnam et al. 2018), and the Artefact Subspace Removal followed by ICA (ASR) approach (Chang et al. 2020). We show that RELAX‐Jr exhibits robust efficacy across a range of cleaning metrics, establishing it as an effective and unbiased automated method for processing EEG data collected from children.

2. Methods

2.1. Dataset

Each pre‐processing pipeline was applied to a dataset of resting‐state EEG recordings collected from a cohort of typically developing children spanning early‐to‐middle childhood (N = 136, age range: 4–12 years; 71 male; average age = 9.42 years, SD = 1.95). All participants were English‐speaking, and none had received a formal diagnosis of any neurological, psychiatric, or genetic disorder. Ethical approval was provided by the Deakin University Human Research Ethics Committee (2017–065), while approval to approach public schools was granted by the Victorian Department of Education and Training (2017_003429). The EEG data were recorded in a dimly lit room using a 64‐channel HydroCel Geodesic Sensor Net (Electrical Geodesics, Inc., USA) containing Ag/AgCl electrodes surrounded by electrolyte‐wetted sponges. Data were acquired either at Deakin University (Melbourne, Australia), or in a quiet room at the participant's school using NetStation software (version 5.0) via a Net Amps 400 amplifier using a sampling rate of 1 KHz. Electrode Cz was used as the online reference. Electrode impedances were checked to ensure they were < 50 kOhms prior to recording. The data were recorded for 2 min while participants sat with their eyes open and focussed their gaze at a fixation cross on a computer screen, and 2 min while participants had their eyes closed. Three of the 136 participants did not have complete eyes‐open recordings and were not included in this dataset. A simplified overview of the data analysis pipeline is provided in Figure 1.

FIGURE 1.

FIGURE 1

Open in a new tab

Overview of the data analysis pipeline. Eyes‐open and eyes‐closed resting‐state EEG recordings were processed using four separate versions of the RELAX‐Jr pipeline (MWF wICA, MWF Only, wICA ADJUST, ICA Subtract), or using one of four comparison pipelines (MADE, APICE, HAPPE, ASR). Quality metrics were then used to assess data cleaning performance.

2.2. RELAX‐Jr Pipelines

The REAX‐Jr pipeline offers a distinctive method for pre‐processing EEG data by employing a series of iterative cleaning stages that leverage multiple toolboxes and algorithms to eliminate artefacts. The process begins with initial data filtering with high, low, and bandpass filters, followed by implementation of the Early‐Stage Electroencephalography Pre‐processing Pipeline (PREP) to discard bad electrodes (Bigdely‐Shamlo et al. 2015). Next, electrode artefacts and noisy data sections are identified based on extreme outlying amplitudes, drift, kurtosis, voltage distribution, or muscle activity. The pipeline then applies MWF filters to address muscle activity, blinks, and eye movements or drift (Borowicz 2018; Somers, Francart, and Bertrand 2018). Subsequently, wavelet‐enhanced ICA is used to eliminate artefacts specifically identified by the adjusted‐ADJUST algorithm (Leach et al. 2020). Notably, REAX‐Jr enables users to easily adjust and customise these steps through an intuitive graphical user interface.

To optimise the RELAX‐Jr pipeline, we tested adaptations of four specific versions of RELAX: MWF wICA, MWF Only, wICA ADJUST, and ICA Subtract. These pipelines were selected from both the optimally performing pipelines suggested by the use of RELAX in adult data (Bailey, Biabani, et al. 2023), and the most commonly applied pipeline in the literature (ICA Subtract). We note that within the original version of RELAX (designed for adult recordings) the wICA and ICA Subtract methods used ICLabel (Pion‐Tonachini, Kreutz‐Delgado, and Makeig 2019) to identify artefact components, while RELAX‐Jr replaces ICLabel with adjusted‐ADJUST (which is designed for use in paediatric EEG data) (Leach et al. 2020). Key details of each of these methods are summarised in Table 1. For brevity, we have not included a detailed description of each method. An overview of the RELAX‐Jr pipeline is provided in Figure 2. A full depiction of the RELAX‐Jr GUI is available in Figures S1–S3, and detailed step‐by‐step details for installing and running the software can be found in the RELAX‐Jr GitHub Wiki. For further specific details of the various methods, the reader is encouraged to refer to our previous publications (Bailey, Biabani, et al. 2023; Bailey, Hill, et al. 2023).

TABLE 1.

Summary of the steps for each of the four implemented RELAX‐Jr pipelines.

MWF wICA ICA Subtract wICA ADJUST MWF Only
Filter 0.25–80 Hz bandpass, 47–53 Hz notch filter 0.25–80 Hz bandpass, 47–53 Hz notch filter 0.25–80 Hz bandpass, 47–53 Hz notch filter 0.25–80 Hz bandpass, 47–53 Hz notch filter
Bad channel rejection PREP's ‘findNoisyChannels’, then for each 1 s epoch, reject if > 5% of data show extreme values (defined in Figure 1) or log‐power log‐frequency slopes > −0.59 PREP's ‘findNoisyChannels’, then for each 1 s epoch, reject if > 5% of data show extreme values (defined in Figure 1) or log‐power log‐frequency slopes > −0.59 PREP's ‘findNoisyChannels’, then for each 1 s epoch, reject if > 5% of data show extreme values (defined in Figure 1) or log‐power log‐frequency slopes > −0.59 PREP's ‘findNoisyChannels’, then for each 1 s epoch, reject if > 5% of data show extreme values (defined in Figure 1) or log‐power log‐frequency slopes > −0.59
Initial outlying data period rejection After bad channels are rejected, mark remaining 1 s periods that exceed the same thresholds as per the bad channel rejection step for exclusion from the MWF cleaning template and rejection prior to wICA After bad channels are rejected, reject remaining 1 s periods that exceed the same thresholds as per the bad channel rejection step After bad channels are rejected, reject remaining 1 s periods that exceed the same thresholds as per the bad channel rejection step

After bad channels are rejected, mark remaining 1 s periods that exceed the same thresholds as per the bad channel rejection step for exclusion from the MWF cleaning template and rejection after cleaning
Initial artefact reduction Three sequential MWF runs, cleaning muscle activity first, then eye blinks, then horizontal eye movement and drift None None Three sequential MWF runs, cleaning muscle activity first, then eye blinks, then horizontal eye movement and drift
Second artefact reduction ICA computed using the PICARD algorithm. Artefacts identified using Adjusted‐ADJUST. Artefactual ICA components reduced using wICA ICA computed using PICARD. Artefactual ICA components subtracted (identified using Adjusted‐ADJUST) ICA computed using PICARD. Artefactual ICA components identified by Adjusted‐ADJUST. Artefacts reduced using wICA None
Open in a new tab

Abbreviations: Hz = hertz; ICA = independent component analysis; MWF = multi‐channel wiener filters; PICARD = preconditioned ICA for real data; PREP = electroencephalography pre‐processing pipeline; s = seconds; wICA = wavelet enhanced independent component analysis.

FIGURE 2.

FIGURE 2

Open in a new tab

Overview of the RELAX‐Jr pipeline. Raw continuous data files (A) can be loaded into the RELAX‐Jr pipeline using the graphical user interface (GUI; B), or via a MATLAB script (not shown). Various data cleaning options can then be selected within the RELAX‐Jr software depending on the specific user's preferences (C). RELAX‐Jr then processes the data to produce continuous EEG files cleaned of artefacts (D). As a final step, the cleaned data can be further segmented into epochs of a desired length, with any remaining extreme segments removed (E). A full depiction of the RELAX‐Jr GUI is available in Figures S1–S3, and detailed step‐by‐step details for installing and running the software can be found in the GitHub Wiki.

2.3. Comparison Pipelines

The comparison pipelines we tested were: (i) the Maryland analysis of developmental EEG (MADE) pipeline (Debnath et al. 2020), (ii) the Automated Pipeline for Infants Continuous EEG (APICE) (Flo et al. 2022), (iii) the Harvard Automated Processing Pipeline for Electroencephalography (HAPPE) (Gabard‐Durnam et al. 2018), and the Artefact Subspace Reconstruction approach followed by ICA subtraction of artefacts identified by Adjusted‐ADJUST (referred to as ASR) (Chang et al. 2020). A simplified summary of key steps for each pipeline is provided in Table 2.

TABLE 2.

Summary of the steps for each of the comparison pipelines.

HAPPE MADE APICE ASR
Filter 0.25 Hz high‐pass, low pass at 80 Hz. CleanLine to remove 50 Hz 0.25 Hz high‐pass, low‐pass 48 Hz 0.25 Hz high‐pass, low pass at 80 Hz 0.25–80 Hz with 47–53 Hz notch filter
Bad channel rejection Rejects channels with joint probability > 3 SD from mean of average log power from 1 to 125 Hz. Performed twice FASTER EEGLAB plugin used to identify bad channels using Hurst exponent, correlation with other channels, and channel variance. Channels with absolute z‐score > 3 on any measure considered bad Relative thresholds used to identify extreme values. Multiple artefact rejection cycles PREP's ‘findNoisyChannels’, then for each 1 s epoch, reject if > 5% of data show extreme values or log‐ power log‐frequency slopes > −0.59
Initial outlying data period rejection None None Bad time periods (algorithm dependent, multiple algorithms used) After bad channels are rejected, reject remaining 1 s epochs that exceed the same thresholds as per the bad channel rejection step
Initial artefact reduction wICA applied to all components ICA run on 1 Hz filtered copy of data. ICA weights then copied back to original data Targeted PCA applied to specific brief events (segments < 100 ms). wICA applied to high‐pass filtered copy of the data (‘good’ time samples) ASR: automatic detection of clean data periods to determine thresholds, followed by rejection of components with large variance (similar to principal component analysis), then reconstruction of the original electrode space data
Second artefact reduction Artefactual artefact components rejected (identified by MARA) Adjusted‐ADJUST plugin used to select ICs containing artefact Adjusted‐ADJUST plugin used to select ICs containing artefact Artefactual ICA components rejected (identified by ICLabel)
Open in a new tab

Note: Filters outlined in the table represent those used to compare with RELAX‐Jr pipeline and might differ from original default settings.

Abbreviations: FASTER = fully automated statistical thresholding for EEG artefact rejection; Hz = hertz; ICA = independent component analysis; MARA = multiple artefact rejection algorithm; MWF = multi‐channel wiener filters; PCA = principal component analysis; PREP = electroencephalography pre‐processing pipeline; s = seconds; wICA = wavelet enhanced independent component analysis.

2.4. Segmentation of the Cleaned Data

After cleaning with each of the pipelines, EEG channels that were rejected during the cleaning steps were interpolated using spherical interpolation and data were segmented into 2‐s non‐overlapping epochs. Epochs rejection thresholds were applied as per the default RELAX settings (i.e., single/all channel improbable data thresholds [SD]: 5 and 3, respectively; single/all channel improbable data thresholds [SD]; 5 and 3, respectively; log‐frequency log‐power slope threshold for detecting muscle activity: −0.31), except for absolute voltage amplitude thresholds, which were set slightly more conservatively (± 100 uV) to avoid removing any epochs containing high‐amplitude alpha activity.

2.5. Cleaning Quality Evaluation Metrics

2.5.1. Multi‐Artefact Cleaning Performance Indicators: SER and ARR

Cleaning quality metrics used previously by our group (Bailey, Biabani, et al. 2023; Bailey, Hill, et al. 2023) and others (Bertrand 2015; Somers and Bertrand 2016; Somers, Francart, and Bertrand 2018) were employed to examine pre‐processing performance by each of the pipelines. Two complimentary metrics; the Signal‐to‐Error Ratio (SER) and Artefact‐to‐Residue Ratio (ARR) were used to estimate cleaning efficacy and preservation of the signal, respectively. SER measures the signal remaining unchanged after cleaning EEG data within periods of the continuous EEG data that were initially identified as free from all artefacts prior to cleaning. Artefact types included in the artefact templates included muscle activity, blinks, horizontal eye movements, and voltage drift. To understand how clean and contaminated periods were determined, refer to Bailey, Biabani, et al. (2023).

The SER measure is derived by dividing the expected value of the squared signal amplitude in clean periods for each electrode in the original (unprocessed) data by the squared signal of the removed artefacts during clean periods in the MWF templates (Somers, Francart, and Bertrand 2018). An average is then taken across all electrodes, with weighting applied according to the amplitude of the artefact signal in each electrode proportional to the amplitude across all electrodes. This results in electrodes containing greater artefact contributing more to the SER score. Larger SER values indicate better cleaning performance. ARR was calculated by obtaining the expected value operator of the square of the signal removed by the artefact reduction processes, divided by the expected value operator of the square of the raw data from the periods defined as containing artefacts prior to cleaning, minus the removed artefact signal from these artefact contaminated periods (Somers, Francart, and Bertrand 2018). As with the SER, larger ARR values are indicative of better cleaning performance. Further, given the complementary nature of these metrics, to perform well, pre‐processing pipelines should be expected to achieve both high SER and ARR values (it is trivially easy but also unhelpful to clean artefacts very effectively if we are not concerned about also preserving the non‐artefact signal).

2.5.2. Eye Blinks

Blink‐amplitude‐ratio (BAR) measures were also used to examine the ratio of blink amplitude to data periods containing no blinks (Robbins et al. 2020). BAR was assessed both across frontal channels (fBAR; average across channels: Fp1, Fp2, F9, F10, AF3, AF4; that is, where blink amplitudes are typically maximal), as well as across all electrodes (allBAR). For these BAR measures, values close to 1 indicate optimal performance, with values below 1 indicating overcleaning and values larger than 1 indicating under cleaning (Robbins et al. 2020). Blink metrics were not assessed for the eyes‐closed recordings, and 35 files were omitted from the blink metric analyses because there were not enough blink‐specific segments of the EEG data left after removing epochs with multiple blinks.

2.5.3. Electromyographic Contamination

In order to provide an indication of how many epochs contained residual electromyographic (EMG) contamination after cleaning, we determined the number of EEG epochs with any electrode showing a log‐power log‐frequency slopes greater than −0.59, which previous research has indicated to reflect muscle activity (Fitzgibbon et al. 2016). Higher values represent poorer cleaning for this metric.

2.5.4. Proportion of Epochs Rejected

An important consideration for data cleaning is the number of artefact free data segments (epochs) that remain and are available for analysis after pre‐processing. It is typically beneficial to retain as much good quality data as possible. This can be particularly important in cases where, for example, recording durations are relatively short, as can be the case with EEG collected in neurodevelopmental cohorts who may only be able to remain still for brief periods. The RELAX‐Jr pipeline performs pre‐processing on, and outputs, continuous EEG data, as do the comparison pipelines we tested. As such, we subsequently segmented the cleaned data files into 2‐s contiguous epochs and then removed any epochs showing values higher than 5 SD from the mean at any electrode or 3 SD from the mean for all electrodes, or epochs showing values outside of a −100 to +100 microvolt window. This provided a metric corresponding to the proportion of data segments that still contained likely artefacts following cleaning, where lower values for the ‘proportion of epochs rejected’ metric (indicating a smaller proportion of epochs rejected) likely reflects better performance.

2.5.5. Alpha Power and Alpha Peak Detection

As a final assessment, we investigated, (i) differences in alpha power values between the eyes‐open and eyes‐closed recordings and (ii) the likelihood of detecting an alpha peak in the data following cleaning. To achieve this, we first parameterised the data into periodic (oscillatory) and aperiodic components using the fitting oscillations and one over f (FOOOF) algorithm (Donoghue et al. 2020). We then calculated the peak alpha spectral power for each participant (i.e., the detected peak with the highest power within the 7–13 Hz range) (Donoghue et al. 2020). This metric provides a test of the potential for the cleaning algorithms to reveal experimental effects, using a well‐established between condition comparison where a reduction in alpha power with eye‐opening is observed, that is, the ‘Berger Effect’ (Kirschfeld 2005). We also further examined the percentage of total electrodes that had a detectable alpha peak following spectral parameterisation for each of the pipelines. As alpha is the dominant cortical rhythm at rest, and alpha oscillations are ubiquitous in resting‐state neural recordings across much of the cortex (although most pronounced in posterior regions) (Edgar et al. 2023; Lew et al. 2021), we used this approach to assess the ability to detect alpha oscillations within the periodic EEG signal following processing with each of the pipelines.

2.6. Statistical Analysis

Statistical analyses were performed in R (version 4.0.3) (R Core Team 2020) and JASP (JASP Team 2023). Robust repeated measures ANOVAs based on trimmed means were used to compare each of the pre‐processing pipelines on the various cleaning quality evaluation metrics using the WRS2 package (Mair and Wilcox 2020). This approach is robust to violations of normality and homoscedasticity, while maintaining equivalent power to traditional parametric tests (Mair and Wilcox 2020). Significant omnibus ANOVAs were followed‐up with pairwise comparisons using robust post hoc t‐tests, which apply Hochberg's method to control for family‐wise error (Hochberg 1988; Mair and Wilcox 2020). We did not perform additional experiment‐wise multiple comparison controls (i.e., for each quality evaluation metric), to emphasise sensitivity for the detection of differences in cleaning outcomes, in alignment with our aim to assess the potential superiority of specific cleaning pipelines (Bailey, Biabani, et al. 2023; Bailey, Hill, et al. 2023; Bender and Lange 2001). Finally, we have provided scatterplots of the SER x ARR values for each pre‐processing pipeline to enable evaluation of these two complimentary metrics together (Figure 3). Heatmaps of the post hoc tests for each metric are also provided in Figures S15–S20. Bayesian analyses were also conducted to examine the strength of evidence supporting a power difference between eyes‐open and eyes‐closed recordings following cleaning by each pipeline.

FIGURE 3.

FIGURE 3

Open in a new tab

Scatterplots depicting the mean signal‐to‐error (SER) and artefact‐to‐residue (ARR) values for the eyes‐open and eyes‐closed recordings for each of the pre‐processing pipelines. Higher SER and ARR values reflect better cleaning of the EEG data.

3. Results

The omnibus ANOVA was significant for all metrics tested for both the eyes‐open and eyes‐closed datasets (specific details in the sections below). Due to the large number of comparisons across the various pipelines, we provide here a summary only of the most relevant results, with more detailed results from the follow‐up post hoc tests provided in Figures S15–S20. An example of a raw (unprocessed) and cleaned EEG trace (single subject) is provided in Figure 4, while further examples for all pipelines are provided in Figures S6–S14.

FIGURE 4.

FIGURE 4

Open in a new tab

Left: EEG signal from a single example subject (eyes‐open recording, 5 s segment) showing the raw data (with DC offset removed [top]) and following pre‐processing with RELAX‐Jr (MWF wICA pipeline [bottom]). Examples of several common artefacts are noted on the EEG trace. All scales are microvolts. Right: Power spectra from the same subject before (top) and after (bottom) cleaning with the same RELAX‐Jr pipeline (power spectra are from the entire 2 min recording).

3.1. Signal‐to‐Error Ratio and Artefact‐to‐Error Ratio

The omnibus ANOVA was significant for both the eyes‐open, F(3.57, 278.78) = 83.58, p < 0.0001, and eyes‐closed, F(3.71, 300.61) = 76.84, p < 0.0001, conditions, indicating a significant difference in SER values between the pipelines (see Figure 5 for plots of SER and ARR values). For the ARR, the omnibus ANOVA was significant for both the eyes‐open, F(3.82, 298.35) = 417.06, p < 0.0001, and eyes‐closed, F(3.61, 292.81) = 524.00, p < 0.0001, conditions. For both the eyes‐open and eyes‐closed data, when assessing the SER and ARR metrics together (Figure 3), the MWF wICA and MWF Only pipelines demonstrated a good middle‐ground between both values, with moderate scores on both metrics. In contrast, the ASR, wICA ADJUST, MADE, APICE, and ICA Subtract pipelines all showed relatively high SER values, but had lower ARR values, indicating that although these pipelines produced little distortion or reduction of the clean EEG signal, they were generally less effective at suppressing artefact. The most extreme examples of this were the ASR pipeline for the eyes‐open data, and the wICA ADJUST, ICA Subtract, and ASR pipelines for the eyes‐closed data, which demonstrated very high SER and low ARR values. In contrast, the HAPPE pipeline demonstrated very high ARR and very low SER values in both the eyes open and closed datasets, indicating that while it was effective at mitigating artefacts, it concurrently removed a substantial amount of the signal from the clean segments of the EEG recordings.

FIGURE 5.

FIGURE 5

Open in a new tab

Plots of (A) signal‐to‐error (SER) and (B) artefact‐to‐residue (ARR) values for the eyes‐open and eyes‐closed recordings. Grey horizontal lines denote the median, while black diamonds denote the mean. Higher values for both SER and ARR concurrently reflect better cleaning performance (higher ARR = better artefact reduction; higher SER = better signal preservation).

It is also interesting to note that for the eyes‐open data, the ASR pipeline resulted in higher values for SER but similar values for ARR compared to ICA Subtract, despite including ICA Subtract as one of the cleaning steps in the ASR pipeline. In contrast, MWF wICA showed higher ARR but lower SER than both pipelines that represented components of the combined MWF wICA pipeline (MWF Only and wICA ADJUST). One possible mechanism by which ASR (which combines ASR and ICA Subtract) could produce increased SER compared to ICA Subtract only is that the initial artefact reduction performed by ASR allowed the ICA Subtract step to better separate artefact components from neural components, better preserving the clean signal. However, MWF wICA applies an analogous approach, first reducing artefacts with MWF prior to wICA, so if this mechanism were accurate, then we might expect MWF wICA and wICA ADJUST to show the same pattern. As such, another potential mechanism worth considering is that ASR may enhance the amplitude of activity in the clean periods of the data, leading to higher SER values through an artificial increase in amplitude in clean periods rather than the preservation of signal in those clean periods.

3.2. Blink Amplitude Ratio

For the fBAR values taken from anterior electrodes in close proximity to the eyes, the omnibus ANOVA was significant F(3.03, 154.63) = 59.80, p < 0.0001. For the allBAR values (all electrodes), the ANOVA was also significant, F(3.32, 169.54) = 111.56, p < 0.0001. The HAPPE pipeline performed the best across the fBAR and allBAR metrics, followed by ASR for fBAR and MWF wICA for allBAR, with APICE performing the worst for both the fBAR and allBAR metrics (values closest to 1 = best performance; Table 3 for Mean and SD values). BAR values are depicted in Figure 6.

TABLE 3.

Means and standard deviations (in parentheses) of the quality evaluation metric values for the eyes‐open data.

Pipeline SER ARR fBAR allBAR EMG remaining Epochs removed
MWF wICA 3.16 (1.37) 10.89 (2.72) 1.27 (0.35) 1.16 (0.21) 0.003 (0.010) 0.07 (0.03)
MWF Only 3.88 (1.90) 9.84 (3.06) 1.40 (0.52) 1.22 (0.29) 0.003 (0.009) 0.09 (0.05)
wICA ADJUST 4.91 (2.95) 5.92 (3.32) 1.73 (0.65) 1.45 (0.35) 0.02 (0.045) 0.15 (0.12)
ICA Subtract 4.44 (3.24) 5.33 (3.17) 1.71 (0.64) 1.44 (0.34) 0.019 (0.044) 0.15 (0.13)
MADE 5.10 (3.73) 6.59 (4.01) 1.67 (0.67) 1.44 (0.39) 0.037 (0.059) 0.19 (0.15)
APICE 4.47 (2.79) 5.58 (2.88) 2.07 (0.63) 1.75 (0.39) 0.015 (0.037) 0.12 (0.10)
HAPPE 0.55 (0.62) 20.72 (5.17) 1.09 (0.08) 1.06 (0.06) 0.988 (0.044) 0.17 (0.31)
ASR 6.39 (3.53) 6.90 (3.46) 1.26 (0.19) 1.18 (0.11) 0.097 (0.097) 0.25 (0.21)
Open in a new tab

Note: Higher values for SER and ARR indicate better cleaning performance. For fBAR and allBAR, values ~1 indicate optimal cleaning performance, while for EMG remaining (i.e., proportion of epochs showing EMG after cleaning) and epochs removed (proportion of epochs removed by cleaning), lower values indicate better performance.

FIGURE 6.

FIGURE 6

Open in a new tab

Blink amplitude ratios across frontal electrodes (average across Fp1, Fp2, F9, F10, AF3, AF4; fBAR) and across all electrodes (allBAR). Grey horizontal lines denote the median while black diamonds denote the mean.

3.3. Proportion of Epochs Showing Muscle Activity After Cleaning

The omnibus ANOVA was significant for both the eyes‐open, F(1.2, 93.67) = 17231.98, p < 0.0001, and eyes‐closed, F(2.08, 168.17) = 48712.90, p < 0.0001, conditions. MWF wICA and MWF Only were top performers in terms of the proportion of epochs showing EMG after cleaning, with MWF Only showing overall strong performance for the eyes‐open dataset, and MWF wICA showing the best performance for the eyes‐closed data. For both pipelines, the amount of EMG remaining was extremely close to zero, and means were identical to three decimal places (see Tables 3 and 4 for Means and SDs). The order of performance was the same for the rest of the pipelines for both the eyes‐open and eyes‐closed datasets, with the next best (in terms of mean rank order) being APICE, followed by ICA Subtract, wICA ADJUST, MADE, ASR, and finally HAPPE (see Figure 7A [for non‐winsorised values see Figure S4]). The HAPPE pipeline showed relatively poor performance on this metric, with the majority of epochs showing log‐power log‐frequency slopes that exceeded thresholds for indicating residual EMG following cleaning.

TABLE 4.

Means and standard deviations (in parentheses) of the quality evaluation metric values for the eyes‐closed data.

Pipeline SER ARR EMG remaining Epochs removed
MWF wICA 4.57 (1.56) 8.57 (2.27) 0.000 (0.001) 0.08 (0.06)
MWF Only 5.42 (1.69) 8.03 (2.24) 0.000 (0.001) 0.10 (0.07)
wICA ADJUST 7.18 (3.88) 3.22 (2.37) 0.011 (0.029) 0.21 (0.17)
ICA Subtract 6.82 (3.89) 3.20 (2.36) 0.011 (0.028) 0.22 (0.17)
MADE 5.50 (4.12) 6.33 (5.58) 0.030 (0.046) 0.22 (0.16)
APICE 5.79 (2.73) 4.73 (2.51) 0.009 (0.023) 0.15 (0.13)
HAPPE 0.49 (0.70) 20.53 (5.71) 0.968 (0.075) 0.21 (0.34)
ASR 7.01 (4.73) 5.09 (3.04) 0.048 (0.066) 0.33 (0.25)
Open in a new tab

Note: For EMG remaining (i.e., proportion of epochs showing EMG after cleaning) and epochs removed (proportion of epochs removed by cleaning), lower values indicate better performance.

FIGURE 7.

FIGURE 7

Open in a new tab

(A) Proportion of epochs showing log‐power log‐frequency values above the −0.59 threshold for EMG activity remaining for each of the pre‐processing pipelines. Lower values reflect more effective cleaning of EMG activity. Grey horizontal lines denote the median while black diamonds denote the mean. Note the data presented here were first winsorised (z = ± 2.5) as outlying values from the MADE and ASR pipelines made it difficult to visualise differences between the pipelines. The HAPPE pipeline was also removed as it had a median value > 0.95. Plots of the full dataset can be found in Figure S4. (B) Proportion of epochs rejected after cleaning for each pipeline.

3.4. Proportion of Epochs Removed by Cleaning

The omnibus ANOVA was significant for both the eyes‐open, F(4.21, 328.75) = 53.70, p < 0.0001, and eyes‐closed, F(3.51, 284.08) = 38.39, p < 0.0001, conditions. For both the eyes‐open and eyes‐closed data, MWF wICA performed the best in terms of mean rank order, followed by MWF Only, both of which resulted in the rejection of only 10% of the data or less on average, with MWF wICA resulting in the preservation of more than 80% of epochs for all EEG files in the eyes open data. In contrast, the worst performers (in terms of mean values) in the eyes open data were MADE, ASR and HAPPE, with ASR and HAPPE pipelines rejecting 100% of epochs for some files, and ASR rejecting more than 25% of epochs on average. In the eyes closed data, wICA ADJUST, ICA Subtract, MADE, and HAPPE all rejected more than 20% of epochs on average, ASR rejected 33% of epochs, and both ASR and HAPPE rejected 100% of epochs for some EEG files (see Figure 7B).

3.5. Variance Explained by Experimental Manipulation (Berger Effect)

Figure 8 shows the frequency spectra of the oscillatory neural activity (i.e., after removal of the aperiodic signal) for both the eyes‐open and eyes‐closed datasets. This figure highlights the alpha band demonstrating the distribution of alpha power across the scalp, while Figure 9 depicts the individual differences between the datasets for each pipeline using a power value derived from all electrodes (root mean square [RMS]). All pipelines showed the expected pattern of enhanced alpha power during eye closure, relative to the eyes‐open condition. To statistically assess differences between the eyes‐open and eyes‐closed conditions in terms of the strength of alpha activity differences between eyes‐open and eyes‐closed conditions, a Bayesian approach was used to determine the strength of the evidence supporting a power difference between the two conditions. Specifically, RMS alpha power values were compared between the eyes‐open and eyes‐closed conditions using Bayesian paired‐samples t‐tests. Bayes factors provided extremely strong support for the alternative hypothesis for all pipelines, indicating an expected pattern of robust alpha power differences between the eyes‐closed and eyes‐open conditions (see Table S1 for complete table of results). APICE showed the highest Bayes factor value, followed by MADE, wICA ADJUST, ICA Subtract, MWF Only, ASR, and MWF wICA, with HAPPE showing the lowest Bayes factor value. Next, we computed difference scores ([eyes‐closed] – [eyes‐open]) for RMS power values for each individual participant after cleaning with each pipeline. We used these difference scores to statistically compare each of the pipelines using a one‐way ANOVA. The omnibus ANOVA was significant, F(7,1059) = 8.952, p < 0.001; however, post hoc tests (Bonferroni corrected) only revealed significant differences between the HAPPE pipeline and all other pipelines (all p < 0.001, with HAPPE showing a smaller difference than all other pipelines), with none of the other pipelines differing in terms of difference scores between eyes‐open and eyes‐closed power.

FIGURE 8.

FIGURE 8

Open in a new tab

(A) Power spectral density plots (POz electrode) of the eyes‐open and eyes‐closed data after removal of the aperiodic signal. Shaded line is 95% confidence interval. Translucent grey bar denotes the alpha frequency range (7–13 Hz). Accompanying topographic plots depict the distribution of alpha power across the scalp. (B) Raincloud plots showing differences in alpha power (root mean square [RMS] values) between the eyes‐open (EO) and eyes‐closed (EC) conditions for each of the pipelines.

FIGURE 9.

FIGURE 9

Open in a new tab

Plots of BF10 values for the Bayesian t‐tests comparing the eyes‐open and eyes‐closed RMS alpha power values. Note, x‐axis is log‐scale to improve visualisation across results for the different pipelines which had large disparity.

3.6. Alpha Peak Detection

Separate one‐way ANOVAs were run for the eyes‐open and eyes‐closed recording conditions comparing each pipeline in terms of the percentage of total electrodes that had a detectable alpha peak following spectral parameterisation with the FOOOF algorithm to account for the aperiodic signal (Figure 10). The omnibus ANOVAs were significant for both conditions (eyes open: F(7,1037) = 41.155, p < 0.001; eyes‐closed: F(7,1062) = 22.898, p < 0.001). For both conditions, Bonferroni corrected post hoc tests indicated that the HAPPE pipeline had a significantly lower percentage of electrodes with a detected alpha peak compared to all other pipelines (all p < 0.001). No significant differences were observed between any of the other pipelines. Additional plots comparing pipelines across theta (4–7 Hz) and beta (13–30 Hz) peaks are also available in Figure S5.

FIGURE 10.

FIGURE 10

Open in a new tab

Percentage of total electrodes having a detected peak within the alpha band after removal of the 1/f‐like aperiodic signal for the eyes‐open and eyes‐closed data after cleaning with each pipeline (average over all participants).

4. Discussion

This study sought to provide an overview and assessment of the newly developed RELAX‐Jr pre‐processing pipeline for use with EEG data recorded in children. This software provides an important extension to the original RELAX software, which was designed for, and extensively evaluated in, adult populations (Bailey, Biabani, et al. 2023; Bailey, Hill, et al. 2023). We have optimised this pipeline through the inclusion of the PICARD ICA algorithm, which has been demonstrated to perform equivalently to extended‐infomax ICA but is computationally much faster (Ablin, Cardoso, and Gramfort 2018a), as well as the adjusted‐ADJUST independent component classification algorithm specifically designed for use with data collected from children. This allows RELAX‐Jr to provide a comprehensive integrated MATLAB‐based approach for cleaning data collected from neurodevelopmental cohorts, which, in addition to the neural signal of interest, often also contain high levels of artefact. We tested four versions of the RELAX‐Jr software on resting‐state EEG data collected in children aged between 4 and 12 years of age. We also compared our software against four popular open‐source automated EEG cleaning pipelines.

Results from our comprehensive assessment indicate that RELAX‐Jr generally performed well overall across the majority of included metrics. In particular, the MWF wICA and MWF Only settings demonstrated exceptional performance in cleaning EMG activity and minimising the number of epochs rejected after data segmentation, while also providing amongst the highest performance for blink artefact reduction. Exceptions to this pattern were: (1) for BAR, where HAPPE outperformed all other pipelines; however, we note that HAPPE likely achieved this high performance at the expense of reducing the neural signal and (2) for the combined SER and ARR metrics, where MWF wICA provided higher ARR values at the expense of lower SER values, while other pipelines made the reverse trade‐off, emphasising SER over ARR. For example, ICA Subtract and wICA ADJUST showed higher SER values, suggesting better signal preservation; however, they were less effective at artefact removal, as evidenced by lower ARR values. This trend was particularly apparent in the eyes‐closed recordings, where these two pipelines achieved amongst the highest SER scores but the lowest ARR scores.

We also note that ASR performed strongly (i.e., ASR showed the best performance for SER), indicative of its robust cleaning performance. In contrast, APICE often performed worse than other pipelines, with an average fBAR value > 2, suggesting that blink periods remained at twice the amplitude of non‐blink periods after cleaning, highlighting the possibility of inadequate blink removal. Additionally, while the HAPPE pipeline excelled in reducing blink artefacts, the low SER values and low amplitude recordings noted post‐cleaning suggest that its excellent blink reduction performance may have come at the cost of over‐cleaning of the EEG data. This outcome aligns with the low amplitude recordings observed with HAPPE in adult studies when compared to the RELAX pipeline (Bailey, Biabani, et al. 2023).

It also remains unclear what the optimal SER:ARR ratio should be, and determinations of superiority from SER and ARR analyses can only be made when one pipeline shows a higher SER value while the ARR value remains constant (or vice versa), or when a pipeline outperforms another in both metrics. For example, the SER:ARR graph suggests ASR is superior to APICE, as ASR has a higher SER with approximately the same ARR value. Consequently, it is difficult to ascertain whether ASR's higher SER, but lower ARR, indicates superior performance compared to MWF Only's lower SER, but higher ARR, or vice versa. However, the results for experimental effects of interest indicated that MWF wICA and MWF Only preserved neural activity of interest for experimental research to a degree that was statistically equivalent to the top performing pipelines, suggesting that these two pipelines did provide equivalent preservation of neural activity of interest to the top‐performing pipelines, while also providing superior artefact cleaning. Finally, since fBAR and allBAR values were calculated only for eyes‐open EEG recordings, we cannot assess RELAX‐Jr's efficacy in removing blinks from eyes‐closed data. Our focus on eyes‐open data stems from the typical occurrence of blink artefacts in these conditions. Nevertheless, it is possible that eyes‐closed data might also sometimes contain some blinks, particularly in young children or infants who might not be capable of accurately following instructions during a recording session.

It may be interesting to consider the contrast to our conclusions for the optimum pipeline in adult data, where MWF wICA or wICA ICLabel were recommended as default pipelines (with wICA ICLabel recommended when a sufficient amount of data are available, as it produced higher effect sizes for task related frequency band power experimental outcomes despite inferior cleaning to MWF wICA). In the current study, although wICA ADJUST was adapted to optimise its application to childhood EEG recordings, MWF Only performed better, despite not being specifically adapted to childhood EEG recordings. It may be that wICA ADJUST was not able to sufficiently clean the childhood EEG data due to the increased frequency and severity of artefacts, allowing MWF Only and MWF wICA to show higher performance. Future research may be able to optimise MWF Only for application to childhood EEG recordings by systematically testing the thresholds used to identify blink, muscle, and horizontal eye movement artefacts. However, we note that the muscle artefact thresholds used within RELAX were identified through research that involved paralysing participant's scalp muscles prior to EEG to eliminate the potential for scalp EMG (Fitzgibbon et al. 2016), an approach that is unlikely to be feasible in children.

In addition to our results for artefact cleaning, it was also encouraging to see that all cleaning pipelines enabled clear differentiation between the eyes‐open and eyes‐closed recordings, indicating sensitivity to this experimental manipulation. Specifically, all pipelines revealed the expected finding of higher alpha power during the eyes‐closed relative to eyes‐open recordings, with Bayes factors showing extremely strong evidence in favour of the alternative hypothesis. Further, when comparing each of the pipelines directly (using difference scores between the [eyes‐closed] − [eyes‐open] RMS power values), there were no significant differences between the pipelines in their ability to detect the Berger effect (with the exception of HAPPE, which showed lower difference values compared to all other pipelines). Similarly, with the exception of HAPPE, the pipelines did not differ in terms of the percentage of electrodes with a detected alpha peak following removal of the aperiodic signal. This observation corroborates our previous results comparing pre‐processing pipelines in adults (Bailey, Biabani, et al. 2023). However, we note that the Berger effect (i.e., alpha blocking) is a robust phenomenon that typically produces large differences in EEG amplitude between conditions (Goncharova and Barlow 1990; Kirschfeld 2005; Niedermeyer 1997). As such, it is possible that more subtle differences, such as disease‐specific alterations in neural activity, or changes following therapeutic interventions, might be more heavily influenced by specific pre‐processing pipelines. Future work comparing pre‐processing pipelines across participants with various neurodevelopmental and/or neuropsychiatric diagnoses might be useful to explore this possibility.

4.1. Limitations and Future Directions

We only tested RELAX‐Jr on 64‐channel Geodesic Sensor‐Net EEG caps. As such, its effectiveness for cleaning data from higher density (e.g., 128 or 256 electrode) montages is uncertain. However, we do not foresee any obvious constraints that might prohibit applying RELAX‐Jr to higher density recordings other than a likely increase in analysis times arising from greater computational burden produced by the inclusion of larger data files. The RELAX‐Jr software was also assessed using a developmental dataset containing a relatively wide participant age range (4–12 years), and we did not assess its performance using data from infants or very young children. As EEG recordings in these populations can deviate substantially from older children and adults, including greater prevalence of low frequency activity (Hrachovy and Mizrahi 2016; Marshall, Bar‐Haim, and Fox 2002), it is possible that artefact cleaning performance in samples of infants might differ from that reported using the present cohort. Future work examining performance in younger cohorts and incorporating datasets with high‐density recordings, would therefore be beneficial.

Additionally, as RELAX‐Jr was assessed using a typically developing sample of children, its cleaning performance in clinical samples (e.g., autism, attention deficit hyperactivity disorder, epilepsy) remains to be established. We also note that lower density recordings, particularly those with < 32 electrodes, could achieve significantly poorer performance resulting from insufficient ICA decomposition, which may not adequately separate neural from artefactual components (Janani et al. 2018; Klug and Gramann 2021). For such recordings, the use of non‐ICA based methods for data pre‐processing should be considered. The efficacy of MWF for artefact reduction with < 32 electrodes is currently untested, but we suspect it may still perform adequately for reducing artefacts in recordings with < 32 electrodes. However, since the MWF approach acts as a spatial filter, its performance will have a lower boundary in terms of numbers of available electrodes, and we encourage future research to test where this lower boundary lies. We also note that recent automated pipelines have been developed specifically for data with low numbers of channels (e.g., HAPPILEE; Lopez et al. 2022) and may be beneficial for investigators analysing very low‐density EEG recordings.

We acknowledge that utilising the ‘proportion of epochs rejected’ as an evaluation metric presumes that any epochs exhibiting improbable voltage distributions or outlying values correspond to data segments containing artefacts. Nonetheless, this is a widely accepted assumption in EEG research. We further observe that the Adjusted‐ADJUST algorithm (Leach et al. 2020) integrated into the RELAX‐Jr pipeline for objective selection of independent components representing artefacts requires the presence of left and right anterior electrodes for classification of blinks and horizontal eye movements. It is therefore possible that, in rare cases, a file with very high levels of artefact across frontal regions could lead to problems if large numbers of electrodes are removed. In such cases, more liberal extreme rejection thresholds could be considered, and might still provide adequate performance (Delorme 2023). We also note that our tests of experimental effects only include the examination of differences in the alpha frequency band, and that the current results may not apply to event‐related potential analyses. In adult populations, our previous research with the RELAX software indicated that the wICA ICLabel setting was optimal for clean data where many epochs were available, and that MWF wICA might be preferred where data are noisier or fewer epochs are available. However, it is worth noting that more aggressive (e.g., 1 Hz) high pass filter settings are not appropriate for event‐related potential analyses (as commonly analysed slow latency event‐related potentials contain frequencies below 1 Hz), but also that all artefact reduction pipelines performed more poorly in adult data when 0.25 Hz high pass filters (which are appropriate for event‐related potential analyses) were implemented. As such, further research is required to determine the best approach for analysing event‐related potentials in childhood EEG data. Finally, we note that while extensive comparisons were made both across versions of the RELAX‐Jr pipeline, and between RELAX‐Jr and several other available pipelines, we did not perform additional comparisons between our automated approach and manual artefact selection/rejection approaches by human EEG experts. While such extensive comparisons would be informative, they were outside the scope of the present study, but could warrant additional investigation in future research.

4.2. Conclusion

Here, we have provided an assessment of the cleaning performance of the RELAX‐Jr software pipeline developed for pre‐processing of EEG data collected from neurodevelopmental populations. The aim of this software is to provide users with a versatile and fully automated toolbox for removing artefacts that frequently contaminate the EEG record, enabling reliable and reproducible cleaning of EEG datasets while preserving the neural signal and minimising user bias (and workload). Based on the results of our analyses, we recommend the MWF wICA implementation of the software for applications to child EEG data, given its strong performance across the range of metrics assessed, including the ability to maximise the number of epochs included for analysis, which can be an important consideration for developmental EEG recordings, which are often limited in length.

Disclosure

The RELAX‐Jr pipeline is publicly available under the terms of the GNU General Public Licence and can be download from: https://github.com/aronthill/RELAX‐Jr. Step‐by‐step instructions for downloading, installing, and implementing the pipeline can be found on the accompanying GitHub Wiki: https://github.com/aronthill/RELAX‐Jr/wiki.

Ethics Statement

The study was approved by the Deakin University Human Research Ethics Committee. Written informed consent was provided by the parent or legal guardian of each child prior to commencement of the study.

Supporting information

Data S1:

HBM-45-e70034-s001.pdf (6.7MB, pdf)

Acknowledgments

This research was supported by a Future Fellowship from the Australian Research Council awarded to P.G.E. (FT160100077). P.B.F. is supported by a National Health and Medical Research Council of Australia Investigator grant (1193596). In the last 3 years, P.B.F. has received equipment for research from Neurosoft and Nexstim. He has served on a scientific advisory board for Magstim and received speaker fees from Otsuka. He has also acted as a founder and board member for TMS Clinics Australia and Resonance Therapeutics. Open access publishing facilitated by Deakin University, as part of the Wiley ‐ Deakin University agreement via the Council of Australian University Librarians.

Funding: This research was supported by a Future Fellowship from the Australian Research Council awarded to Peter G. Enticott (FT160100077). Paul B. Fitzgerald is supported by a National Health and Medical Research Council of Australia Investigator grant (1193596).

Data Availability Statement

The data that support the findings of this study are available on reasonable request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.

References

  1. Ablin, P. , Cardoso J. F., and Gramfort A.. 2018a. “Faster ICA Under Orthogonal Constraint.” In Paper Presented at the 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 4464–4468. Calgary, Canada: IEEE. [Google Scholar]
  2. Ablin, P. , Cardoso J. F., and Gramfort A.. 2018b. “Faster Independent Component Analysis by Preconditioning With Hessian Approximations.” IEEE Transactions on Signal Processing 66, no. 15: 4040–4049. 10.1109/TSP.2018.2844203. [DOI] [Google Scholar]
  3. Bailey, N. W. , Biabani M., Hill A. T., et al. 2023. “Introducing RELAX: An Automated Pre‐Processing Pipeline for Cleaning EEG Data ‐ Part 1: Algorithm and Application to Oscillations.” Clinical Neurophysiology 149: 178–201. 10.1016/j.clinph.2023.01.017. [DOI] [PubMed] [Google Scholar]
  4. Bailey, N. W. , Hill A. T., Biabani M., et al. 2023. “RELAX Part 2: A Fully Automated EEG Data Cleaning Algorithm That Is Applicable to Event‐Related‐Potentials.” Clinical Neurophysiology 149: 202–222. 10.1016/j.clinph.2023.01.018. [DOI] [PubMed] [Google Scholar]
  5. Bell, A. J. , and Sejnowski T. J.. 1995. “An Information‐Maximization Approach to Blind Separation and Blind Deconvolution.” Neural Computation 7, no. 6: 1129–1159. [DOI] [PubMed] [Google Scholar]
  6. Bell, M. A. , and Cuevas K.. 2012. “Using EEG to Study Cognitive Development: Issues and Practices.” Journal of Cognition and Development 13, no. 3: 281–294. 10.1080/15248372.2012.691143. [DOI] [PMC free article] [PubMed] [Google Scholar]
  7. Bender, R. , and Lange S.. 2001. “Adjusting for Multiple Testing—When and How?” Journal of Clinical Epidemiology 54, no. 4: 343–349. 10.1016/S0895-4356(00)00314-0. [DOI] [PubMed] [Google Scholar]
  8. Berger, H. 1929. “Über das Elektrenkephalogramm des Menschen.” Archiv für Psychiatrie und Nervenkrankheiten 87, no. 1: 527–570. 10.1007/bf01797193. [DOI] [Google Scholar]
  9. Bertrand, A. 2015. “Distributed Signal Processing for Wireless EEG Sensor Networks.” IEEE Transactions on Neural Systems and Rehabilitation Engineering 23, no. 6: 923–935. 10.1109/TNSRE.2015.2418351. [DOI] [PubMed] [Google Scholar]
  10. Bigdely‐Shamlo, N. , Mullen T., Kothe C., Su K. M., and Robbins K. A.. 2015. “The PREP Pipeline: Standardized Preprocessing for Large‐Scale EEG Analysis.” Frontiers in Neuroinformatics 9: 16. 10.3389/fninf.2015.00016. [DOI] [PMC free article] [PubMed] [Google Scholar]
  11. Borowicz, A. 2018. “Using a Multichannel Wiener Filter to Remove Eye‐Blink Artifacts From EEG Data.” Biomedical Signal Processing and Control 45: 246–255. 10.1016/j.bspc.2018.05.012. [DOI] [Google Scholar]
  12. Brooker, R. J. , Bates J. E., Buss K. A., et al. 2020. “Conducting Event‐Related Potential (ERP) Research With Young Children: A Review of Components, Special Considerations and Recommendations for Research on Cognition and Emotion.” Journal of Psychophysiology 34, no. 3: 137–158. 10.1027/0269-8803/a000243. [DOI] [PMC free article] [PubMed] [Google Scholar]
  13. Buzzell, G. A. , Morales S., Valadez E. A., Hunnius S., and Fox N. A.. 2023. “Maximizing the Potential of EEG as a Developmental Neuroscience Tool.” Developmental Cognitive Neuroscience 60: 101201. 10.1016/j.dcn.2023.101201. [DOI] [PMC free article] [PubMed] [Google Scholar]
  14. Castellanos, N. P. , and Makarov V. A.. 2006. “Recovering EEG Brain Signals: Artifact Suppression With Wavelet Enhanced Independent Component Analysis.” Journal of Neuroscience Methods 158, no. 2: 300–312. 10.1016/j.jneumeth.2006.05.033. [DOI] [PubMed] [Google Scholar]
  15. Cellier, D. , Riddle J., Petersen I., and Hwang K.. 2021. “The Development of Theta and Alpha Neural Oscillations From Ages 3 to 24 Years.” Developmental Cognitive Neuroscience 50: 100969. 10.1016/j.dcn.2021.100969. [DOI] [PMC free article] [PubMed] [Google Scholar]
  16. Chang, C. Y. , Hsu S. H., Pion‐Tonachini L., and Jung T. P.. 2020. “Evaluation of Artifact Subspace Reconstruction for Automatic Artifact Components Removal in Multi‐Channel EEG Recordings.” IEEE Transactions on Biomedical Engineering 67, no. 4: 1114–1121. 10.1109/TBME.2019.2930186. [DOI] [PubMed] [Google Scholar]
  17. Daube, J. R. 2009. “Waveforms and Artifacts.” In Clinical Neurophysiology, edited by Daube J. R. and Rubin D. I., 3rd ed., 103–1144. New York: Oxford University Press. [Google Scholar]
  18. De Blasio, F. M. , and Barry R. J.. 2023. “It's Time to RELAX and Smell the Roses!” Clinical Neurophysiology 149: 176–177. 10.1016/j.clinph.2023.02.169. [DOI] [PubMed] [Google Scholar]
  19. Debnath, R. , Buzzell G. A., Morales S., Bowers M. E., Leach S. C., and Fox N. A.. 2020. “The Maryland Analysis of Developmental EEG (MADE) Pipeline.” Psychophysiology 57, no. 6: e13580. 10.1111/psyp.13580. [DOI] [PubMed] [Google Scholar]
  20. Delorme, A. 2023. “EEG is Better Left Alone.” Scientific Reports 13, no. 1: 2372. 10.1038/s41598-023-27528-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
  21. Delorme, A. , and Makeig S.. 2004. “EEGLAB: An Open Source Toolbox for Analysis of Single‐Trial EEG Dynamics Including Independent Component Analysis.” Journal of Neuroscience Methods 134, no. 1: 9–21. 10.1016/j.jneumeth.2003.10.009. [DOI] [PubMed] [Google Scholar]
  22. Dimigen, O. 2020. “Optimizing the ICA‐Based Removal of Ocular EEG Artifacts From Free Viewing Experiments.” NeuroImage 207: 116117. 10.1016/j.neuroimage.2019.116117. [DOI] [PubMed] [Google Scholar]
  23. DiStefano, C. , Dickinson A., Baker E., and Jeste S. S.. 2019. “EEG Data Collection in Children With ASD: The Role of State in Data Quality and Spectral Power.” Research in Autism Spectrum Disorder 57: 132–144. 10.1016/j.rasd.2018.10.001. [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. Donoghue, T. , Haller M., Peterson E. J., et al. 2020. “Parameterizing Neural Power Spectra Into Periodic and Aperiodic Components.” Nature Neuroscience 23, no. 12: 1655–1665. 10.1038/s41593-020-00744-x. [DOI] [PMC free article] [PubMed] [Google Scholar]
  25. Edgar, J. C. , Franzen R. E., McNamee M., et al. 2023. “A Comparison of Resting‐State Eyes‐Closed and Dark‐Room Alpha‐Band Activity in Children.” Psychophysiology 60, no. 6: e14285. 10.1111/psyp.14285. [DOI] [PMC free article] [PubMed] [Google Scholar]
  26. Fitzgibbon, S. P. , DeLosAngeles D., Lewis T. W., et al. 2016. “Automatic Determination of EMG‐Contaminated Components and Validation of Independent Component Analysis Using EEG During Pharmacologic Paralysis.” Clinical Neurophysiology 127, no. 3: 1781–1793. 10.1016/j.clinph.2015.12.009. [DOI] [PubMed] [Google Scholar]
  27. Fitzgibbon, S. P. , Powers D. M. W., Pope K. J., and Clark C. R.. 2007. “Removal of EEG Noise and Artifact Using Blind Source Separation.” Journal of Clinical Neurophysiology 24, no. 3: 232–243. 10.1097/WNP.0b013e3180556926. [DOI] [PubMed] [Google Scholar]
  28. Flo, A. , Gennari G., Benjamin L., and Dehaene‐Lambertz G.. 2022. “Automated Pipeline for Infants Continuous EEG (APICE): A Flexible Pipeline for Developmental Cognitive Studies.” Developmental Cognitive Neuroscience 54: 101077. 10.1016/j.dcn.2022.101077. [DOI] [PMC free article] [PubMed] [Google Scholar]
  29. Frank, G. , Makeig S., and Delorme A.. 2022. “A Framework to Evaluate Independent Component Analysis Applied to EEG Signal: Testing on the Picard Algorithm.” arXiv. doi:arXiv:2210.08409.
  30. Gabard‐Durnam, L. J. , Mendez Leal A. S., Wilkinson C. L., and Levin A. R.. 2018. “The Harvard Automated Processing Pipeline for Electroencephalography (HAPPE): Standardized Processing Software for Developmental and High‐Artifact Data.” Frontiers in Neuroscience 12: 97. 10.3389/fnins.2018.00097. [DOI] [PMC free article] [PubMed] [Google Scholar]
  31. Gabard‐Durnam, L. J. , Wilkinson C., Kapur K., Tager‐Flusberg H., Levin A. R., and Nelson C. A.. 2019. “Longitudinal EEG Power in the First Postnatal Year Differentiates Autism Outcomes.” Nature Communications 10, no. 1: 4188. 10.1038/s41467-019-12202-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
  32. Goncharova, I. I. , and Barlow J. S.. 1990. “Changes in EEG Mean Frequency and Spectral Purity During Spontaneous Alpha Blocking.” Electroencephalography and Clinical Neurophysiology 76, no. 3: 197–204. 10.1016/0013-4694(90)90015-C. [DOI] [PubMed] [Google Scholar]
  33. Herve, E. , Mento G., Desnous B., and Francois C.. 2022. “Challenges and New Perspectives of Developmental Cognitive EEG Studies.” NeuroImage 260: 119508. 10.1016/j.neuroimage.2022.119508. [DOI] [PubMed] [Google Scholar]
  34. Hochberg, Y. 1988. “A Sharper Bonferroni Procedure for Multiple Tests of Significance.” Biometrika 75, no. 4: 800–802. 10.1093/biomet/75.4.800. [DOI] [Google Scholar]
  35. Hrachovy, R. A. , and Mizrahi E. M.. 2016. Atlas of Neonatal Electroencephalography. 4th ed. New York: Demos Medical. [Google Scholar]
  36. Janani, A. S. , Grummett T. S., Bakhshayesh H., Lewis T. W., Willoughby J. O., and Pope K. J.. 2018. “How Many Channels Are Enough? Evaluation of Tonic Cranial Muscle Artefact Reduction Using ICA With Different Numbers of EEG Channels.” In Paper Presented at the 2018 26th European Signal Processing Conference (EUSIPCO). Rome, Italy: IEEE. [Google Scholar]
  37. Jas, M. , Engemann D. A., Bekhti Y., Raimondo F., and Gramfort A.. 2017. “Autoreject: Automated Artifact Rejection for MEG and EEG Data.” NeuroImage 159: 417–429. 10.1016/j.neuroimage.2017.06.030. [DOI] [PMC free article] [PubMed] [Google Scholar]
  38. JASP Team . 2023. “JASP (Version 0.18.1).” [Computer Software].
  39. Jiang, X. , Bian G. B., and Tian Z.. 2019. “Removal of Artifacts From EEG Signals: A Review.” Sensors 19, no. 5: 987. 10.3390/s19050987. [DOI] [PMC free article] [PubMed] [Google Scholar]
  40. Kirschfeld, K. 2005. “The Physical Basis of Alpha Waves in the Electroencephalogram and the Origin of the ‘Berger Effect’.” Biological Cybernetics 92, no. 3: 177–185. 10.1007/s00422-005-0547-1. [DOI] [PubMed] [Google Scholar]
  41. Klug, M. , and Gramann K.. 2021. “Identifying Key Factors for Improving ICA‐Based Decomposition of EEG Data in Mobile and Stationary Experiments.” European Journal of Neuroscience 54, no. 12: 8406–8420. 10.1111/ejn.14992. [DOI] [PubMed] [Google Scholar]
  42. Leach, S. C. , Morales S., Bowers M. E., et al. 2020. “Adjusting ADJUST: Optimizing the ADJUST Algorithm for Pediatric Data Using Geodesic Nets.” Psychophysiology 57, no. 8: e13566. 10.1111/psyp.13566. [DOI] [PMC free article] [PubMed] [Google Scholar]
  43. Lew, B. J. , Fitzgerald E. E., Ott L. R., Penhale S. H., and Wilson T. W.. 2021. “Three‐Year Reliability of MEG Resting‐State Oscillatory Power.” NeuroImage 243: 118516. 10.1016/j.neuroimage.2021.118516. [DOI] [PMC free article] [PubMed] [Google Scholar]
  44. Lopez, K. L. , Monachino A. D., Morales S., Leach S. C., Bowers M. E., and Gabard‐Durnam L. J.. 2022. “HAPPILEE: HAPPE in Low Electrode Electroencephalography, a Standardized Pre‐Processing Software for Lower Density Recordings.” NeuroImage 260: 119390. 10.1016/j.neuroimage.2022.119390. [DOI] [PMC free article] [PubMed] [Google Scholar]
  45. Mair, P. , and Wilcox R.. 2020. “Robust Statistical Methods in R Using the WRS2 Package.” Behavior Research Methods 52, no. 2: 464–488. 10.3758/s13428-019-01246-w. [DOI] [PubMed] [Google Scholar]
  46. Marshall, P. J. , Bar‐Haim Y., and Fox N. A.. 2002. “Development of the EEG From 5 Months to 4 Years of Age.” Clinical Neurophysiology 113, no. 8: 1199–1208. 10.1016/S1388-2457(02)00163-3. [DOI] [PubMed] [Google Scholar]
  47. Mumtaz, W. , Rasheed S., and Irfan A.. 2021. “Review of Challenges Associated With the EEG Artifact Removal Methods.” Biomedical Signal Processing and Control 68: 102741. 10.1016/j.bspc.2021.102741. [DOI] [Google Scholar]
  48. Niedermeyer, E. 1997. “Alpha Rhythms as Physiological and Abnormal Phenomena.” International Journal of Psychophysiology 26, no. 1: 31–49. 10.1016/S0167-8760(97)00754-X. [DOI] [PubMed] [Google Scholar]
  49. Pion‐Tonachini, L. , Kreutz‐Delgado K., and Makeig S.. 2019. “ICLabel: An Automated Electroencephalographic Independent Component Classifier, Dataset, and Website.” NeuroImage 198: 181–197. 10.1016/j.neuroimage.2019.05.026. [DOI] [PMC free article] [PubMed] [Google Scholar]
  50. R Core Team . 2020. R: A Language Environment for Statistical Computing. Vienna, Austria: R Foundation for Statistical Computing. [Google Scholar]
  51. Robbins, K. A. , Touryan J., Mullen T., Kothe C., and Bigdely‐Shamlo N.. 2020. “How Sensitive Are EEG Results to Preprocessing Methods: A Benchmarking Study.” IEEE Transactions on Neural Systems and Rehabilitation Engineering 28, no. 5: 1081–1090. 10.1109/TNSRE.2020.2980223. [DOI] [PubMed] [Google Scholar]
  52. Roy, V. , Shukla P. K., Gupta A. K., Goel V., Shukla P. K., and Shukla S.. 2021. “Taxonomy on EEG Artifacts Removal Methods, Issues, and Healthcare Applications.” Journal of Organizational and End User Computing (JOEUC) 33, no. 1: 19–46. 10.4018/JOEUC.2021010102. [DOI] [Google Scholar]
  53. Sazgar, M. , and Young M. G.. 2019. “EEG Artifacts.” In Absolute Epilepsy and EEG Rotation Review. Switzerland: Springer. [Google Scholar]
  54. Somers, B. , and Bertrand A.. 2016. “Removal of Eye Blink Artifacts in Wireless EEG Sensor Networks Using Reduced‐Bandwidth Canonical Correlation Analysis.” Journal of Neural Engineering 13, no. 6: 066008. 10.1088/1741-2560/13/6/066008. [DOI] [PubMed] [Google Scholar]
  55. Somers, B. , Francart T., and Bertrand A.. 2018. “A Generic EEG Artifact Removal Algorithm Based on the Multi‐Channel Wiener Filter.” Journal of Neural Engineering 15, no. 3: 036007. 10.1088/1741-2552/aaac92. [DOI] [PubMed] [Google Scholar]
  56. Tandle, A. , and Jog N.. 2016. “Classification of Artefacts in EEG Signal Recordings and EOG Artefact Removal Using EOG Subtraction.” Communications on Applied Electronics 4: 12–19. [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Data S1:

HBM-45-e70034-s001.pdf (6.7MB, pdf)

Data Availability Statement

The data that support the findings of this study are available on reasonable request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.

Articles from Human Brain Mapping are provided here courtesy of Wiley

RESOURCES