We collected public datasets for motor tasks, including MI and ME, from several resources, including journals Scientific Data (Scientific Data, n.d.) and Gigascience (GigaScience | Oxford Academic, n.d.), research projects (Deep BCI and BNCI Horizon), and dataset platforms (IEEE DataPort (IEEE DataPort, n.d.) and MOABB). The characteristics of each resource are as follows. Gigascience and Scientific Data are journals that provide the source of the dataset, and Gigascience shares the dataset through its database, while Scientific Data shares the dataset through such repositories as Figshare without requiring a separate license for both. Deep BCI and BNCI Horizon are research projects that collect shared datasets published in a paper or requested by the projects. Researchers can download data directly from the BNCI Horizon website, while Deep BCI can receive the dataset after obtaining consent from the data owner. IEEE Dataport is a data platform that provides a dataset and simple information, and MOABB is a BCI benchmark that offers data and available Python APIs. Among them were datasets released multiple times from other resources and datasets the projects requested.

We collected a total of 25 datasets based on the following criteria:

There were 17 MI datasets that met these criteria (Leeb et al., 2007; Grosse-Wentrup et al., 2009; Faller et al., 2012; Tangermann et al., 2012; Ahn et al., 2013a; Yi et al., 2014; Lee et al., 2016, 2019; Steyrl et al., 2016; Zhou et al., 2016; Cho et al., 2017; Shin et al., 2017; Kim et al., 2018; Ma et al., 2020; Wu, 2020; Zhou, 2020; Stieger et al., 2021), 4 ME datasets (Luciw et al., 2014; Brantley et al., 2018; Wagner et al., 2019; Schwarz et al., 2020), and 4 both MI and ME datasets (Schalk et al., 2004; Ofner et al., 2017; Kaya et al., 2018; Jeong et al., 2020). In this study, the naming of the dataset took the form of representing the first author’s last name and the year of the published paper or dataset. For example, Stieger (2021) indicates the dataset in Stieger et al. (2021).

We conducted a quantitative analysis of eight representative MI datasets to compare their compatibility: Ahn et al., 2013a; Yi et al., 2014; Cho et al., 2017; Shin et al., 2017; Kaya et al., 2018; Kim et al., 2018; Lee et al., 2019; Stieger et al., 2021. These were selected as comparable according to various conditions, such as device, instruction method, and cue type, and were measured by the four EEG devices used most commonly—Neuroscan SynAmps2, BranProduct BrainAmp, Neurofax EEG-1200, and Biosemi—and included the basic MI paradigm, left- and right-handed imagination. The paradigm of the dataset is similar in the framework of imagination according to the instruction (cue) after rest (fixation cross), but the details differed, such as cue display and imagination methods.

To assess the data quality and compatibility with other datasets, we calculated the classification accuracy of the binary class problems and the feature distance among other datasets. As each dataset has a different number of channels, sampling rate, and signal scale, each dataset needs to be projected onto the same feature space for comparison. We extracted the features using the CSP, which is one of the methods to classify binary MI problems used most commonly (Aggarwal and Chugh, 2019). Our goal was to quantify it as objectively as possible by analyzing different datasets comparatively under common conditions. The CSP constructs a spatial filter to maximize one class and minimize the other class on the binary conditions (Ramoser et al., 2000).

The time-series data that have undergone a series of pre-processing processes extracts the CSP features through the following process. Let x_i ∈ R^N×S be the pre-processed EEG signal that consists of channels N -by-sample S as a multi-channel time series for a particular class motor imagery i such as the left- and right-hand. The spatial covariance matrix is Cⁱ ∈ R^N×N, given in Eq. 1:

in which ⊤ presents the transpose of matrix. The objective function J of CSP to optimize the spatial filter follow Eq. 2 and w ∈ R^N×N is spatial filter:

We used filters to generate features in two ways depending on the purpose. The first feature f_i,j follows Eq. 3 given as a feature calculation method to calculate accuracy.

in which j = {1,…,m} present the dimensional feature vector, and the first and last m rows of the CSP filter maximize the features of each class by the objective function. The second feature f¯i,j is the normalized feature, which is divided by the number of samples S to consider the sampling rate difference between the datasets, and follows Eq. 4:

We compared the eight features vector generated from the CSP filter to analyze the dataset’s compatibility. The upper filter maximizes the left-hand imagery, while the lower filter maximizes the right-hand imagery. We selected the two upper and two lower filters, and the left- or right-hand signals were projected on the filters to yield eight features. We extracted the features and calculated the accuracies for each subject, and the same subject was considered a different subject if measured on any other day. Therefore, the number of subjects in some data increased compared to the results Kaya et al. (2018), Lee et al. (2019), and Stieger et al. (2021) reported previously.

Because each dataset has a different reference electrode, the raw signal was re-referenced through a common average reference, and then bandpass filtered by 8–35 Hz. This frequency band was chosen to cover a sufficient interval (Alpha and Beta rhythms) and was based on the literature (Naeem et al., 2009; Bai et al., 2014; Cho et al., 2017). The segmented and released datasets, they were filtered for each trial, and the data released with continuous raw signals were filtered and then subjected to epoching. The signal was segmented from 500 to 3000 ms based on the cue (onset). The window size may vary in the optimal period for each dataset, but this was chosen for common conditions because the maximum window size obtained from Ahn et al. (2013a) was up to 3 s. Cho et al. (2017) had an exceptionally additional pre-process, and the signal was divided by the signal gain value of 32.

We also selected four filters to extract features to calculate accuracy, divided each dataset into ten sets, and used seven to train and three to test a classifier model. Here, the CSP filter was constructed using only training data. We reported mean accuracy by shuffling the data every time, repeating this process ten times, and the classifier model used linear discriminant analysis (LDA) (Mika et al., 1999). We used t-distributed stochastic neighbor embedding (t-SNE), a non-linear dimension reduction technique that show into low-dimensional space to visualize high-dimensional CSP features (van der Maaten and Hinton, 2008). As it maps data to minimize the difference between the probability distributions in a high and low-dimensional space, the similarity relation between points is maintained while the dimensionality is reduced simultaneously.

This review divided the specifications into public, environmental, and experimental specifications. We defined them by naming them intuitively, as several papers often use different words to describe the same specification. Each definition is in the paper, and a complete list of definitions is provided in the Supplementary material.

Table 4 presents the brief recording settings of representative datasets for CSP analysis based on actual data that can be mismatched from what the references reported. The table consists of the parameters that may affect the compatibility between the datasets, such as device, paradigm, and the parameters. The signal power indicates the mean of the signal envelope of the pre-processed Cz channels during the entire session. However, in Ahn et al. (2013a), seven measured by Biosemi were calculated as the mean of the entire channel because there was no channel information. It can infer the range of amplitudes in each dataset.

As shown in Figure 3 and Yi et al. (2014) had the highest accuracy of 75.71 ± 14.69%, followed by 71.92 ± 15.70% for Lee et al. (2019), 69.27 ± 13.03% for Cho et al. (2017), 65.65 ± 15.17% for Stieger et al. (2021), 62.54 ± 9.01% for Kaya et al. (2018), 62.08 ± 10.20% for Ahn et al. (2013a), 62.47 ± 17.22% for Kim et al. (2018) and 58.45 ± 15.17% for Shin et al. (2017). Yi et al. (2014), Cho et al. (2017), and Lee et al. (2019) have high accuracy and the distribution of accuracy forms a normal distribution, including both good and poor performers.

The MI-BCI poor performer is typically one who demonstrates a 60%–70% accuracy (Ahn et al., 2013b), and we considered the minimum critical criterion 60%. Yi et al. (2014) had the lowest rate of 20%, followed by 28.9% for Cho et al. (2017), 29.63% for Lee et al. (2019), 40.0% for Ahn et al. (2013a), 46.51% for Kaya et al. (2018), 49% for Stieger et al. (2021), 54.55% for Kim et al. (2018), and 62.1% for Shin et al. (2017). The grand mean accuracy of 861 subjects (or sessions) was 66.28%, and the BCI poor performer ratio was 45.30%. However, of 861 subjects, Stieger’s work included 589, which may have biased the grand mean. We calculated the mean of each dataset’s normalized accuracy distribution (right panel of Figure 3) was 66.49%, and the BCI poor performer ratio, the black-filled area, was 36.49%.

Figure 4 illustrates eight CSP features for each dataset that are the grand mean of all subjects within the dataset overall. This figure indicates directly how well the CSP filter constructed the features to maximize the characteristics of each class in each dataset. Although the pattern of the features was similar according to the filter, the feature scales differed for each dataset.

Figure 5A presents a t-SNE map of the eight-dimensional CSP features using the Mahalanobis distance. The reason why it was used here is that we examined the distribution between the datasets considering the distance of each subject. Each circle is a single subject or session, and the marginal distribution represents the distribution of datasets and the relation between them datasets. As there is a scale difference between the features (Figure 4), it was remapped to normalize them by variance (Figure 5C). This mitigates the difference in the scale between datasets, and it is possible to confirm the t-SNE map based on the features’ pattern.

Stieger et al. (2021) was widely distributed in the t-SNE maps of CSP features, while the other datasets were generally cohesive between same datasets (Figures 5A, C). There were cases in which the distributions of datasets overlapped, such as Cho et al. (2017) with Lee et al. (2019), and Shin et al. (2017) with Kaya et al. (2018) and Yi et al. (2014). Ahn et al. (2013a) forms two clusters, and Kim et al. (2018) is distributed widely compared to the number of data points.

Figures 5B, D represents the Euclidean distance between datasets based on the center points as each dataset’s mean in the eight-dimensional features’ space. This was used here rather than the Mahalanobis distance because the straight-line distance of each dataset’s central point was obtained. It is the distance in the feature space, so the unit and the size of the number are meaningless and can be interpreted as a relative size.

In the distance between the raw CSP features (Figure 5B), the mean of the log distance value from the other datasets was 12.75 for Stieger et al. (2021), 7.56 for Lee et al. (2019), 7.41 for Yi et al. (2014), 7.36 for Kim et al. (2018), 7.35 for Cho et al. (2017), 7.31 for Kaya et al. (2018), 7.26 for Shin et al. (2017), and 7.21 for Ahn et al. (2013a). The dataset closest to each other were as follows: 4.61 between Ahn et al. (2013a) and Kaya et al. (2018), 5.20 between Shin et al. (2017) and Ahn et al. (2013a), 5.45 between Yi et al. (2014) and Cho et al. (2017), 5.58 between Lee et al. (2019) and Yi et al. (2014), 6.2 between Kim et al. (2018) and Shin et al. (2017), and 12.75 between Stieger et al. (2021) and Lee et al. (2019).

In the distance between the normalized CSP features (Figure 5D), the mean of log distance value from the other datasets was 1.57 for Ahn et al. (2013a), 1.56 for Cho et al. (2017), 1.34 for Stieger et al. (2021), 1.332 for Shin et al. (2017), 1.31 for Lee et al. (2019), 1.29 for Kim et al. (2018), 1.27 for Yi et al. (2014), and 1.21 for Kaya et al. (2018). The datasets closet to each other were as follows: −0.272 between Shin et al. (2017) and Kaya et al. (2018), 0.19 between Ahn et al. (2013a) and Cho et al. (2017), 0.24 between Kim et al. (2018) and Lee et al. (2019), 0.29 between Yi et al. (2014) and Kaya et al. (2018), 0.67 between Stieger et al. (2021) and Yi et al. (2014).

The number of public datasets has been increasing recently, but for them to be used actively, researchers need sufficient information to be available for comparative selection. Therefore, we scrutinized the details of datasets, which have been covered before rarely. This provides guidelines to which researchers can refer when using public datasets and confirms the trends in movement-related studies.

Among the datasets collected, the MI datasets were more numerous than the ME datasets, and both were focused on the MI paradigm (Table 1). Schalk et al. (2004) had the only data with more than 100 subjects for MI and ME. Except for that datum, there were only four MI datasets with 50 or more, and the remainder had 30 or fewer. Schwarz et al. (2020) had 30 or more ME datasets. It was confirmed that datasets that recruited far more subjects than BCI competitive datasets were released.

We compared and analyzed the BCI competitive datasets that are used frequently and the groups of datasets based on the public year they were released publicly (Figure 1). Although the number of samples is insufficient to analyze the trends by year, research trends can be confirmed by comparing BCI competitions released approximately 10 years ago, datasets released at that time, and those released recently.

We found that not only has the number of published datasets increased each year, but also the size of a single dataset has increased with respect to the number of subjects, electrodes, and sampling rate. Recent datasets include a larger mean number of trials than do old ones, and the datasets in 2020 and 2021 consisted often of multi-sessions measured over several days. In particular, the number of subjects per dataset has also increased in the recent datasets. Compared to the number of BCI competitive datasets with 9 subjects, the mean number of subjects in datasets in the past 5 years was 29.5, a significant increase. The dataset released recently tend to collect big data because data-driven deep learning and machine learning are used actively now in BCI research (Altaheri et al., 2021).

In the measurement environment, a relatively heavy device was preferred to a lightweight one. However, recently, an increasing number of BCI research studies are using lightweight EEG devices. LaRocco et al. (2020) reported the possibility of drowsiness detection studies based on lightweight EEGs, such as MindWave, EmotionalEpoc, and OpenBCI, and Krigolson et al. (2017) quantified ERP using a portable MUSE EEG system. On the other hand, high-cost, high-resolution EEG devices are relatively preferred in the MI and ME research studies. The datasets published more recently show a trend toward higher sampling rates and a greater number of electrodes. The public datasets were also measured in whole brain regions, except for Leeb et al. (2007) and Tangermann et al. (2012), which released data from constrained environments for competition purposes.

In addition, MI and ME had differences in the difficulty of conducting experiments by subjects. MI included primarily simple and discrete tasks because it is difficult when confirming the conduction of the work, while ME consists of relatively complex tasks that are performed as delicate movements, such as the object’s weight and tactile differences. Because both paradigms are sensitive to noise due to movement, additional electrode measurements have been preferred rather than lightening the environment in practical terms. Many datasets also measured electrodes capable of post-processing noise, such as EOG and EMG electrodes, and in particular, all ME datasets included motion-tracking sensors. This trend overall requires signals obtained with reliable EEG devices, as biomarker extraction in MI-BCI is complex and requires as many EEG resources as possible for accurate analysis (Lee et al., 2019).

The basic framework of the current MI paradigm is a cue-based format in which a subject gazes at the fixation cross (pre-rest), instructions are presented, such as directions (cue), and then imagination is conducted for a certain period. There are often paradigms in which feedback from images is presented using models learned from data obtained during calibration sessions (Mousavi et al., 2017). Many MI studies follow the conventional Graz BCI motor imagery paradigm (Pfurtscheller et al., 1993). Thus far, information about paradigms in MI experiments has not been considered important. The MI paradigm is fixed only in composition; according to this investigation, there are no specific established details.

Each experimental parameter can influence brain activity as a form of stimulus, and ultimately is likely to lead EEG signal differences between datasets, even when based on the same paradigm. The public datasets collected had enormous differences in the details about the paradigm parameters, such as block duration and cue type (Figure 2 and Table 2). The stimulus presented on the screen during the break before and after imagination largely used a fixation cross or blank. They could be used as a stimulus to minimize the effect and serve merely for eye fixation. There were various cue types in the imagination block: 14 for arrow; 3 for text; 3 for object; and 2 for figure. The length of the imagination block ranged from 1 to 10 s, and the length of one trial varied considerably from a minimum of 2.5 to a maximum of 29 s.

Brain-computer interface (BCI) researchers overlook the method of imagination among parameters often, which can make the differences in brain waves. It has been reported that the EEG of KMI corresponds to sensory imagination performs better in classifying MI than the EEG of visual motor imaging (VMI) corresponds to the visual network (Yang et al., 2021). Among the datasets in this study, seven provided instruction using only class information (SMI), eight clarified movement tasks (EMI), and six instructed the subjects to focus their muscles on movement (KMI). In the case of SMI datasets, information may have been omitted because the matter was not considered an essential factor to researchers.

Practicing the same movement to improve the imagination before the experiment, or helping the subject by clarifying the way to imagine, can also affect the results. Some datasets, such as those of Yi et al. (2014), Cho et al. (2017), and Ofner et al. (2017), explained that the subjects engaged in a practice process before MI, but most datasets cannot confirm this explanation. It was confirmed that the activity in the motor-related area increased more with the method of performing MI after the actual motion, not with a simple verbal instruction method (Hasegawa et al., 2017). Lorey et al. (2009) found that brain activation occurs more strongly in MI from the first-person perspective than the third-person perspective. Thus, researchers can improve MI performance through paradigm details that aid imagination.

We confirmed that the detailed elements of the experimental paradigm differed across datasets, and we believe that these need to be considered because it is quite possible that they affect brain waves. O’Shea and Moran (2017) pointed out that MI and ME research has been conducted continuously according to the motion simulation theory (Jeannerod, 2001), but there is a lack of understanding of the cognitive mechanisms that underlie MI. To study this in more detail, researchers need to share sufficient information about the paradigms they employ. Therefore, in the future, they need to design paradigm details more delicately in MI experiments and describe them in their papers.

We confirmed the dataset’s quality by comparing the CSP accuracies calculated under the same conditions. There is one limitation, in that we did not calculate each dataset’s optimal accuracy to equalize the conditions. If the length of the CSP input signal is increased further or the frequency band is widened, then some datasets may have increased accuracy. We also agree that the quality of signals and datasets can be quantified in many ways, and it is inappropriate to define quality unconditionally according to only mean accuracy. We evaluated the data’s quality on the assumption that researchers who use public datasets collected high-quality data for MI classifier model generation or explored MI biomarkers, and thus evaluated the data quality based on the mean accuracy and MI-BCI poor performer rate based on a threshold of 60%.

The mean accuracy was good in the order of Yi et al. (2014), Lee et al. (2019), Cho et al. (2017), Stieger et al. (2021), Kaya et al. (2018), Ahn et al. (2013a), Kim et al. (2018), and Shin et al. (2017). The low performer rate was low in the order of Yi et al. (2014), Cho et al. (2017), Lee et al. (2019), Ahn et al. (2013a), Kaya et al. (2018), Stieger et al. (2021), Kim et al. (2018), and Shin et al. (2017). Lee et al. (2019) and Cho et al. (2017) can be considered to have obtained high-quality data because they had a vast number of samples that were distributed normally. The mean accuracy and poor performer rates of the entire datasets were 66.28 and 45.30%, respectively, but this may be a biased result because Stieger et al. (2021) accounted for more than the majority of the entire sample. Therefore, the mean accuracy and poor performer rate obtained from the dataset’s normalized distribution were 66.49% and 36.49%, respectively. A meta-analysis of 861 people in different environments confirmed that approximately one-third of all MI-BCI users were the poor performers.

Poor BCI performer is a challenge that must be solved for BCI technology to develop. As with Kim et al. (2018)’s research objectives, many studies have attempted to improve BCI performance (Vidaurre and Blankertz, 2010), and we found that one-third of the many datasets we collected corresponded to theirs. In addition, although there was a difference in the ratio for each dataset, the fact that all of them included a certain ratio of poor performers suggests that they cannot continue to be excluded simply from the analysis.

When public datasets are used in MI studies, it may be necessary to take poor performers into account and analyze them. It has been found that it is better to use only subjects who have achieved a certain level of performance or more than all subjects to create a generalized model for MI (Won et al., 2021). As in this study, researchers would need to choose the optimal data because they vary depending on the study’s purpose. They can select datasets with a large number of high performers to form an MI-optimized model or datasets with a high proportion of poor performers to analyze their neurophysiological characteristics. In particular, the poor performers will be very important research subjects in future efforts to commercialize BCI. If this is not addressed, one-third of those who attempt to BCI technology will not be able to do so, even if they purchase it, which could be fatal to the interface technology. Because our study evaluated data quality based on accuracy, undervalued datasets may be a suitable sample for this study.

We inferred compatibility between datasets by referring to CSP features extracted under the same conditions. The expectation might be that the closer the distance to the other datasets within the feature space, the higher their compatibility, while the more heterogeneous the distribution, the lower the compatibility. This assumes that the signals can be compared in the same dimension because they were moved to the feature domain using the same methodology, but still reflected dataset-specific characteristics.

Figure 4 shows the difference in the scale of the CSP feature in each dataset that is related to the power value of the pre-processed signal in Table 4. The sum of the mean and the standard deviation of the signal power is more prominent in Stieger et al. (2021) than others, and the scale of the CSP feature is also the largest. Yi et al. (2014) and Lee et al. (2019) showed a relatively similar mean of CSP features value, but there were significant difference in the variance of the CSP features and the signal powers: 3.28 ± 1.52 for Lee et al. (2019) and 7.75 ± 5.30 for Yi et al. (2014). It is expected that the CSP features reflect the signal’s power and the dispersion of the individual data that constitute the dataset. However, the CSP feature is an indicator that is not affected by power alone. Kaya et al. (2018) and Kim et al. (2018) had similar signal power ranges, but there was a large difference in the feature scale. This is because the signal reflects multiple data information in addition to signal power as it transposes to the CSP feature domain.

We analyzed the dataset’s feature distribution using raw features, including each signal’s the power information (Figure 5A). In addition, we examined the distribution based on the pattern using the normalized CSP features (Figure 5C). Individual sessions within the same dataset may have formed clusters because they may be superior to that in other datasets. Although the shape of individual sessions’ (subjects) the distribution in Figures 5A, C differ the overlapping dataset distributions are nearly the same. There is overlap between Yi et al. (2014), Shin et al. (2017), and Kaya et al. (2018), and between Cho et al. (2017), Lee et al. (2019). Interestingly, Ahn et al. (2013a)’s data formed two distribution clusters, which may be due to measurements with two devices. In Figure 5, one cluster overlaps Cho et al. (2017) measured by Biosemi, and the other overlaps Shin et al. (2017) measured by BrainProduct. This is consistent with the finding that EEG devices can affect the data significantly (Melnik et al., 2017).

This is confirmed more clearly in Figure 5D, which shows the distance between datasets in a more eight-dimensional normalized feature space. Ahn et al. (2013a) and Cho et al. (2017) measured by Biosemi, Kim et al. (2018) and Lee et al. (2019) measured by Brainproducts, and Yi et al. (2014) and Stieger et al. (2021) measured by Neuroscan were the closest datasets to each other. Figure 5B shows that the scale of the signal affects the compatibility between the data greatly. Stieger et al. (2021) has a large signal power compared to other datasets and was the farthest from all other data. It is expected that if the signal scale is solved, the compatibility between datasets measured with the same company’s equipment will be the best.

As explained in section “4.2. Importance of paradigm parameters,” the paradigm parameters, such as the cue type, may affect the distance. Figure 5B shows that Shin et al. (2017) was closest to Ahn et al. (2013a) and Kim et al. (2018) was closest to Shin et al. (2017), who used an arrow stimulus. Yi et al. (2014) was closest to Cho et al. (2017), who a used text stimulus. Figure 5D shows that Kim et al. (2018) was closest to Lee et al. (2019) who used a text stimulus. Another factor is the method of instruction used. Figure 5B shows that Stieger et al. (2021) is closest to Lee et al. (2019), and in Figure 5D and citeBR80 is closest to Kaya et al. (2018), who used the same instruction method as EMI.

Although not confirmed in this study, details of EEG devices, such as electrode types, may affect signal compatibility as well. EEG was sensitive to noise previously, so electrodes were attached with gel to reduce impedance. Recently, however, several pin types have been developed as dry EEG with wet device quality for convenience (Di Flumeri et al., 2019; Heijs et al., 2021). Habibzadeh Tonekabony Shad et al. (2020) confirmed that the input impedance of the latest EEG system of dry electrodes and amplifiers is more than 100 MΩ, while the impedance of conventional wet electrodes is less. The electrode type can affect the signal power because of impedance, and the resulting baseline difference between EEG signals can affect compatibility. However, in our study, we could not confirm this because only some studies provided information about the electrode type, such as Kaya et al. (2018), Jeong et al. (2020), and Ma et al. (2020), who used a gel type electrode.

We found that there are cases in which data are disclosed that lack important details (see Supplementary Material 2). These can have a significant effect on the analysis, and some can lead to distorted results if not taken into account. To avoid this problem, we suggest that BCI researchers give attention to certain considerations when they release future datasets.

The EEG-BIDS project was established to develop a standard format to organize and share brain imaging datasets between laboratories better (Pernet et al., 2019). The format contains sufficient information for data analysis. We believe that BIDS could facilitate better use of public datasets and ultimately foster an open science trend in the field. However, some information may not be available in certain experiments, and collecting all of the information for BIDS or related standard formats may be another burden to researchers. Indeed, among the datasets, only Ma et al. (2020) was provided in BIDS format. Therefore, using a well-organized standard format is obviously encouraged, but what to include is another important issue when it is not an option to use standard. Here, we will discuss the minimal components that researchers should consider when they share their datasets with the public.

Here, we present components that researchers who cannot follow the BIDS form must have when they release data. The list includes only the bare minimum of information to encourage the dataset to be published. The essential components that a dataset must contain are as follows:

We recommend sharing both the MATLAB file format, used conventionally in EEG analysis, and the Python file format, whose use is increasing with deep learning research. It is recommended that the signal be released as consecutive signals rather than as separate signals by trial. As there is a concern about the edge effect in the bandpass filtering used commonly for short-length signals in EEG analysis, we encourage opening the raw signal to maximize the possibility of analysis. In addition, this may help infer paradigm specifications, such as ITI, which is not presented to the user through event information. Approximately 71% of the MI and ME datasets collected met this recommendation, and the datasets that can be loaded into the MATLAB, even if they are not “mat” format exceptionally, were included.

One crucial point is that all information should be included in the data structure so that experimental information can be confirmed only with a data file alone (Figure 6). Information in collected the public datasets we collected was scattered, for example, papers’ references, separate text or word files, and variables within the data file. To prevent the users from having to make unnecessary efforts, the essential information at least needs to be presented intuitively within the data.