Noise Reduction on the Raw Signal of Emotiv EEG Neuroheadset

Transcription

1 Noise Reduction on the Raw Signal of Emotiv EEG Neuroheadset Raimond-Hendrik Tunnel Institute of Computer Science, University of Tartu Liivi 2 Tartu, Estonia jee7@ut.ee ABSTRACT In this paper, we describe the classification of thought patterns from the electroencephalography readings from the consumerlevel Emotiv EEG neuroheadset. We use support vector machine (SVM) classification on the frequency representation of the raw signal and try to improve the F1 score of our classification by applying noise reduction methods. Categories and Subject Descriptors I.5.4 [Pattern Recognition Applications]: Signal Processing low-pass filter, high-pass filter, baseline filter, kernel density estimation, support vector machine. General Terms Algorithms, Measurement, Experimentation. Keywords Emotiv EEG headset, classification, noise reduction. 1. INTRODUCTION Emotiv EEG is a low-cost neuroheadset that reads the voltage activity from the surface of the human head using 14 electrodes. Emotiv s consumer level devices have been used for example to direct movement of virtual objects [3; 10], real robots [2] and also as a way to analyze mental responses for different multimedia [4; 7]. Varying methods have been used to achieve those application, but all of them include some noise filtering and pattern classification. This research will focus on SVM [9] and applying different noise reduction techniques to improve pattern classification. First we classify 3 distinct mental states without any noise reduction and then see how the classification rate changes after applying filters on the raw data. Assessment of the classification is made using F1 score measure on the SVM model. This can be calculated as follows: precision recall F1 2 precision recall Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Ilya Kuzovkin Institute of Computer Science, University of Tartu Liivi 2 Tartu, Estonia ilja.kuzovkin@ut.ee We begin by describing the data format and the instances that we are trying to classify. Then we see how SVM performs on a given dataset and what parameters for the model give empirically good results. Next we try out different noise reduction methods and compare F1 scores of models trained on filtered data and the model trained with unfiltered data. Lastly, we perform a principal component analysis (PCA) to see if the tested noise reduction methods achieve a representation of the data that needs less components to be accurately described than the original data. We used Python s machine learning toolkit Scikit-learn [5]. 2. DATA FORMAT Data from the Emotiv EEG devices comes in a form of a 14- element vector, where the components represent relative values of voltages compared to the measures from 2 reference electrodes that are located behind the ears and are not included in the data. Table 1: Example of a raw data vector with values rounded Channel AF3 F7 F3 FC5 T7 P7 O1 Value O2 T8 P8 FC6 F4 F8 AF Each value in the data vector has ~4100 μv added to it based on the specification of the device. From this data we could read that the difference between sensors AF3 and the reference signal was: V One measurement doesn t give us much information about the mental state. Different types of brain activities vary from 3 Hz (delta waves) to 100+ Hz (gamma waves). Epoc EEG device has a bandwidth of 0.2 ~ 45 Hz with higher frequencies already removed. In order to get the frequency representation from our data vectors we use Short-time Discrete Fourier Transform that finds the occurring frequencies inside a small time window. To get higher time resolution we would like our window to be minimized. Because of the Gabor limit (i.e. uncertainty principle) smaller window sizes give us inaccurate spatial resolution. Furthermore because the sampling rate of the Emotiv EEG device is fixed, by the Nyquist-Shannon sampling theorem smaller window sizes limit the maximum detectable frequency [6]. Table 2 illustrates how these values change with the change in our window size.

2 Table 2: Maximum detectable frequency and the minimum distinguishable frequency depending on the window size Window size (s) Maximum detectable frequency (Hz) ~ Accuracy (Hz) ~3 2 ~ Based on this we see that our desired window size should be in the range (0.7, 1.0). The minimum bound comes from the fact that smaller window sizes will not be able to detect frequencies higher than ~45 Hz, but these frequencies contain useful information for mental state classification. Larger window sizes than 1.0 seconds are also discarded because of the loss in time resolution. We define our window stepsize with the window width, but use a parameter called shift in order to achieve overlapping windows. This allows us to get more information out of the same data and thus potentially improve the classification. Formula (1) describes our ith window s starting position, W0 = 0. W W i i W W i1 i1 stepsize (1 shift) windowsize After taking the samples from the window we perform the Discrete Fourier Transform (DFT) on them to find out the frequency components of our signal. In our data instances we use the magnitude of the frequency components. EEG signal often contains a linear drift that is not part of the useful information. This causes low-frequency components with high magnitudes as seen on Figure 1. We initially tried to remove it from the raw data via a operation called detrend, but that caused non-determinism in the later classification because of round-off errors. So we used a high-pass filter (HPF) on the frequency data to achieve the same result deterministically (Figure 2). (1) Figure 2. Frequency representation with HPF applied In the case where windowsize is 1 second, after performing the DFT we get 14 vectors of 64 frequency components. So our instance vector length would be 910. These will be the vectors we will try to classify. Our data is collected from a test subject who had one image, from a set of 3, shown to for 10 seconds. This trial was repeated several times with a randomly chosen image. The EEG data was collected during those trials. We have two data sets recorded with the same test subject and environment, but with a 10 minute break in between. Thus each of the data vectors in both sets also have a trial counter and the class (which image was shown) assigned to it. We use our first data set as training data and the second one, that was recorded after 10 minutes from the first, as test data to estimate the goodness of the trained model. 3. CLASSIFICATION We used support-vector machine (SVM) classification to build our model based on the frequency representation instances of our data. Different parameters were empirically tested to determine a good configuration of them. SVM uses a pseudo random number generator in order to shuffle the data before training the model. We tested several different seeds when finding a good parameter configuration. Tested parameters are in Table 3. Note that all applicable combinations of those were applied. Table 3. Different parameters that were tested Kernel (with degree) C Gamma Linear Polynomial (2) Polynomial (3) Figure 1. Frequency representation without applying HPF Polynomial (4) Polynomial (5) 10.0 Polynomial (10) RBF Sigmoid

3 When performing the testing we discovered that the best results are given by the linear and sigmoid kernels. Comparison of the best performances is in Table 4. We determined that the sigmoid kernel with window size 1.0, C 1.0, and gamma 10.0 performed reasonably well (e.g. F-score with seed 17) in the tested cases. So this was the parameter configuration we used to compare noise reduction techniques. Table 4. Example of SVC performance with seed=17 Kernel Window size Shift C Gamma F-score Sigmoid Sigmoid Sigmoid Sigmoid Linear None Linear None Prior to classification with SVM we are using a StandardScaler trained on the test data to remove the mean and reduce variance to unit size for both training and test data. Scaling based on the test data seems to give better results than scaling based on the training data. 4. NOISE REDUCTION 4.1 High-Pass Filter A filter that at least removes the zero frequency component of a signal can be considered a HPF [1]. We are using one to remove the linear drift that occurs in our signal and keep only the frequency components that have higher frequency than the zero frequency (threshold 1). This threshold gave roughly the same frequency representation as the detrend operation. Here we tried with different other thresholds. Based on the results in Table 5, we see that 1 is a good threshold. HPF threshold Table 5. Comparison of HPF thresholds F1 score Low-Pass Filter LPF-s remove the higher frequencies from a given threshold [1]. Because the highest detectable frequency we have with 1.0 second time window is 64 Hz, threshold 65 does not do any filtering. We tried with the following thresholds: 65, 60, 55, 50, 45, 40, 35, 30, 25. Table 6 shows the thresholds that achieved the highest F1 score. Threshold 65 (no LPF) is determined to be the best. LPF threshold Table 6. Comparison of LPF thresholds F1 score Baseline Filtering One of the mental states that we classified was a neutral state where the test subject saw an image of a circle and was asked to relax. We assumed that the measurements taken at that time could be considered as a baseline that on removal from the other states would improve our classification. We took means of each frequency component measured during the neutral state and subtracted them from all of the training and test data. Results were also clipped to be positive, without clipping the F1 score was worse. This didn t improve the result as can be seen from Table 7. Table 7. Comparison of baseline subtraction percentages. Baseline subtraction percentage F1 score We also tried subtracting the mean of the neutral and other class mean baselines from each instance on the training data, but this didn t give any improvement either. Figure 3. Shows the frequency components of an instance prior to baseline filtering (blue) and after (green). Figure 4. Shows the frequency components of an instance prior to baseline filtering (blue) and after (green) without the clipping of values.

4 4.4 Kernel Density Estimation Given a sampled signal we can estimate the probability distribution function of the given signal using a method called kernel density estimation (KDE) [8]. When estimating the value at an arbitrary point x, KDE-s kernel takes into account a number of neighboring points (inside a given bandwidth) of x with different weights. Several such kernel functions can be used, very common is a Gaussian bell-shaped function. This technique can be used to smooth the otherwise rough signal. We used KDE to smooth our frequency representation of each electrode s signal in a hope to better the classification. Effect of such smoothing can be seen from Figures 4 and 5. We also tried KDE on the raw data channels, but without any significant results. There are also scott factor for determining the bandwidth: bandwidth n d bandwidth n 4 1 ( d ) 4, and a silverman (SM) factor given by: 1 2 d 4 samples and d is the dimensionality of the data., where n is the total number of Figure 5. One data instance without KDE applied. Frequencies from 14 electrodes can be easily seen. We tried with different bandwidths, but smoothing of the signal didn t seem to give any improvement in the classification. This can be seen from Table 8. Kernel bandwidth Table 8. Comparison of KDE bandwidths. None F1 score Scott SM BASELINE FILTERING WITH KDE Baseline filtering and KDE smoothing alone didn t give an improvement in the classification. We tried a combination of those two methods that first performs the KDE smoothing on the frequency representation and then does baseline subtracting without clipping to zero. Example of an instance after this filtering can be seen in Figure 7. This gave an improvement for the classification s F1 score as can be seen from Table 9. Improvement also persisted in the case of other random seeds. Table 9. Comparison of baseline subtraction percentages and KDE bandwidths. F1 score has improved. Baseline subtraction percentage KDE kernel bandwidth F1 score PCA m None Figure 6. Instance with KDE applied on each electrode's data frame. Bandwidth size: 0.1. Figure 7. KDE with bandwidth 1.0 and baseline subtraction with coefficient 0.5.

5 6. PRINCIPAL COMPONENT ANALYSIS PCA finds n orthogonal vectors (principal components) that are ordered in a way where each preceding vector describes the data better than the next. We can define m to be a minimal number of components needed to represent our data with accuracy 99.9%. Frequency representation of the data with windowsize 1.0, shift 0.6, that has only HPF with threshold 1 applied to it, has 575 samples and 910 features. To convey 99.9% of that information we need 333 components according to PCA. After applying baseline filtering with subtraction percentage 0.5 and KDE with bandwidth 1.0, the number of components dropped to 16. This means that our noise reduction technique transforms the data into a format where we could reduce the feature count into 4.8% of its original size and only lose 0.01% of information. This can be observed from Figure 8. Figure 8. Original data instance (green) and data instance restored from 16 features after PCA transform (blue). When testing the classification on the transformed data, F1 score of the SVM dropped to This means that further parameter configuration search is necessary. 7. CONCLUSION In this paper we described the data format received from Emotiv EEG device, its representation in the frequency domain and researched mental state classification with SVM-s. Several noise reduction techniques were tested out: ideal high- and low-pass filters, baseline filtering and KDE smoothing. With only HPF removing the zero frequency component, a sigmoid kernel with C 10.0 and gamma 1.0 was empirically determined to give a good result in our environment with a resulting F1 score Combination of baseline filtering with coefficient 0.5 and KDE smoothing with kernel bandwidth 1.0 improved the F1 score to Principal component analysis of this method showed that we only need 4.8% data to convey the 99.9% information needed for the classification. Results show that different noise reduction and signal processing methods can improve classification of mental states from an EEG signal. Techniques tested and combined here are only a few that could be used for data preprocessing. 8. REFERENCES [1] Gonzalez, R. C., Woods, R. E Digital Image Processing, Third Ed., Prentice Hall [2] Grude, S., Freeland, M., Yang, C., Ma, H. Controlling mobile Spykee robot using Emotiv Neuro Headset, 32nd Chinese Control Conference (CCC) (China, 2013), [3] Hazrati, M. Kh., Hofmann, U. G. Avatar Navigation in Second Life using Brain Signals, IEEE 8th International Symposium on Intelligent Signal Processing (WISP) (Funchal, 2013), 1-7. [4] Moldovan, A.-N., Ghergulescu, I., Weibelzahl, S., Muntean, C. H. User-centered EEG-based Multimedia Quality Assessment, IEEE International Symposium on Broadband Multimedia Systems and Broadcasting (BMSB) (London, 2013), 1-8. [5] Pedregosa et al., Skikit-learn: Machine Learning in Python. JMLR 12 (2011) [6] Quatieri, T. F Discrete-Time Speech Signal Processing: Principles and Practice. Prentice Hall. Ch 2. [7] Sinharay, A., Chatterjee, D., Sinha, A. Evaluation of Different Onscreen Keyboard Layouts using EEG signals, IEEE International Conference on Systems, Man, and Cybernetics (Manchester, 2013), [8] Smolka, B., Plataniotis, K. N., Lukac, R., Venetsanopoulos, A. N. New Class of Impulsive Noise Reduction Filters Based on Kernel Density Estimation, IEEE International Conference on Acoustics, Speech, and Signal Processing Vol 3 (2003), [9] Tan, P.-N., Steinbach, M., Kumar, V Introduction to Data Mining. Addison-Wesley, Boston, MA [10] Wijayasekara, D., Manic, M. Human Machine Interaction via Brain Activity Monitoring, The 6th International Conference on Human System Interaction (HSI) (Poland, 2013),