The performance of a low-cost bio-amplifier on 3D human arm movement reconstruction

Kayode P. Ayodele 1 , Eniola A. Akinboboye 1  and Morenikeji A. Komolafe 2
  • 1 Department of Electronic and Electrical Engineering, Obafemi Awolowo University, Ile-Ife, Nigeria
  • 2 Department of Medicine, Obafemi Awolowo University, Ile-Ife, Nigeria
Kayode P. Ayodele
  • Corresponding author
  • Department of Electronic and Electrical Engineering, Obafemi Awolowo University, Ile-Ife, Nigeria
  • Email
  • Search for other articles:
  • degruyter.comGoogle Scholar
, Eniola A. Akinboboye
  • Department of Electronic and Electrical Engineering, Obafemi Awolowo University, Ile-Ife, Nigeria
  • Search for other articles:
  • degruyter.comGoogle Scholar
and Morenikeji A. Komolafe

Abstract

Objectives

In this study, the performance of OpenBCI, a low-cost bio-amplifier, is assessed when used for 3D motion reconstruction.

Methods

Eleven scalp electrode locations from three subjects were used, with sampling rate of 125 Hz, subsequently band-pass filtered from 0.5 to 40 Hz. After segmentation into epochs, information-rich frequency ranges were determined using filter bank common spatial filter. Simultaneously, the actual hand motions of subjects were captured using a Microsoft Kinect sensor. Multimodal data streams were synchronized using the lab streaming layer (LSL) application. A modified version of an existing multiple linear regression models was employed to learn the relationship between the electroencephalography (EEG) feature input and the recorded kinematic data. To assess system performance with limited data, 10-fold cross validation was used.

Results

The most information-rich frequency bands for subjects were found to be in the ranges of 5 – 9 Hz and 33 – 37 Hz. Hand lateralization accuracy for the three subjects were 97.4, 78.7 and 96.9% respectively. 3D position reconstructed with an average correlation coefficient of 0.21, 0.47 and 0.38 respectively along three pre-defined axes, with the corresponding average correlation coefficients for velocity being 0.21, 0.36 and 0.25 respectively. The results compare favourably with a cross-section of existing results, while cost-per-electrode costs were 76% lower than the average per-electrode cost for similar systems and 44% lower than the cheapest previously-reported system.

Conclusions

This study has shown that low-cost bio-amplifiers such as the OpenBCI can be used for 3D motion reconstruction tasks.

Introduction

Background

A brain-computer interface (BCI) is a combination of hardware and software that allows direct communication between a human brain and some other system external to the body, bypassing at least a portion of the brain’s normal output pathways of peripheral nerves and muscles [1], [2].

An ideal BCI is accurate, cheap, easy to use, and safe [2], [3]. The balance of these four factors achievable by any particular BCI is largely determined by the neuroimaging technique used for neurosignal acquisition. A large number of neuroimaging modalities have been used for implementing BCIs. These include magnetoencephalography (MEG), functional magnetic resonance imaging (fMRI), positron emission tomography (PET), functional near-infrared spectroscopy (fNIRS), and electroencephalography (EEG). There are three variants of EEG differentiated primarily by the location of the electrodes, which can be on the scalp (scalp EEG, SEEG), the surface of the cortex (electrocorticography, ECoG), or within the cortex (intracortical EEG, ICE). MEG, PET and fMRI are currently not suitable for widespread use as BCIs due to their cost, limited real-time capabilities, complex technical requirement [4] and bulk. Although fNIRS is less affected by the above problems, it is still not as commonly used for BCI as systems based EEG, which collectively have the best balance of features. EEG-based systems are generally compact, capable of high resolution, and relatively cheap. SEEG is very safe, although ECoG and ICE are less so. ICE acquisition systems are capable of the highest spatial resolution of 0.05 mm and temporal resolution of 0.003 s [5], but the required surgical implantation of electrodes is so risky that it is only considered when there are no suitable non-invasive alternatives [6].

Although SEEG is very safe but does not offer the resolution or signal-to-noise ratios of ECoG or ICE. There is therefore a lot of interest in improving the performance of SEEG-based BCI, while retaining its advantages. This has become even more important due to the rise in the use of continuous decoding systems, which carry out continuous reconstruction or estimation of neural parameters rather than the classical approach of treating the BCI problem as a discrete multi-class classification task.

Classical discrete decoders are specifically tuned to decipher state changes (for example, left against right hand movement/imagery) [7], while the underlying continuous neurodynamic parameters involved in the changes are not monitored. As such, the technique leads to inability to represent multidimensional output effectively, by not having enough information for reconstruction. Researchers previously tackled this problem by linking different and unrelated input sources (tongue, hand and leg) to generate discrete commands necessary to control end-effectors. These forms of control entailed long training periods: majorly because it deviates from the central nervous system control strategies the subject is previously used to [8]. This procedure saps the subjects’ energy and might bring discouragement, leaving not much to be desired in the long run.

In continuous-decoding paradigms, subjects control the BCI system by carrying out normal limb movements similar to those used in activities of daily living. This thus would make BCI’s more natural to use, and more likely to deliver on its long-term promise. In order to make BCI adoption faster though, the issue of cost has to be addressed. There is thus an interest in finding out if low cost scalp-EEG systems can be used as a basis for a continuous decoding BCI. Consequently, the aim of this study was to explore the suitability of a low-cost scalp EEG system for designing continuous decoding BCI.

Related works

With respect to BCI, the reconstruction of human arm movement involves the continuous extraction of features and the decoding of kinematic variables (velocity, acceleration, direction and position) defining the arm movement from the brain rhythms elicited during the movement [9], [10], [11].

One of the earliest attempts to decode hand movement direction during a centre-out experiment was reported in [12]. Using 20-electrode EEG with 2.5 cm spacing above the contralateral motor along with MEG, nine right-handed subjects were asked to move a joystick in one of four directions in a self-paced manner. The study demonstrated that the direction of small-hand movements could be inferred with a reliability of 67%. The study also found that frequencies less than 7 Hz were found to contain features which correlated highly with movement, which was similar to previous results using more invasive methods, although in addition, the study of [12] also found discriminative features in the 10 – 30 Hz (beta) and 62 – 87 Hz (high gamma) frequency ranges.

Low-frequency discriminants were also observed by [13] in the frequency range of 2 – 5 Hz in an experiment using MEG. This study also confirmed the link between continuously monitored hand movement and motor cortex data on the side contralateral to the hand being used. In a follow-up [10], the authors noted that decoding accuracy and information rate did not differ significantly between EEG and MEG-based BCIs.

During a reaching motion of hands of human subjects toward randomly displayed reach targets, [9] decoded both reach intentions and actual reach movements and the successful classifications of several classes of motions involved in reach. The group classified left vs. right reach, top vs. down reach, the three and four reach targets in both experiments and the starting pose collapsed across target end points. The experiment utilized meta-classifiers: a combination of classifiers using probability to select the most probable classification accuracy. A Northern Digital Optrack Certus 3-D camera-based tracking system was used to record subject movements. It was concluded that properly-cleaned EEG is very capable of being used for classifying the dynamics involved in reach motion. But it was opined that enough of the movement related artifacts especially due to the EOG and EMG must be removed before further processing, independent component analysis (ICA) and dipole fitting methods for that purpose.

Similar results were obtained by [11] in an experiment to reconstruct hand velocity during a 3D centre-out movement profile. A self-initiated movement paradigm was adopted, in which subjects were allowed to choose target and the time to reach towards the target of choice. The velocity of hand movement was reconstructed using a 34-electrode array and a discriminant frequency of 1 Hz, with the components of the velocity along three perpendicular axes being reproduced with Pearson’s correlation coefficients of up to 0.19, 0.38 and 0.32 respectively.

Event related synchronization and desynchronization information extracted from EEG encode hand clenching speed along with the traditional hand lateralization information, as revealed in the study of [14]. The study developed a linear equation with independent parameters for movement-related information. Similarly, the experiment reported in [15] successfully extracted hand-clenching information from low-frequency EEG signals, but utilized a real-time Kalman filter and a smoother.

In [16], two movement speeds and along four directions in a horizontal plane were adopted along with a centre-out paradigm. Two orthogonal components of hand motion were measured using a robotic device strapped to the hand of the subject. These components were then used as predicted variables in an adaptive Kalman filter for which EEG features were predictor variables. Model performance was estimated using the correlation between the predicted and observed movement parameters.

Most of the existing studies in the literature use either joystick or some other device requiring physical contact, or sophisticated camera-based approaches to recording subject movement. Since one of the promises of continuous decoding systems is the freedom for BCI users to make more natural movements, movement monitoring systems based on physical contact have to be seen as sub-optimal. On the other hand, camera-based systems tend to be expensive. The Microsoft Kinect however offers a relatively cheap, viable option. For example, the usefulness of the Microsoft Kinect sensor for arm movement speed was evaluated in [17] and found it to be more suitable than any previous approach that utilized accelerometer among other technologies.

Materials and methods

Instrumentation/equipment

One goal of this study was to keep costs as low as possible. For that reason, free and open source technologies were adopted with only one major exception: a Microsoft Kinect sensor. The OpenBCI 32-bit 16-channel bio-amplifier was used for scalp EEG data capture. The OpenBCI features an open design based on the Texas Instruments ADS1299 analogue frontend with a 24-bit Delta-Sigma analogue-to-digital converter. Its 8-channel Ganglion board was combined with an 8-channel Cyton biosensing board and custom headset fabricated out of ABS plastic using an Ultimaker II 3D printer. Dry Ag/AgCl electrodes with diameter of 7 mm and spike dimension of 100 μm were used as the interface to the subjects’ scalps, with the effective scalp-electrode resistance measured to be below 50 kΩ at any instance. A Microsoft Kinect sensor was used for contactless measurement of body part coordinates, with video data acquired at a rate of 30 frames per second. The system was capable of acquiring data at the rate of 125 Hz with a serial baud rate of 115200 over the Bluetooth connection.

Release 3.0.2 of OpenSesame, an open-source, graphical experiment builder was used for designing the interface for stimulus presentation to subjects. For the data acquisition and signal analysis, Python script from OpenBCI repository and custom developed scripts were utilized respectively. The custom Python script was written to exploit subjects’ specific frequency bands using filter bank common spatial pattern (FBCSP) algorithm. The lab streaming layer (LSL) multimodal datastream synchronization package was used to coordinate and synchronize data from the OpenBCI, the OpenSesame interface, and the Kinect sensor. All data processing algorithms were run on a HP z640 workstation with 32 GB of RAM, and an NVidia Titax Xp GPU.

Scenario description

The experiment required subjects to move their arms up from their torso in response to cues presented by OpenSesame. The presented cue informed the subject to either move the left or the right arm at random speeds. After the disappearance of the cue, the subjects were required to bring the arms back to the initial position parallel to the torso.

The subjects sat on a comfortable, armless and non-reclining chair. The height of the chair was adjustable, so that subject’s feet could lie flat against the floor. They were required to put on a flip-flop for convenience (to reduce the stress on the feet) and safety (to insulate them from ground).

The experiment was carried out in a sleep lab and the subjects were required to have rested for at least a period of 4 h before the commencement of the experiment. This was done to relax the subjects in order to eliminate stress-evoked noise that might confound the analysis later on.

Subjects and EEG data acquisition

Three healthy, right-handed male subjects with ages between 20 and 30 took part in the experiment. They all had normal (or corrected to normal) vision and were selected at random. None of them had any previous experience working with BCI. The approval of the ethics committee of Obafemi Awolowo University, Ile-Ife was sought and obtained in writing before commencement of experimentation, and all subjects were properly informed, and gave their written informed consent to the study.

The experimental EEG dataset were acquired at the rate of 125 Hz, band-pass filtered from 0.5 to 40 Hz and subsequently subsampled to 31.25 Hz using a 4th order Butterworth filter. No notch filter was applied to remove contamination from the mains due to the pass-band used for the band-pass filter. The Kinect data were acquired at 30 frames per second and band-pass-filtered from 0.5 to 3 Hz.

A custom 3D designed headset served as an interface between the scalp of the subjects and the OpenBCI amplifier as shown in Figure 1. The headset was designed to host a spring-loaded assembly of dry electrode placed in conformance to 10–20 international standard of electrode placement. Eleven electrodes were located at the following scalp sites: F3, Fz, F4, C3, Cz, C4, P3, Pz, P4, O1 and O2 with ear linked reference. Alongside, the blinking eye movements were recorded even though subjects were instructed to delay blinking during the cue presentations within a trial. The recording was done by implementing a basic electrooculography (EOG) system attaching wet disposable EEG electrodes placed directly above and below each eye. Trials having excessive blinks were later excluded from further analysis.

Figure 1:
Figure 1:

Custom designed headset worn by a subject.

Citation: Biomedical Engineering / Biomedizinische Technik 2020; 10.1515/bmt-2019-0085

The experiment was carried out in two sessions for each subject. Each session contained 10 runs and each run had 10 trials. As shown in Figure 2, each trial took 9 s to complete. The experiment observed an inter-run rest of 1 min. The sessions were spread over a week. Each trial started with an arrow informing the subject to use either the left or the right hand for the experiment, and then a beep followed to get the subjects ready for the coming cue. The cue was a fixation cross which told the subject to move the cued arm up and keep there for the duration of the cue. As soon as the cue disappeared, the subject was to move the arm back to the starting point parallel to the torso.

Figure 2:
Figure 2:

Experimental timeline for a trial.

Citation: Biomedical Engineering / Biomedizinische Technik 2020; 10.1515/bmt-2019-0085

Data processing

The signal processing was done with a Python programming language version 3.5 script using the processing pipeline presented in Figure 3. The EEG data preprocessing involved the removal of the three frontal electrodes from further analysis because they were discovered to be contaminated by eye movement artifacts. Then a manual inspection was employed to determine the extent of contamination of the other channels by noise and artifacts. Any trial affected was removed from further processing. The datasets were segmented into epochs using the appearance of the cue as the reference. Each epoch included 2s baselines pre and post trials to aid such analysis as the time-frequency decomposition and baseline corrections later on.

Figure 3:
Figure 3:

Data processing pipeline.

Citation: Biomedical Engineering / Biomedizinische Technik 2020; 10.1515/bmt-2019-0085

An implementation of the FBCSP algorithm (see Figure 4) was used to determine the subject’s specific frequency [18], [19], [20], [21], [22] from 4 to 40 Hz with the bandwidth of each bin being 4 Hz. In addition, the optimal temporal features were determined by running several instances of the FBCSP algorithm on time samples taken by window length and offset.

Figure 4:
Figure 4:

FBCSP implementation for feature selection.

Citation: Biomedical Engineering / Biomedizinische Technik 2020; 10.1515/bmt-2019-0085

After determining the optimal time and subjects’ specific frequency bands the potential time series (PTS) was converted to band power time series (BTS) [16] for further analysis using the continuous wavelet reconstruction of the decomposed PTS signal. Therefore, the features were extracted from the BTS as a combination of spectral-temporal signatures selected at the output of the FBCSP algorithm by the mutual information-based best individual feature (MIBIF) algorithm [20]. The features were ranked, with only the two highest features selected and combined into feature matrix across the trial epochs.

A modified version of the multiple linear regression model (mLR) presented by [17] was employed to learn the relationship between the EEG feature input and the registered kinematic data that served as the output.

xi[t]=ai+b=1Bn=1Nk=1LbntkiSnt[tk]+ε[t]
where xi[t] = {vx, vy,vz, v, dx,dy,dz,d} R8TR, v and d are the Euclidean norms of the velocity and position parameters obtained from the Kinect, T is the number of samples in the trials, and R is the number of trials. In addition, ai and bnki are the regression parameters that learnt the relationship between the independent variable Sn[t-k] and the dependent variable xi. N is the number of EEG channels, B is the number of sub-bands, L is the number of time lags, index i denotes the three coordinates registered by the Kinect and ε[t] is the residual error. The x axis was defined as the “left-right” axis defined by the intersection of a single coronal plane and a transverse plane through the subject, movement along the y-axis was defined as a purely “up-down” vertical movement on any sagittal plane, while the z axis was defined as the “in-out” axis defined by the horizontal component of movement along the mid-sagittal plane.

Sn [t-k] is the standardized temporal difference of EEG potentials at the EEG sensor n and at time lag k given by:

S[t]=vBTS[t]σvBTS
where vBTS [t] is the value of band-power time series input at time t and σBTS is the standard deviation of vBTS.

To solve the multidimensional problem in equation (1), a partial least square (PLS) solver was employed for this particular case of multidimensional predicted and predictor space. PLS is well suited to situations where there are large dimensions of observation space but limited data available for such predictions. Just like PCA, the PLS gives a small set of principal directions called the latent vectors which on the contrary takes care of the observed variations in both predictor and predicted spaces [27].

Due to noise and the lack of computational platforms capable of infinite precision arithmetic, embedding parameters were estimated, with false nearest neighbour (FNN) used to estimate the time lag, τ , while the first minimum average mutual information algorithm was used for estimating the embedding dimension, L + 1. By plotting the average mutual information against time lag, a first minimum of one was obtained, which translated to 8 ms of time lag. The optimal lag time per subject was subsequently determined.

A 10-fold cross-validation scheme was employed for estimating the performances of the algorithms used. In this scheme, the dataset was partitioned into 10 folds, 9 of the 10 were used for training and the remaining one for testing. The process was repeated 10 times and the results obtained were averaged to obtain the performance metric. The scheme was applied to both the classification and the regression analysis. For the classification, the metric was the accuracy while the metric for the regression was the correlation coefficient between the measured and reconstructed variables.

The performance of the device with respect to the discriminative power of the classifier and the correlation coefficient of decoded kinematic parameters was evaluated and reported, with receiver operating curve (ROC) of the classifiers per subject, and the correlation coefficients indicating the strength of agreement between the measured and decoded kinematic parameters. In addition, the effect of the experimental setup and the continuous decoding of the kinematic parameter on the speed of the system were examined, their effects on the training time required for the subjects were speculated.

Results and discussion

The optimal time lag and the embedding dimensions were determined to be 64 ms and 15 respectively. The other optimal parameter selections are presented in Table 1, including the window sizes (the chunk of a trial’s time series taken for the analysis) and the subjects’ specific frequency. Classification reports for all the subjects are presented in Table 2. For each subject, the precision, recall, f1-score and support for left and right hand are presented, along with an average of the left and right hand scores. The combined confusion matrix is presented in Table 3. Clearly, the OpenBCI-based system was able to discriminate quite well between left- and right-hand movement using EEG data with an overall accuracy of 89.7%. Even with the variation across subjects, the worst-performing subject (Subject B) still had an accuracy of 78.7%. Subjects A and C had accuracies of 97.4 and 96.9% respectively.

Table 1:

Optimal classification and regression parameters for each subject.

ParameterSubject ASubject BSubject C
Window size1.5s2.5s1s
Specific frequency5–9, 33–375–9, 33–375–9, 33–37
Embedding dimension151515
Time lag64ms64ms64ms
Number of Components used for PLS Regression solver505050
Table 2:

Classification reports for hand laterization.

SubjectHandPrecisionRecallf1-scoresupport
ALeft hand1.00.950.9720
Right hand0.951.00.9718
Average/Total0.980.970.9738
BLeft hand0.810.740.7723
Right hand0.770.830.8024
Average/Total0.790.790.7947
CLeft hand1.00.940.9716
Right hand0.941.00.9716
Average/Total0.970.970.9732
Table 3:

Combined confusion matrix for hand lateralization.

SubjectPredicted Right HandPredicted Left Hand
AActual Right Hand191
Actual Left Hand018
BActual Right Hand176
Actual Left Hand420
CActual Right Hand151
Actual Left Hand016

A summary of the performance of the system on hand movement reconstruction for all subjects is presented in Table 4. The correlations are generally weakly positive, with y-axis (up-down) and z-axis (in-out) movements consistently marginally better than x-axis (left-right) movement for reasons which could not be determined. Similar to the classification/hand lateralization tasks, Subject B had the lowest performance on movement reconstruction. This suggests some unique underlying user-specific physiological or anatomical variables at play, and is in line with the well-known observation that BCI-naive subjects perform differently on BCI tasks, and require different training times to reach similar performance levels [23]. While the EEG of all subjects contained discriminatory information in the 5 – 9 Hz band, the most velocity information-rich supplementary band differed for Subject A (17 – 21 Hz) from the 33 – 37 Hz range of Subject B and Subject C.

Table 4:

Summary of velocity and positional parameter reconstruction performance.

SubjectParameterCoefficientFrequency Range
AVx0.236 (0.03)5–9, 17–21
Vy0.511 (0.053)5–9, 17–21
Vz0.35 (0.021)5–9, 17–21
Px0.266 (0.028)5–9, 33–37
Py0.56 (0.057)5–9, 33–37
Pz0.361 (0.018)5–9, 33–37
BVx0.154 (0.015)5–9, 33–37
Vy0.228 (0.018)5–9, 33–37
Vz0.196 (0.031)5–9, 33–37
Px0.153 (0.015)5–9, 33–37
Py0.347 (0.041)5–9, 33–37
Pz0.249 (0.030)5–9, 33–37
CVx0.250(0.036)5–9, 33–37
Vy0.331(0.048)5–9, 33–37
Vz0.200(0.027)5–9, 33–37
Px0.218(0.029)5–9, 33–37
Py0.510(0.083)5–9, 33–37
Pz0.523(0.079)5–9, 33–37

The slightly lower coefficients of velocity parameters is very likely due to the fact that velocity data were computed as derivatives of position time series, leading to the introduction of computational noise. For comparison, the correlation coefficients of similar studies are represented in Table 5, along with information about the electrode densities and cost of bio-amplifiers. The average per-electrode cost of previous studies was $243.41, while the cheapest was $105.63, vs. the per-electrode cost of $59.40 achieved for the current study. Consequently, the current study was able to perform generally favourably for arm reconstruction for BCI when compared with previous studies, even though the total bio-amplifier cost was much lower, the per-electrode cost was 76% lower than the average per-electrode cost for previous studies, and was 44% cheaper than the previous cheapest per-electrode cost achieved in the literature [26]. Table 5 shows a mild correlation between the number of electrodes and the reconstruction accuracy, with the 64-, 49- and 35-electrode systems achieving the best overall performance. Notwithstanding, the performance of the 12 electrode system used for this study is as good as previous studies using 21, 34 and 40 electrodes. This suggests that the effect of number of electrodes on reconstruction accuracy is mitigated by other factors such as the amplifier and the processing pipeline.

Table 5:

Comparison of selected motion reconstruction results.

AuthorMethodElectrode DensityEEG DeviceCost/channel ($)Reconstruction accuracy (averages of correlation coefficients for position and velocity)
This studyMultilinear regression model using Partial Least Square (PLS) solver12-electrodesOpenBCI 16 channel amplifier59.40x: 0.22

y: 0.42

z: 0.32
Lv et al. [15]Spatial filtering, Kalman filtering and smoothing algorithm40-channel EEG cap LT37 from CompumeticsNeuroScan NuAmps299.87x: 0.37

y: 0.24
Bradberry et al. [11]Linear decoding model, 8 × 8 cross-validation, Source estimation with sLORETA.34 sensorsNot availableNot availablex: 0.19

y: 0.38

z: 0.32
Kim et al. [24]Multiple linear regression21 electrodesActiveTwo, BioSemi218.23x: 0.31

y: 0.27

z: 0.15
Kim et al. [25]Linear and non-linear (kernerl ridge regression) decoding models with 8-fold cross-validation64 electrodesBrainAmp system421.88x, y and z greater than 0.6
Ofner and Muller-Putz [26]Multilinear regression model49 electrodesg.USBamp amplifiers (g.tec, Graz, Austria).105.63x: 0.70

y: 0.77

z: 0.62
Robinson et al. [16]Multiple linear regression and Kalman filter models35 electrodesNeuroscan SynAmps EEG amplifier171.43x and y gave a mean of 0.57

This has important implications for future applications of BCI technology. Clearly, one of the biggest application areas for BCIs based on 3D hand movement reconstruction is in the development of neuroprostheses which can be controlled with more natural motion (or motor imagery). Cost is a major barrier to achieving that goal. This study currently has the lowest per-electrode cost of any study that has demonstrated any ability to reconstruct hand motion from EEG signals. More work needs to be done to improve performance, but the results definitely provide a platform that can be built upon towards lower cost neuroprostheses controlled with natural motion or motor imagery of natural movements.

Conclusion

In this study, the effectiveness of a low-cost bio-amplifier in reconstructing 3-D components of hand motion from EEG signals was evaluated using 11 electrode locations from the scalps of three subjects. The system was assessed in terms of classification of hand lateralization and continuous reconstruction of position and velocity. The performance was comparable to previous studies using up to double the number of electrodes, while using a total setup that was by far the cheapest used till date. The primary implication is that neuroprostheses using electroencephalography and subject control based on natural hand gestures can be made much cheaper.

Acknowledgments

The workstation used for this research was donated by the Sanofi Family Foundation, Wollongong. The Titan Xp GPU was donated by the NVIDIA Corporation. The authors thank the Akinlolu Foundation, Lagos for its support.

Research funding: Authors state no funding involved.

Conflict of interest: Authors state no conflict of interest.

References

  • 1.

    Allison BZ, Wolpaw EW, Wolpaw JR. Brain–computer interface systems: progress and prospects. Expert Rev Med Devices 2007;4:463–74. https://doi.org/10.1586/17434440.4.4.463.

    • Crossref
    • PubMed
    • Export Citation
  • 2.

    Wolpaw JR, Birbaumer N, Heetderks WJ, McFarland DJ, Peckham PH, Schalk G, et al. Brain-computer interface technology: a review of the first international meeting. IEEE Trans Rehabil Eng 2000;8:164–73. https://doi.org/10.1109/TRE.2000.847807.

    • Crossref
    • PubMed
    • Export Citation
  • 3.

    Van Dokkum LE, Ward T, Laffont I. Brain computer interfaces for neurorehabilitation–its current status as a rehabilitation strategy post-stroke. Ann Phys Rehabil Med 2015;58:3–8. https://doi.org/10.1016/j.rehab.2014.09.016.

    • Crossref
    • PubMed
    • Export Citation
  • 4.

    Daly JJ, Wolpaw JR. Brain–computer interfaces in neurological rehabilitation. Lancet Neurol 2008;7:1032–43. https://doi.org/10.1016/S1474-4422(08)70223-0.

    • Crossref
    • PubMed
    • Export Citation
  • 5.

    Purkayastha SS, Jain VK, Sardana HK. Topical review: a review of various techniques used for measuring brain activity in brain computer interfaces. Adv Electron Electr Eng 2014;4:513–22.

  • 6.

    Hildt E. Brain-computer interaction and medical access to the brain: individual, social and ethical implications. Stud Ethics Law Technol 2010;4. https://doi.org/10.2202/1941-6008.1143.

  • 7.

    Sleight J, Pillai P, Mohan S. Classification of executed and imagined motor movement EEG signals. Ann Arbor, MI: University of Michigan; 2009.

  • 8.

    Liao K, Xiao R, Gonzalez J, Ding L. Decoding individual finger movements from one hand using human EEG signals. PLoS One 2014;9:e85192. https://doi.org/10.1371/journal.pone.0085192.

    • Crossref
    • PubMed
    • Export Citation
  • 9.

    Hammon PS, Makeig S, Poizner H, Todorov E, De Sa VR. Predicting reaching targets from human EEG. IEEE Signal Process Mag 2008;25:69–77. https://doi.org/10.1109/MSP.2008.4408443.

    • Crossref
    • Export Citation
  • 10.

    Jerbi K, Vidal JR, Mattout J, Maby E, Lecaignard F, Ossandon T, et al. Inferring hand movement kinematics from MEG, EEG and intracranial EEG: from brain-machine interfaces to motor rehabilitation. IRBM 2011;32:8–18. https://doi.org/10.1016/j.irbm.2010.12.004.

    • Crossref
    • Export Citation
  • 11.

    Bradberry TJ, Gentili RJ, Contreras-Vidal JL. Reconstructing three-dimensional hand movements from noninvasive electroencephalographic signals. J Neurosci 2010;30:3432–7. https://doi.org/10.1523/JNEUROSCI.6107-09.2010.

    • Crossref
    • PubMed
    • Export Citation
  • 12.

    Waldert S, Preissl H, Demandt E, Braun C, Birbaumer N, Aertsen A, et al. Hand movement direction decoded from MEG and EEG. J Neurosci 2008;28:1000–8. https://doi.org/10.1523/JNEUROSCI.5171-07.2008.

    • Crossref
    • PubMed
    • Export Citation
  • 13.

    Jerbi K, Bertrand O, Schoendorff B, Hoffmann D, Minotti L, Kahane P, et al. Online detection of gamma oscillations in ongoing intracerebral recordings: from functional mapping to brain computer interfaces. In: Joint meeting of the 6th international symposium on non-invasive functional source imaging of the brain and heart and the international conference on functional biomedical imaging. IEEE, Hangzhou, China; 2007. 330–3 pp.

  • 14.

    Yuan H, Perdoni C, He B. Decoding speed of imagined hand movement from EEG. In: annual international conference of the IEEE engineering in medicine and biology. IEEE, Buenos Aires, Argentina; 2010. 142–5 pp.

  • 15.

    Lv J, Li Y, Gu Z. Decoding hand movement velocity from electroencephalogram signals during a drawing task. Biomed Eng Online 2010;9:64. https://doi.org/10.1186/1475-925X-9-64.

    • Crossref
    • PubMed
    • Export Citation
  • 16.

    Robinson N, Guan C, Vinod AP. Adaptive estimation of hand movement trajectory in an EEG based brain–computer interface system. J Neural Eng 2015;12:066019. https://doi.org/10.1088/1741-2560/12/6/066019.

    • Crossref
    • Export Citation
  • 17.

    Elgendi M, Picon F, Magnenat-Thalmann N, Abbott D. Arm movement speed assessment via a Kinect camera: a preliminary study in healthy subjects. Biomed Eng Online 2014;13:88. https://doi.org/10.1186/1475-925X-13-88.

    • Crossref
    • Export Citation
  • 18.

    Chin ZY, Ang KK, Wang C, Guan C, Zhang H. Multi-class filter bank common spatial pattern for four-class motor imagery BCI. In: Annual international conference of the IEEE engineering in medicine and biology society. IEEE, Minneapolis, MN, USA; 2009. 571–4 pp.

  • 19.

    Park GH, Lee YR, Kim HN. Improved filter selection method for filter bank common spatial pattern for EEG-based BCI systems. Int J Electron Electr Eng 2014;2:101–5. https://doi.org/10.12720/ijeee.2.2.101-105.

  • 20.

    Ang KK, Chin ZY, Zhang H, Guan C. Filter Bank Common Spatial Pattern (FBCSP) algorithm using online adaptive and semi-supervised learning. In: International joint conference on neural networks. IEEE, San Jose, CA, USA; 2011. 392–6 pp.

  • 21.

    Bentlemsan M, Zemouri ET, Bouchaffra D, Yahya-Zoubir B, Ferroudji K. Random forest and filter bank common spatial patterns for EEG-based motor imagery classification. In: 5th International conference on intelligent systems, modelling and simulation. IEEE, Langkawi, Malaysia; 2014. 235–8 pp.

  • 22.

    Blankertz B, Lemm S, Treder M, Haufe S, Müller KR. Single-trial analysis and classification of ERP components—a tutorial. NeuroImage 2011;56:814–25. https://doi.org/10.1016/j.neuroimage.2010.06.048.

    • Crossref
    • PubMed
    • Export Citation
  • 23.

    Blankertz B, Losch F, Krauledat M, Dornhege G, Curio G, Müller KR. The Berlin brain-computer interface: accurate performance from first-session in BCI-naive subjects. IEEE Trans Biomed Eng 2008;55:2452–62. https://doi.org/10.1109/TBME.2008.923152.

    • Crossref
    • PubMed
    • Export Citation
  • 24.

    Kim JH, Chavarriaga R, Millán JDR, Lee SW. 3D trajectory reconstruction of upper limb based on EEG. In: Proceedings of the Fifth International Brain-Computer Interface Meeting 2013 (No. CONF). Graz University of Technology Publishing House; 2013. https://doi.org/10.3217/978-3-85125-260-6-137.

  • 25.

    Kim M, Kim BH, Jo S. Quantitative evaluation of a low-cost noninvasive hybrid interface based on EEG and eye movement. IEEE Trans Neural Syst Rehabil Eng 2015;23:159–68. https://doi.org/10.1109/TNSRE.2014.2365834.

    • Crossref
    • PubMed
    • Export Citation
  • 26.

    Ofner P, Müller-Putz GR. Decoding of velocities and positions of 3D arm movement from EEG. In: Annual international conference of the IEEE engineering in medicine and biology society. IEEE, San Diego, CA, USA; 2012. 6406–9 pp.

  • 27.

    Abdi H. Partial least squares regression and projection on latent structure regression (PLS Regression). Wiley Interdiscip Rev Comput Stat 2010;2:97–106. https://doi.org/10.1002/wics.51.

    • Crossref
    • Export Citation

If the inline PDF is not rendering correctly, you can download the PDF file here.

  • 1.

    Allison BZ, Wolpaw EW, Wolpaw JR. Brain–computer interface systems: progress and prospects. Expert Rev Med Devices 2007;4:463–74. https://doi.org/10.1586/17434440.4.4.463.

    • Crossref
    • PubMed
    • Export Citation
  • 2.

    Wolpaw JR, Birbaumer N, Heetderks WJ, McFarland DJ, Peckham PH, Schalk G, et al. Brain-computer interface technology: a review of the first international meeting. IEEE Trans Rehabil Eng 2000;8:164–73. https://doi.org/10.1109/TRE.2000.847807.

    • Crossref
    • PubMed
    • Export Citation
  • 3.

    Van Dokkum LE, Ward T, Laffont I. Brain computer interfaces for neurorehabilitation–its current status as a rehabilitation strategy post-stroke. Ann Phys Rehabil Med 2015;58:3–8. https://doi.org/10.1016/j.rehab.2014.09.016.

    • Crossref
    • PubMed
    • Export Citation
  • 4.

    Daly JJ, Wolpaw JR. Brain–computer interfaces in neurological rehabilitation. Lancet Neurol 2008;7:1032–43. https://doi.org/10.1016/S1474-4422(08)70223-0.

    • Crossref
    • PubMed
    • Export Citation
  • 5.

    Purkayastha SS, Jain VK, Sardana HK. Topical review: a review of various techniques used for measuring brain activity in brain computer interfaces. Adv Electron Electr Eng 2014;4:513–22.

  • 6.

    Hildt E. Brain-computer interaction and medical access to the brain: individual, social and ethical implications. Stud Ethics Law Technol 2010;4. https://doi.org/10.2202/1941-6008.1143.

  • 7.

    Sleight J, Pillai P, Mohan S. Classification of executed and imagined motor movement EEG signals. Ann Arbor, MI: University of Michigan; 2009.

  • 8.

    Liao K, Xiao R, Gonzalez J, Ding L. Decoding individual finger movements from one hand using human EEG signals. PLoS One 2014;9:e85192. https://doi.org/10.1371/journal.pone.0085192.

    • Crossref
    • PubMed
    • Export Citation
  • 9.

    Hammon PS, Makeig S, Poizner H, Todorov E, De Sa VR. Predicting reaching targets from human EEG. IEEE Signal Process Mag 2008;25:69–77. https://doi.org/10.1109/MSP.2008.4408443.

    • Crossref
    • Export Citation
  • 10.

    Jerbi K, Vidal JR, Mattout J, Maby E, Lecaignard F, Ossandon T, et al. Inferring hand movement kinematics from MEG, EEG and intracranial EEG: from brain-machine interfaces to motor rehabilitation. IRBM 2011;32:8–18. https://doi.org/10.1016/j.irbm.2010.12.004.

    • Crossref
    • Export Citation
  • 11.

    Bradberry TJ, Gentili RJ, Contreras-Vidal JL. Reconstructing three-dimensional hand movements from noninvasive electroencephalographic signals. J Neurosci 2010;30:3432–7. https://doi.org/10.1523/JNEUROSCI.6107-09.2010.

    • Crossref
    • PubMed
    • Export Citation
  • 12.

    Waldert S, Preissl H, Demandt E, Braun C, Birbaumer N, Aertsen A, et al. Hand movement direction decoded from MEG and EEG. J Neurosci 2008;28:1000–8. https://doi.org/10.1523/JNEUROSCI.5171-07.2008.

    • Crossref
    • PubMed
    • Export Citation
  • 13.

    Jerbi K, Bertrand O, Schoendorff B, Hoffmann D, Minotti L, Kahane P, et al. Online detection of gamma oscillations in ongoing intracerebral recordings: from functional mapping to brain computer interfaces. In: Joint meeting of the 6th international symposium on non-invasive functional source imaging of the brain and heart and the international conference on functional biomedical imaging. IEEE, Hangzhou, China; 2007. 330–3 pp.

  • 14.

    Yuan H, Perdoni C, He B. Decoding speed of imagined hand movement from EEG. In: annual international conference of the IEEE engineering in medicine and biology. IEEE, Buenos Aires, Argentina; 2010. 142–5 pp.

  • 15.

    Lv J, Li Y, Gu Z. Decoding hand movement velocity from electroencephalogram signals during a drawing task. Biomed Eng Online 2010;9:64. https://doi.org/10.1186/1475-925X-9-64.

    • Crossref
    • PubMed
    • Export Citation
  • 16.

    Robinson N, Guan C, Vinod AP. Adaptive estimation of hand movement trajectory in an EEG based brain–computer interface system. J Neural Eng 2015;12:066019. https://doi.org/10.1088/1741-2560/12/6/066019.

    • Crossref
    • Export Citation
  • 17.

    Elgendi M, Picon F, Magnenat-Thalmann N, Abbott D. Arm movement speed assessment via a Kinect camera: a preliminary study in healthy subjects. Biomed Eng Online 2014;13:88. https://doi.org/10.1186/1475-925X-13-88.

    • Crossref
    • Export Citation
  • 18.

    Chin ZY, Ang KK, Wang C, Guan C, Zhang H. Multi-class filter bank common spatial pattern for four-class motor imagery BCI. In: Annual international conference of the IEEE engineering in medicine and biology society. IEEE, Minneapolis, MN, USA; 2009. 571–4 pp.

  • 19.

    Park GH, Lee YR, Kim HN. Improved filter selection method for filter bank common spatial pattern for EEG-based BCI systems. Int J Electron Electr Eng 2014;2:101–5. https://doi.org/10.12720/ijeee.2.2.101-105.

  • 20.

    Ang KK, Chin ZY, Zhang H, Guan C. Filter Bank Common Spatial Pattern (FBCSP) algorithm using online adaptive and semi-supervised learning. In: International joint conference on neural networks. IEEE, San Jose, CA, USA; 2011. 392–6 pp.

  • 21.

    Bentlemsan M, Zemouri ET, Bouchaffra D, Yahya-Zoubir B, Ferroudji K. Random forest and filter bank common spatial patterns for EEG-based motor imagery classification. In: 5th International conference on intelligent systems, modelling and simulation. IEEE, Langkawi, Malaysia; 2014. 235–8 pp.

  • 22.

    Blankertz B, Lemm S, Treder M, Haufe S, Müller KR. Single-trial analysis and classification of ERP components—a tutorial. NeuroImage 2011;56:814–25. https://doi.org/10.1016/j.neuroimage.2010.06.048.

    • Crossref
    • PubMed
    • Export Citation
  • 23.

    Blankertz B, Losch F, Krauledat M, Dornhege G, Curio G, Müller KR. The Berlin brain-computer interface: accurate performance from first-session in BCI-naive subjects. IEEE Trans Biomed Eng 2008;55:2452–62. https://doi.org/10.1109/TBME.2008.923152.

    • Crossref
    • PubMed
    • Export Citation
  • 24.

    Kim JH, Chavarriaga R, Millán JDR, Lee SW. 3D trajectory reconstruction of upper limb based on EEG. In: Proceedings of the Fifth International Brain-Computer Interface Meeting 2013 (No. CONF). Graz University of Technology Publishing House; 2013. https://doi.org/10.3217/978-3-85125-260-6-137.

  • 25.

    Kim M, Kim BH, Jo S. Quantitative evaluation of a low-cost noninvasive hybrid interface based on EEG and eye movement. IEEE Trans Neural Syst Rehabil Eng 2015;23:159–68. https://doi.org/10.1109/TNSRE.2014.2365834.

    • Crossref
    • PubMed
    • Export Citation
  • 26.

    Ofner P, Müller-Putz GR. Decoding of velocities and positions of 3D arm movement from EEG. In: Annual international conference of the IEEE engineering in medicine and biology society. IEEE, San Diego, CA, USA; 2012. 6406–9 pp.

  • 27.

    Abdi H. Partial least squares regression and projection on latent structure regression (PLS Regression). Wiley Interdiscip Rev Comput Stat 2010;2:97–106. https://doi.org/10.1002/wics.51.

    • Crossref
    • Export Citation
FULL ACCESS

Journal + Issues

Biomedical Engineering / Biomedizinische Technik (BMT) is a high-quality forum for the exchange of knowledge in the fields of biomedical engineering, medical information technology and biotechnology/bioengineering. BMT addresses engineers, natural scientists, and clinicians working in research, industry, or clinical practice. All articles are peer-reviewed.

 

Search