A New Compound-Limbs Paradigm: Integrating Upper-Limb Swing Improves Lower-Limb Stepping Intention Decoding From EEG

Brain-computer interface (BCI) systems based on spontaneous electroencephalography (EEG) hold the promise to implement human voluntary control of lower-extremity powered exoskeletons. However, current EEG-BCI paradigms do not consider the cooperation of upper and lower limbs during walking, which is inconsistent with natural human stepping patterns. To deal with this problem, this study proposed a stepping-matched human EEG-BCI paradigm that involved actions of both unilateral lower and contralateral upper limbs (also referred to as compound-limbs movement). Experiments were conducted in motor execution (ME) and motor imagery (MI) conditions to validate the feasibility. Common spatial pattern (CSP) proposed subject-specific CSP (SSCSP), and filter-bank CSP (FBCSP) methods were applied for feature extraction, respectively. The best average classification results based on SSCSP indicated that the accuracies of compound-limbs paradigms in ME and MI conditions achieved 89.02% ± 12.84% and 73.70% ± 12.47%, respectively. Although they were 2.03% and 5.68% lower than those of the single-upper-limb mode that does not match human stepping patterns, they were 24.30% and 11.02% higher than those of the single-lower-limb mode. These findings indicated that the proposed compound-limbs EEG-BCI paradigm is feasible for decoding human stepping intention and thus provides a potential way for natural human control of walking assistance devices.


I. INTRODUCTION
B RAIN-COMPUTER interface (BCI) system detects, analyzes, and converts brain signals to commands for controlling external devices [1].Different techniques have been used to measure brain signals for BCIs.Electroencephalography (EEG) is one of such methods and has been widely applied because of its non-invasiveness, portability, and high temporal resolution [2].In general, there are four basic rhythms in EEG, mainly including delta (0-4 Hz), theta (4-7 Hz), alpha (8)(9)(10)(11)(12)(13), and beta (13-30 Hz) rhythms [3].Specially, brain waves over the motor cortex in the band of (8)(9)(10)(11)(12) are also referred to as the mu rhythm [3].As spatiotemporal amplitudes of mu and beta rhythms are modulated differently by moving different body parts, thus movement of specific limbs can normally be classified by analyzing the EEG rhythms [4].Then the output of classification is converted into commands for an external device.
Motor imagery (MI) is one of the most investigated BCI paradigms.In this paradigm, participants perform imagery of moving different body parts (e.g.left hand, right hand, tongue, and foot) without actually moving or muscle contraction [5].Compared to motor execution (ME) which involves actual movement, MI also can induce similar EEG rhythms [6].Currently, one of the most important BCI applications is leveraging MI information to control walking-assisted exoskeletons that enable paralyzed patients to perform walking training with their proactive intentions [7], [8].However, it should be noted that most BCI-MI paradigms lack the naturality to meet people's daily interaction habits, which might reduce comfort and effectiveness in practical applications.An intuitive and natural MI process, i.e., matching the way how human limbs naturally move [9], might be beneficial for promoting practical applications of BCI systems [10], [11].
Previous studies involving human walking MI-BCI can be generally divided into three types.The first type is imagining the start or stop of walking to activate or halt lower-limb rehabilitation exoskeletons [12].While the start and stop of human walking have been recognized with acceptable average accuracy, i.e., 78.96% in [12], 81.81% in [13], and 83.5% in [14], this paradigm did not take the unilateral discrimination of the legs into account.As for the second type, most studies have applied MI of the left hand, the right hand, and both feet to achieve multi-command decoding.For example, Lee et al. [15] instructed human subjects to imagine moving both hands, the left hand and right hand, mapping intentions of walking forward, turning left, and turning right, respectively, in which the overall accuracy was 91%.Gao et al. [16] applied MI of the left hand, the right hand, and both feet to reflect intentions of ascending stairs, descending stairs, and walking, respectively.Wang et al. [17] utilized the same imagined actions as in [15] to represent intentions of sitting down, standing up, and walking, respectively.Their best classification accuracy was 85.33%.However, walking is a movement that related to lower limbs.This type of paradigm is not a direct and natural way to reflect the human walking process.
The third type is imagining the specific action process of lower limbs, which is closer to walking movement.The stateof-the-art literature about this type can be further generalized into two major categories: 1) imagine the dorsiflexion or/and plantarflexion of feet; 2) imagine the extension or/and flexion of thighs and calves.In the first category, Hashimoto and Ushiba [18] analyzed the MI features of left-and right-foot dorsiflexion.Their results suggested that there was potential to differentiate the left and right foot MI via the beta rebound, with an average accuracy of 69.3%.Tariq et al. [19] suggested that the best average accuracy was observed for beta ERS, given as 74.9%.Gu et al. [20] used a similar MI paradigm of feet and graph theory features were exerted to find out the difference in cortical signal characteristics.Their classification accuracy averaged 67.13%.For the second category (leftand right-leg MI), Hsu et al. [21] implemented a lower-limb MI paradigm in which human participants lifted their left or right legs on a step tool.Their results indicated that the mu rhythm contributed more to the higher classification accuracy (71.25%).Similar research was conducted by Liu et al. [22], but it was found that there exhibited no distinct event-related desynchronization/synchronization (ERD/ERS) difference between the MI of the left and right leg.The classification results were not provided.Kline et al. [23] developed an fMRI-informed EEG approach to enable an accurate mapping of spatial brain activity corresponding to the EEG electrodes, thereby improving the accuracy of lowerlimb MI-BCI.Their average classification accuracy achieved 66.5%.A recent study guided participants to lie down on a specially-designed device [24], where they imagined leg flexion and extension processes of their legs.The average accuracy of this paradigm was 59.75%.
Compared with upper-limb paradigms, the classification accuracies of lower-limb MI paradigms were lower.There are two possible reasons: 1) the positions of the motor cortex controlling the movement of lower limbs are deeper and smaller (medial surface of the motor strip in the interhemispheric fissure) than those of upper limbs, which results in a minor spatial distance between the cortex areas corresponding to the left and right lower limb movements; 2) EEG does not have enough spatial resolution to probe the brain activity about 5cm perpendicular to the human scalp [25].Therefore, it could be hypothesized that a paradigm that integrates actions of both upper and lower limbs might benefit to improve the classification accuracy.Because this paradigm could leverage the significant unilateral activation characteristics of upper limbs.Besides, since natural human walking is a coordinated movement involving both upper and lower limbs, the participation of upper limbs should not be ignored while designing the natural human walking MI-BCI.
Several recent studies have considered investigating the brain response to multi-limbs tasks.Yi et al. implemented a series of studies [26], [27], [28] to explore characteristics of hand-foot combination tasks.These tasks were the imitation of playing drums by single-hand or compound actions of the unilateral hand and the contralateral foot.Their findings indicated the imagination of one hand and one foot contributed to the simultaneous activation of contralateral and mid-central brain areas.Besides, Weersink et al. [29], [30] suggested that there was stronger ERD when humans walked accompanied by natural arm swing than that of without-arm swing in the supplementary motor area (SMA).These findings provide a feasibility foundation for investigating the compound-limbs paradigm.
In our preliminary study [31], we designed an arm-leg combination ME paradigm to investigate the characteristics of compound-limbs tasks on five subjects.In this work, we (1) extended the participants of the ME experiment and provided a more thorough feature and classification analysis, and (2) expanded the EEG-based compound-limbs paradigm to the MI condition.This compound-limbs paradigm involved the execution/imagination of stepping forward the unilateral lower limb alternatively and swinging forward the contralateral upper limb simultaneously.The main contributions of this study are as follows: 1) a new EEG-BCI paradigm of compound-limbs tasks integrating upper-limb swing and lower-limb stepping has been proposed; 2) the effect of compound-limbs tasks on the cerebral cortex is illustrated in terms of the time-frequency analysis of EEG; 3) extensive experiments were conducted on two representative EEG-based BCI conditions, namely motor execution and motor imagery, to verify the feasibility of the proposed paradigm.It is hoped that this work will promote the practical application of natural human stepping MI-BCI in the field of lower-extremity exoskeletons.

II. METHODS
This section includes five parts.Parts A and B involve the introduction of the experiment protocol, data collection, and processing.The methods of feature extraction and classifiers were depicted in part C. The analytical methods of time-frequency characteristics and distinguishability of channels are explained in parts D and E.

A. Participants and Experiment Protocol
Sixteen healthy and BCI-naïve human subjects (seven males, nine females; mean ± SD aged: 23.8 ± 1.9 years) were recruited in ME (seven subjects) and MI (nine subjects) experiments, respectively.They all had a normal or correctednormal vision.Prior to the experiment, each of them gave their written informed consent and was confirmed to be right-handed according to the Edinburgh Handedness Inventory.The experiment protocol was carried out in accordance with the Declaration of Helsinki.The Human Participant Ethics Committee of the Southern University of Science and Technology approved this study (No. 20210024).
To make comparisons, four single-limb tasks and two compound-limbs tasks were designed in Fig. 1(a).Subjects were expected to execute or imagine the following actions: 1) the single upper limb swings forward and back (LA/RA); 2) the single leg kicks forward with the foot dorsiflexion and back to the original point (LL/RL); 3) the unilateral lower limb and the contralateral upper limb move simultaneously as described in single-limb action (LLRA/RLLA).The specific action process of one task was shown in Fig. 1   ME or MI, and rest.The total duration of each experiment was about 120 minutes, which included the device setting and instruction time for subjects.In each experiment, there was a video shown on the screen.In ME, subjects were required to perform actions shown in the video synchronously.In MI, subjects were instructed to familiarize themselves with the corresponding actual movements of imagery tasks for about 15 min.Then, at the formal MI experiment, the participants watched the video at first; after 2 s of baseline time, there would be a static picture of the task and then subjects started to imagine the specific action for about 4 s.In both ME and MI tasks, participants executed or imagined movement only once in a single trial.Table I summarized the important protocol parameters in ME and MI.

B. Data Acquisition and Preprocessing
The experiment was conducted in an electromagneticshielded environment, as shown in Fig. 2(a).Subjects sat on a chair located approximately 1.5m away from a 43-inch display monitor.They needed to lean against the chair with their backs straight, keeping their arms naturally vertical at two sides and Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.their legs relaxed.This sitting state could avoid significant artifacts caused by limb movement in ME and the fatigue that might be caused by the long-time standing posture in both ME and MI.A 56-electrode EEG head cap shown in Fig. 2(b) was used and the locations of electrodes were set according to the 10-20 international placement system [32].The referenced electrodes were placed on the earlobes, and the grounding electrode was placed on the forehead.EEG data were recorded by the 256-channel g.HIAMP biosignal amplifier (g.tec medical engineering GmbH, Austria) in a frequency range from 0.1Hz to 100Hz with a 1200Hz sampling frequency.Electrode impedances were kept to less than 30K .
Besides, four electrooculography (EOG) electrodes were used in both ME and MI experiments.Two of them were placed above and below the left eye, and the other two were put on the outer canthus of eyes, which was shown in Fig. 2(c).The left EOG channel was referenced with the right EOG channel to represent the horizontal EOG component.Similarly, the upper one was referenced with the lower channels to capture the vertical EOG component.In addition, there were four surface-electromyography (sEMG) channels set at trapezius near the neck in pairs respectively (Fig. 2(d)) in ME experiments.The sEMG channels were referenced with their same-side channels, capturing the muscle electrical activities.
All data preprocessing was conducted via the EEGLAB toolbox [33] in MATLAB 2020a (The MathWorks, Natick MA).The EOG and EMG signals were filtered from 0.1Hz to 20Hz and 5Hz to 120Hz, respectively.The abnormal EEG channels were detected visually and interpolated.Then, the EEG signals were band-pass filtered from 1Hz to 30Hz.
Next, we used the ASR function for correcting signals with artifacts.Subsequently, the EEG signals were re-referenced by the common average reference method [34].Then, the independent component analysis (ICA) was applied and the artifactual component was deleted by making a comparison with the processed EOG and sEMG signals.After that, we also used the ICALABEL function [35] to remove the other residual artifacts.Finally, the EEG data were down-sampled to 512Hz.

C. Time-Frequency Analysis
Event-related spectral perturbation (ERSP) presents mean spectral power variations during the execution time compared with baseline time (pre-task).It described the even-related (de)synchronization (ERD/ERS) features of brain signals in time-frequency domain.To calculate the ERSP, the power spectrum is computed over a sliding window and then averaged across all trials, which can be defined as follows [36]: where F k (k, f, t) is the spectral estimation of the channel k at frequency f f and time t, T denotes the number of discrete time points and t denotes the time point index.The B(k, f)B(k, f ) is the mean values of baseline interval, defined as follows: In this research, the ERSP values were calculated from the 0.5 s before cue onset to 4.5 s after cue onset between 1 Hz and 30 Hz.The data form 0.5 s before cue onset to cue onset was determined as baseline data of each trial.Since the ERD/ERS phenomenon induced by one subject was not apparent, this study superimposed and averaged the time-frequency spectra of the evoked signals under all trials of all subjects in same tasks.Topographic maps represent the mean ERSP values of all electrodes in the specific frequency band and time interval of EEG signals.In this research, we averaged ERSP values across all subjects to obtain a group-level ERSP topographic map that helped analyze the spatial scalp distribution of ERD/ERS information among the limb-related motor cortex.

D. Distinguishability of Motor-Related Channels
The EEG signals of different tasks correspond to different types of classes.Suppose the ERD amplitude of one task at all EEG channels in the specific frequency band is different from that of the other task at some electrodes.Then the two tasks can be distinguished.The Fisher score was used as the criterion of distinguishability in this paper.It is defined as [37]: Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.The greater the Fisher score corresponding to a channel, the separability of the two tasks was greater.

E. Feature Extraction and Classification Methods
The Common Spatial Pattern (CSP) algorithm has been widely applied for EEG signal feature extraction.This method designs a spatial filter related to a specific task to maximize the variance of the one-class signal while minimizing the variance of the other-class signals.The K components related to a specific task were extracted, eliminating irrelevant components and noise [38].Filter bank CSP (FBCSP) [39] is a is a variation of CSP.The full frequency bands were divided into multiple sub-bands; then the features of each band were extracted by the CSP filter; finally, the optimal K features among all sub-bands are selected by the feature selection algorithm.
In this study, the traditional CSP method was applied to data that was filtered from 4Hz to 30 Hz, which was the first feature extraction method.The second CSP method was named subject-specific CSP (SSCSP).The EEG data were filtered by a 1Hz interval from 1Hz to 30Hz, resulting in 29 frequencyband datasets.Next, the CSP was used to extract features on each sub-frequency band.Features of each sub-band were then used for classification through machine learning algorithms.The sub-frequency band that showed the best classification accuracy was determined as the subject-specific frequency band.The third feature extraction method is FBCSP.The EEG data were filtered by six filter banks, which were composed of 4-8 Hz, 8-12 Hz, 12-16 Hz, 16-20 Hz, 20-24 Hz, and 24-28 Hz.Then the best four features were selected from the 4 * 6=24 features by the Mutual Information-Based Best Individual Feature (MIBIF) algorithms.According to the ERSP analysis, preprocessed EEG data were extracted from 1 s to 4 s after the cue onset for feature extraction and classification.
Then, to avoid potential biases that may be introduced by using only one classifier, three classic machine learning algorithms were applied to classify the features of each paradigm.These algorithms included support vector machine (SVM), K-nearest neighbor (KNN), and random forest (RF).The classification tasks in each paradigm are as follows: (1) single-upper-limb tasks: LA and RA; (2) single-lower-limb tasks: LL and RL; (3) compound-limbs tasks: LLRA and.RLLA.The 10-fold stratified cross-validation was applied to evaluate the classification performance that presented by the average accuracy.

III. RESULTS
To investigate the feasibility of the proposed paradigm, this section first analyzes the results of time-frequency domain features in each type of task and then presents the final classification results.In Parts A, B, and C, the characteristics of the event-related spectral perturbation, the brain spatial activation distribution, and the separability of motor-related channels are described, respectively.Finally, the classification results are shown in part D.

A. Event-Related Spectral Perturbation
The event-related spectral perturbation across 1Hz to 30Hz at electrodes C3, C4 and Cz of all subjects were shown in Fig. 3. Noticeably, for all action patterns, long-lasting power decreases around 1s at C3, Cz, and C4 electrodes after cue onset, in both MI and ME.In the single-upper-limb swing task (LA and RA), for example, when subjects executed or imagined the action of the left upper limb, stronger ERD was shown at C4 than at C3.In contrast, no obvious difference between tasks of the single left leg and right leg (LL and RL) could be found.However, while combining actions of both arm and leg, it was easy to differentiate LLRA and RLLA tasks by contrasting the ERD responses at C3 and C4.Specifically, for the LLRA task, there was more lasting and distinct ERD at electrode C3.Its feature bands included 10-15Hz and 20-28Hz, which covered the mu and beta rhythms.For the RLLA task, the feature bands were similar to the LLRA task, but its feature electrode was C4.In terms of Cz, there showed stronger ERD around 20Hz in RL, LLRA and RLLA tasks during the ME condition; while in the MI state, the stronger ERD phenomenon only showed in single-upper-limb tasks.
The grand average amplitude of ERD(S) at mu and beta frequency bands were presented in Fig. 4. It could be seen that at the mu band, the difference between (1) LA and RA or (2) LLRA and RLLA at C3 and C4 were significant.For the electrode Cz, the significance level between LL and RL was higher than the other two types of tasks in ME.However, there was no significant difference between LL and RL tasks at the beta band in MI.

B. Brain Spatial Activation Distribution
From Fig. 5, it was found that the ERD phenomenon is more obvious in the mu and higher beta frequency bands.Therefore, the following contents will discuss and analyze the ERD topography in these two frequency bands.The one-dimensional characteristic data of ERD was obtained by averaging the time-frequency data in frequency and time windows.In Fig. 5, the brain topographies of six tasks were averaged across all subjects for ME and MI, respectively.
In both MI and ME conditions, there was a larger ERD area in the motor cortex at mu rhythm.To be specific, for the single-upper-limb task, the brain topographic map of the left upper limb showed that the attenuation of EEG power at the right brain region was more obvious than that of the left region.A similar contralateral dominance phenomenon also showed in the right upper-limb task.For single-lower-limb tasks, the difference between LL and RL was not distinct as that of single-upper-limb tasks.The common characteristic was that broader ERD showed at the central area both in LL and RL tasks.Nevertheless, compared with the single-lower-limb task, the substantial ERD phenomenon of compound limbs showed both at the contralateral hemisphere and the central SMA area, which reflected the activities of both the unilateral lower limb and the contralateral upper limb.However, as for the beta rhythm, the distinct contralateral dominant phenomenon only appeared in the compound-limbs tasks.
Furthermore, we compared the difference in induced ERD between MI and ME.Overall, the ERD feature difference was consistent in MI and ME.However, there still had some distinctions.For example, the noticeable ERD was also shown at the frontal and occipital lobes on the LA task.For the singleleg task, the ERD of LL was intense around electrode C2.For the compound-limbs task, unlike the ME condition, the ERD in the beta band presents contralateral distribution in compound-limbs MI tasks.

C. Separability Analysis of Motor-Related Channels
Fig. 6 and Fig. 7 showed the Fisher analysis of motor-related channels in three two-class classification tasks for all subjects in two situations.The top sub-figures in each figure presented the Fisher scores of each EEG electrode.The blue dotted lines were the subject-specific threshold levels.From Fig 6 (a) and Fig 7 (a), it could be seen that the number of channels that exceeded the threshold level in single-upper-limb paradigm and compound-limbs paradigm was larger than that of single-lower-limb paradigm for most of the subjects.
Fig. 6 (b) and Fig. 7 (b) showed the distribution of channels (red) that exceeded the threshold level.In single-upper-limb paradigm, the highly separable channels usually appeared at/around C3 and C4.For single-lower-limb paradigm, it showed that most of the red channels neared channel Cz.In terms of compound-limbs paradigm, the highly separable channels showed not only on the upper-limb controlled area but also around Cz.In addition, more than half of the subjects had overlapping red channels between single-upper-limb and compound-limbs paradigms.
Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.

D. Classification Performance
The performance of all subjects who participated in ME and MI experiments was illustrated in Table II and Table III, respectively.First of all, it is apparent that there was the same classification performance phenomenon in both ME and MI: single-upper-limb task > compound-limbs task > singlelower-limb task.Besides, for all three paradigms in ME and MI, the average accuracies of the SSCSP method were always the highest, followed by the FBCSP, and the CSP is the worst.
Take the results of the SSCSP as examples, mean accuracy of compound-limbs tasks was as well as that of single-upperlimb tasks in ME, with about 2% difference.Five of seven subjects' accuracies were higher than 90% in compound-limbs tasks.In contrast, the best average accuracy of single-lowerlimb tasks was 64.91% ± 6.67%.In MI, accuracies of

TABLE II AVERAGE ACCURACY IN MOTOR EXECUTION (%)
all tasks had declined.The highest average accuracies of single-upper-limb paradigm, single-lower-limb paradigm, and compound-limbs paradigm declined to 79.38%, 62.98% and 73.70%, respectively.However, there were still three subjects whose classification performance of compound-limbs tasks was between 80% and 90%.
One-way ANOVA statistical analysis was performed to further compare the classification performance differences between compound-limbs paradigm and single-limb paradigms across combinations of three feature extraction methods and three classification algorithms, which was shown in Fig. 8.It could be found that there was no significant difference between compound-limbs tasks and single-upperlimb tasks in both ME and MI.As for comparison with single-lower-limb paradigm, the classification performance of compound-limbs paradigm was significantly (p < 0.05) better than that of single-lower-limb tasks across all methods in ME.But in MI, its performance was only significant in (1) combinations of KNN and three feature extraction methods, and (2) combination of RF and FBCSP methods.

IV. DISCUSSION
This study proposed a compound-limbs EEG-BCI paradigm that integrated upper-limb swing and lower-limb stepping.This paradigm not only activated more brain regions but also had better classification performance.The feasibility and effectiveness of the proposed paradigm were verified in four aspects: (1) the event-related spectral perturbation characteristics, (2) the brain spatial activation distribution, (3) the separability of EEG electrodes, and (4) the classification performance.These results are discussed in detail as follows.
From the ERSP result, the contralateral dominance effect was not found in the single-lower-limb tasks.The ERD amplitude results of LL and RL tasks suggested that more neural activities were induced at C3 and C4 than that at Cz, which was further verified in the ERSP topographical distribution.As for the single-upper-limb tasks, the significant contralateral dominance characteristic was consistent with the previous studies [40], [41].In terms of the focus paradigm of this study, as we expected, when combing the unilateral lower limb with the contralateral upper limb, the brain activation level was increased compared with the single-lower-limb tasks.The possible explanation was that the participation of upper limb swing action made a contribution to the ERD increase, which was also found in [30].In addition, the ERSP topographical distribution maps (Fig. 4) among all tasks verified that compound-limbs tasks activated not only the contralateral hemisphere areas that were related to upper-limb actions but also activated the central parietal region that controls lower limbs, which was a similar phenomenon in [26].However, according to ERSP analysis of the single limb task, contralateral lower limb movements may also activate the ipsilateral upper limb control area.Therefore, it could be noted that there was an overlap of activated electrode areas between the LLRA and RLLA tasks.
The separability analysis of motor-related channels from all subjects revealed that the number of distinguishable channels of compound-limbs tasks was much more than that of single-lower-limb tasks in ME and MI, explaining the higher classification results of compound-limbs tasks.A comparison of this result with those of [25] confirmed that the regions related to the left and right lower limbs were overlapping and challenging to differentiate based on EEG signals.Besides, for single-upper-limb and compound-limb paradigms, most of electrodes that passed the threshold were the lateral ones.It is consistent with the brain spatial ERSP distribution result that the ERD phenomenon was distinct on the lateral side of motor cortex area.
Three feature extraction methods and three machine learning algorithms were adopted in this study.It is consistent with the suggestion of previous studies that the performance of the FBCSP method was better than that of traditional CSP method.However, there is a new finding in this study that using the SSCSP method improved the classification performance.Especially in the MI condition, the accuracies of SSCSP were about 16% higher than those of FBCSP.This suggested that extracting features from subject-specific 1Hz-width frequency bands might provide more effective and task-related information.Therefore, the matching of the subject-specific frequency band and paradigms probably should be considered.
In terms of the performance of the proposed compoundlimbs paradigm, the proposed compound-limbs EEG-BCI paradigm realized a higher classification accuracy of human stepping patterns than that of the single-lower-limb EEG-BCI paradigm.Based on the result of the SSCSP method, the best average accuracies of the compound-limbs paradigm under two conditions were 24.30% (ME) and 11.02% (MI) higher than those of the single-lower-limb paradigm.Besides, the average performance of the compound-limbs MI paradigm was 2.45%, 7.2%, and 13.95% higher than that of the previous single-leg paradigms [21], [23], [24], respectively.In Fig. 8, although the classification performance of the compound-limbs paradigm was not always significantly better than that of the single-lower-limb paradigm in MI, it could be verified that the involvement of upper-limb action is advantageous for classifying the patterns of human stepping based on statistical significance test results of ME.In fact, some people might not have the adequate ability to execute the MI task.These people are defined as BCI illiteracy [42].Although the multi-limb task could activate more motor cortex [43], [44], it also depends on more concentration.Therefore, subjects with lower accuracies of compound-limbs MI tasks may have a poor ability to imagine multi-limbs movement at the same time.These subjects might need more practice for the adaptation of these imagination tasks.Besides, it should also be noted that the highest average accuracies of the compound-limbs paradigm were 2.03% (ME) and 5.68% (MI) lower than those of the single-upperlimb paradigm.The overlapping activation regions between the LLRA and RLLA tasks may affect the degree of difference in features, making it more difficult to distinguish between them than the single-upper-limb tasks.Some studies found that the ipsilateral cortex was also activated during the single upper-limb movement [45].There existed similar neural signal activities between ipsilateral and contralateral hemispheres [46].Besides, it was reported that the lower-limb movement showed less lateralization in the motor cortex [47].Based on these single upper-limb findings, we proposed a hypothesis that compound-limbs tasks might produce more similar neural activities in both upper-limb and lower-limb motor control cortex areas.Therefore, the classification accuracy of compound-limbs tasks could be lower than that of the single-upper-limb tasks.However, benefiting from larger activation levels of upper limbs, the difference between compound-limbs tasks was easier to be decoded than that of single-lower-limb tasks.
Although the classification accuracy of the single-upperlimb paradigm was better than the proposed paradigm, the single-upper-limb paradigm is not a natural and direct way to reflect human walking intention.In terms of the previous hand-foot paradigm [26], [27], [28], the imagination of playing drums is still not similar with compound-limbs action of human walking.In contrast, the proposed paradigm involved simultaneous actions of the unilateral lower limb and the contralateral upper limb, which is closer to the natural human stepping posture.Besides, it is suggested that the swing of the upper limbs plays an important role in human walking [30], [48].One of the advantages of this proposed compound-limbs paradigm is that it might be beneficial for gait rehabilitation, such as improving interlimb coordination and trunk stability after stroke [49], [50].Therefore, the possible application of the proposed compound-limbs paradigm is to drive lower-limb rehabilitation exoskeletons by decoding human MI information.This system might not only present superior performance in decoding human stepping intentions but also contribute to coordination ability improvement between upper and lower limbs.
Despite the fact that this work proposed and proved the feasibility of the new paradigm for recognizing human walking intention based on EEG signals, there are several limitations worthing mentioned for practical applications.First, developing a fast and effective EEG online preprocessing platform is challenging.Signal quality plays an important role in online classification performance.Therefore, we have to exploit a real-time artifact-reduction algorithm to improve the signal quality.Second, considering BCI systems requiring robust decoding abilities in real-time scenarios [51], [52], the current offline decoding accuracy (73.70%) needs to be further improved in terms of feature extraction and decoding models.Third, the current compound-limbs ME paradigm may not be appropriate for patients since they generally have difficulty in moving their upper and lower limbs coordinatively and naturally.

V. CONCLUSION
This study proposed a new compound-limbs BCI paradigm to decode the human lower-limb stepping intention from EEG.The paradigm involved both the stepping action of the unilateral lower limb and the swing action of the contralateral upper limb simultaneously.With the analysis of EEG characteristics, it can be seen that compound-limbs tasks activated both upper-and lower-limb-related brain regions and generated an evident difference between the stepping intention of the left and right sides.Furthermore, the average classification accuracies of compound-limbs paradigm were higher than those of the single-lower-limb paradigm on both the ME-and MI-EEG dataset.These findings demonstrated that decoding step intention from compound-limbs tasks is feasible and outperforms the traditional single-lower-limb paradigm.Future work will focus on integrating this EEG-BCI paradigm in real-life walking-assistive exoskeletons.

VI. SUPPLEMENTARY MATERIALS
The experimental video is attached along with the paper.

Fig. 1 .
Fig. 1.Experiment protocol.(a): The design of single-limb and compound-limbs tasks.; (b): The action process shown in the video, tasking the LLRA task as example; (c) (d): The paradigm timeline of motor execution and motor imagery respectively.Abbreviation notes: LA: left arm; RA: right arm; LL: right leg; RL: right leg; LLRA: left leg with right arm; RLLA: right leg with left arm.
(b), taking the LLRA task as an example.The processes of ME and MI experiments designed in this research were shown in Fig. 1 (c) and Fig. 1 (d).The whole ME and MI experiments were divided into eight and six sessions, respectively, and each section consisted of 36 trials.During each session, six tasks were randomly displayed.Between the two sessions, there would be an intermediate two-minute rest.Each trial included three stages: baseline,

Fig. 2 .
Fig. 2. The actual experimental environment (a), electrodes setup of EEG (b), EOG (c) and sEMG (d).Given the information of motor planning and control, electrodes in the red rectangle area were used for decoding.

Fig. 3 .
Fig. 3. Average ERSP figures of six tasks across all subjects at key electrodes C3, Cz, and C4 during ME (a) and MI (b) conditions.Red and blue indicates ERS and ERD respectively.The different time-frequency areas were marked with black borders.LA: left arm; RA: right arm; LL: right leg; RL: right leg; LLRA: left leg with right arm; RLLA: right leg with left arm.

Fig. 5 .
Fig. 5.The average brain topographical distribution based on ERSP values for the motor execution and motor imagery conditions respectively.

Fig. 6 .
Fig. 6.Fisher scores of electrodes covering motor cortex in ME condition.The spatial maps showed the EEG electrodes that exceeded the threshold level (red) in different tasks and conditions.The blue dotted lines represented threshold levels.

Fig. 7 .
Fig. 7. Fisher scores of electrodes covering motor cortex in MI condition.The spatial maps showed the EEG electrodes that exceeded the threshold level (red) in different tasks and conditions.The blue dotted lines represented threshold levels.

TABLE I THE
EXPERIMENTAL PROTOCOL PARAMETERS IN ME AND MI