^{®}EEG Headset and Data Analysis Based on Empirical Mode Decomposition

^{1}

^{1}

^{*}

^{2}

^{3}

^{1}

^{3}

This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/).

This paper presents a project on the development of a cursor control emulating the typical operations of a computer-mouse, using gyroscope and eye-blinking electromyographic signals which are obtained through a commercial 16-electrode wireless headset, recently released by Emotiv. The cursor position is controlled using information from a gyroscope included in the headset. The clicks are generated through the user's blinking with an adequate detection procedure based on the spectral-like technique called Empirical Mode Decomposition (EMD). EMD is proposed as a simple and quick computational tool, yet effective, aimed to artifact reduction from head movements as well as a method to detect blinking signals for mouse control. Kalman filter is used as state estimator for mouse position control and jitter removal. The detection rate obtained in average was 94.9%. Experimental setup and some obtained results are presented.

In the last decades there has been a growing effort from the research community to develop human-computer interfaces (HCI), aiming to provide convenient communication alternatives for disabled persons. Several approaches of user-friendly interfaces using voice, vision, gesture, and other modalities, can be found in recent literature [

In this work we present a project focusing on the development of a hands-free mouse emulation using the EEG headset recently released by Emotiv [

EMD is a technique used to decompose a time series into a finite number of functions called intrinsic mode functions (IMF) using an empirical identification based on its characteristic time scales [

In this work, a feature extraction procedure based on the spectral-like technique EMD is described. Results obtained using the proposed technique indicate an adequate process to hand the non-stationarity characteristic of EEG signals. Additional experiments using DWT were carried out for comparison purposes. Typical mouse-like function is a sequential process in which the user performs a movement to locate the cursor in the required position, and then selects an operation by applying a click action. Experiments were carried out considering extreme situations, where the subject is instructed to move the head at different speeds while applying a double click in indicated times. Section 2 describes some theoretical background on the used techniques. Section 3 presents a description of the experimental setup. Section 4 describes some obtained results, and section 5 presents some concluding remarks and future work about the described project.

EMD was first introduced by Huang [_{min(}_{t}_{)} and maxima _{max(}_{t}_{)}). Generate the envelope by connecting maxima and minima points with a curve, for instance, cubic spline interpolation, although other interpolation techniques are allowed. Determine the mean by averaging;

There are iteration stopping criteria such as establishing a certain number of siftings, thresholds, or minimum amplitude of residual. EMD satisfies completeness and orthogonality properties in the same way as spectral decompositions such as Fourier or wavelet transform. The completeness property is satisfied by EMD, in the sense that it is possible to reconstruct the original signal based on their decompositions. These decomposition functions should all be locally orthogonal to each other, as expressed in

An orthogonality index expressed in

Occasionally, the consideration of a local EMD is necessary. In this case, sifting operations are not applied to the full length signal. Sometimes, a better local approximation is obtaining through over-iteration of a specific zone; however, this process produces contamination in other signal zones and in consequence over-decomposing. Thus, the algorithm must keep iterating only over zones where the error remains large. Local EMD is implemented introducing a weighting function (

In EEG signal detection, it is important to get consistent records of electric brain activity from specific surface electrode location. For that purpose, scientists and physicians rely on a standard system for accurately placing electrodes, which is called the International 10–20 System, generally used in clinical EEG recording and EEG research.

The modules proposed to detect double-blinking event and to process gyroscope data are shown in the block diagram of

EEG signals provided by the EMOTIV headset EEG acquisition system are contaminated by noise produced by different sources such as: muscular movements (head movement, breathing,

As previously stated, noise produced by head or body movement will appear in all electrodes of the system with small variations, therefore, correlation analysis using Pearson coefficient is used for noise detection purposes. Pearson correlation coefficient provides a measure of dependence between two random variables. _{X}_{Y}_{X}_{Y}

Correlation function applied directly to the signals obtained from each electrode will state dependence between channels. Common signals detected would have to be removed; however, applying directly an operation to separate those signals could cause removing also important information. Therefore, decomposing the signal from each electrode will reduce the loss of information, allowing the system to distinguish between artifacts from head movements and double blinking signals. That decomposition has been carried out using EMD technique.

In order to find the amount of similarity or dependence, the Pearson correlation is calculated from corresponding IMF functions. Additionally, a

Once the noise is reduced, a second derivative is obtained in order to determine whether a critical point is a local maximum or a local minimum. A typical double blinking event will have two local max points inside a 0.5 s window.

The gyroscope IC embedded in the Emotiv headset provides information about head movements through a speed signal. An integration step is then required in order to obtain the cursor relative position.

In this stage, Kalman filter was used as a state estimator for mouse position control and jitter removal. Kalman filter is a recursive estimator that is used for computing a future estimate of the dynamic system state from a series of noisy measurements, minimizing the mean of the squared estimate error, between the prediction of the system's state and the measurement. Estimated state from the previous time step and the new measurements are used to compute a new estimate for the current dynamic system's state. Kalman filter has been used as estimator to perform smooth tracking and jitter remotion in several contexts such as image stabilization [_{s}

The Kalman filter is a recursive predictive filter that utilizes minimization of covariance error becoming it in an optimal estimator. The filter is based on the system definition using state space variables and recursive algorithms for the minimization process [

_{k}_{k}

Noise process variance was established from measurements using the data from gyroscope at rest, which is considered the process noise. Noise measurement variance was determined from data obtained during the experiment related to measurement or sensor noise.

Amplitude and time duration of double blinking vary among different person depending on physiological characteristics and intensity of the action. A test window of 2 s (256 samples) starts when a mark is sent to the recording system, which is then processed through the described EMD decomposition. A complete view of the system is depicted in

Subjects under testing were seated in a comfortable position using the Emotiv headset with a laser pointer attached at the top, as shown in

System test was performed using fold validation, dividing in a random way the data set in two groups of 10 vectors each. The system is tested using 5 complete sequence of 91 s, in which 10 double blinking events occur randomly. A double blinking event was experimentally found to fall inside a distance value of 0.95. The Mahalanobis distance is defined by

System performance is analyzed through a Receiver Operation Characteristic plot (ROC) [_{m} obtained from the classifier output. The curves indicate a common tendency of an increasing rate of true positives events with simultaneous increasing rate of false positives. A small increase of the FP rate compared to the variation of TP rate can be noticed. Changes on FPR and TPR are results of variations in threshold value.

According to

The Emotiv system could detect velocities in the range from 0 to 600 degrees per second with a 12 bit resolution ADC, with velocity increments up to 0.14 degrees per second. This variation is mapped to an integer range from 0 to 4,000. A simple approach for jitter removal reported in similar systems is the use of a dead band with a previously established threshold. It is evident that the use of a dead band implies losing precision in cursor location and speed. A direct evaluation based on simple experiments indicates that it would be necessary at minimum a dead-band of 10 steps to eliminate shakiness at best scenario and almost 30 steps for fast movements. In that scenario, the system would lose information on a range of variations from 1.4 degrees per second up to 4.2 degrees per second. Using the Kalman estimator, which takes in account the noise variance, the loss would be only in the order of 0.7 degrees per second.

We have presented an initial prototype of a mouse control which takes advantage of gyroscope and EEG signals obtained from the commercial Emotiv headset. Emulation of mouse-clicks using double blinking is detected using Empirical Mode Decomposition. Analysis on detection rate indicated that EMD provided an efficient, effective and quick computational tool, adequate to non-stationary signals. Despite movements of the testing subject during double clink event, the performance system shows excellent results. About the task of transforming gyroscope data into mouse device movements, Kalman filter as a state estimator for mouse position control and jitter removal offers a better approach to increase mouse pointer resolution in comparison to consider a threshold-based dead band to eliminate noise and shakiness. The proposed noise reduction method based on information available from multiple electrodes is a preprocessing technique which can be adapted to different EEG systems, when noise caused by head or body movement is required to be removed. Double blinking detection using a feature extraction technique based on Empirical Mode Decomposition provided a detection rate of 94.9% in average using a Mahalanobis-distance based classifier. Additional experiments exploring the incorporation of classifiers such as Support Vector Machine and Neural Networks are currently in progress.

The first author acknowledges the financial support from the Mexican National Council for Science and Technology (CONACYT), scholarship No. 347624. This research has been partially supported by CONACYT Grant No. CB-2010-155250. The authors would like to thank the anonymous reviewers for their helpful comments.

The author declares no conflict of interest.

(

International system 10–20.

Proposed scheme, blinking detection and gyroscope processing system.

Head movement noise during double blinking events.

Preprocessing to reduce head movement noise.

EMD decomposition from four different electrodes near AF3. (

Noise reduction based on correlation function removing, (

Double blinking detection with noise reduction.

Gyroscope data and velocity target movement from subject head movement; target movement (red line), head movement (blue line).

Gyroscope data and velocity target movement from four different subjects head movements.

Simplified flow diagram for Kalman filter.

Kalman filtering as state estimator in mouse control and jitter removal; target movement (black line), observed movement (blue line), and filtered movement (red line).

General scheme of detection system proposed.

Typical EMD decompositions (

Experimental setup of EEG-based mouse emulation.

Testing setup system.

Range of double blink detection for the classification module.

Average ROC curves obtained through measurements from AF4 (blue) and AF3 (red) electrodes.

Average ROC curves obtained for EMD decomposition from AF4 (blue) and for Wavelet decomposition (red) from the same electrode.

Comparison of several reported methods for blinking detection using electrodes as element sensor.

N. Kurian, |
None | Amplitude values | Thresholding | Not specified |

T. Wissel, |
Bessel filtering | Wavelet Transform | 1NN/LDA/Neural Networks | 90%–94% |

R. Barea, |
None | Wavelet Transform | Neural Networks | 92% |

B. Paulchamy, |
Not specified | Wavelet Transform | Adaptive Noise Cancellation | Based on SNR values |

L.F. Araghi, 2010 [ |
None | Wavelet Transform | ADALINE (adaptive linear neuron) | Not specified |

P. Kumar, |
None | Wavelet Transform | Thresholding by statistical parameters | Not specified |

P. SenthilKumar, |
None | Wavelet Transform | ADALINE (adaptive linear neuron) | Supression ratio: 3–71 dB |

W. Hsu, |
Surface Laplacian | Wavelet Transform | Support Vector Machine | 84% average |

X. Yong, |
None | Morphological Component Analysis | Creation of dictionary/template | Not specified |

J. Lin, |
Not specified | FFT | Simple Threshold | Results in average time consumed: 4.15–13.35 min |

H. Shahabi, |
None | Kalman Filter modeling | Simple Threshold | 98% modeling fitting |

M.K.I. Molla, |
None | EMD | Thresholding by statistical parameters | Not specified |

L. Ming-Ai, |
None | EMD | Simple Threshold | RRMSE against ICA: 0.1143 and 01186 |

T. Jung, |
None | Statistical parameters/ICA/E-ICA | Threshold filtering | Expert manual evaluation |

S. Woltering, |
None | Statistical parameters | Correlation | Correlation values for several electrodes |

P. Balaiah, |
Not specified | Statistical parameters | ADALINE (adaptive linear neuron) | SNR average 10.29 |

H. Nolan, |
Filtering not specified | Statistical parameters/ICA | Thresholding by statistical parameters | Specificity > 90% |

H. Cai, |
Not specified | ICA based features | Thresholding by statistical parameters | Correlation values: 0.8457 |