IROS2019国际学术会议论文集0926

Abstract Brain-computer interface is a framework which provides a communicating pathway between the human brain and neuroprosthetic devices. In this work, we have performed signal smoothing to fix unregulated electroencephalogram (EEG) fluctuations and to effectively collect hidden patterns in corresponding EEG spectra. To do it, we have applied a Savitzky Golay function because of its ability to preserve spectral properties without distorting it much. For feature selection, we have applied the Filter Bank Common Spatial Pattern (FBCSP) algorithm with the Principle Component Analysis (PCA). FBCSP creates 11352 features from EEG signals and PCA gradually reduces the features to 185 most significant features. These features are utilized in the classification process by the eXtreme Gradient Boosting (XGBoost) algorithm with suitable node split criteria to manage optimal tree height. A five-fold cross-validation approach concludes the superior performance of XGBoost in terms of minimizing execution time (3.7 times faster in the training phase) and providing improved accuracy as compared to existing results. Our approach enhances classification accuracy (88.80%) by approximately 10% over regularized common spatial patterns (78.01% accuracy), 15% over shift variance approximation (73.84% accuracy) and 15% over Riemannian approach (74.77% accuracy). It also concludes that the pre-consideration of the noise level in EEG spectra provides a better approximation. Index Terms - Brain-computer interface, Signal processing, Savitzky-Golay approximation, Filter bank common spatial pattern, Gradient Boosting. I. INTRODUCTION Brain-computer interface (BCI) or brain-machine interface (BMI) is a communicating framework (Fig. 1) which facilitates the human brain to communicate with external devices such as the wheelchair system 1, neuroprostheses 2 and robots 3 via a dedicated communication channel. This communication system takes advantage of the tempo- spatial characteristics of electroencephalogram (EEG) neural signals for the screening of several mental state patterns. In the current scenario, neurophysiologists are highly intended 1Department of Computer Science and Engineering, Indian Institute of Technology (BHU), Varanasi 221005 India anuragtiwari.rs.cse17itbhu.ac.in 2Department of Computer Science and Engineering, Indian Institute of Technology (BHU), Varanasi 221005 India amrita.cseiitbhu.ac.in to identify the deep view of these hidden patterns within these EEG signal spectrum. As a primary objective, the BCI framework translates aforesaid patterns into a set of device controlled commands so that different intelligent devices can be operated even remotely. Based on the design and EEG signal modeling approaches, we can categorize BCI systems into two major classes. In the first class of BCI systems, a subject performs an action after getting visual or auditory cue impulse generated by the computer. Thus these systems analyze neural activities within a predefined cue window and perform the windows feature extraction and their processing in a sequential manner 4. These systems are studied under synchronous BCI (cue- based) 5 that facilitate the BCI system to initialize the trials. In this paper, we use a dataset that has been extracted using cue-based BCI paradigms for demonstration purposes (described in section 3). In the second class, BCI systems provide a more natural mode of communication between brain and external applications because it functions independently to any external cue stimulus. Contrary to cue based systems, it can extract and process EEG sample features without any sequence. The functioning of these BCI systems depends on their ability to recognize cue on the screen where event-related synchronization (ERS) and event-related desynchronization (ERD) occur 6. During these events, it is highly complicated to recognize available patterns within a corresponding mental state because of the presence of several erratic features (for ex. spectral density, energy, spectral centroid, and zero crossing rate) in EEG spectra 7. The nonstationary chaotic behavior of these EEG signals (existing in both synchronous as well as asynchronous types) also plays a vital role to depreciate the accuracy of a classification approach which further limits the performance of the BCI framework over real-time streaming data. Although these studies have contributed a variety of new hypotheses to recognize the internal dynamics of human brain A Multiclass EEG Signal Classification Model using Spatial Feature Extraction and XGBoost Algorithm Anurag Tiwari1 and Amrita Chaturvedi2 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) Macau, China, November 4-8, 2019 978-1-7281-4003-2/19/$31.00 2019 IEEE4169 still they have illustrated a range of challenges also. Some noticeable primary challenges are the unavailability of common noise recognition framework 8, uncertain characteristics of the EEG spectrum, bulky data, high dimensionality, intrinsic non-stationary nature of neural data, missing values, non-robust classification rules and poor cross validation measures. Commonly, these challenges decelerate the performance of decision tree based classification approaches in terms of poor execution rate and improper recognition of the chaotic nature of input data. To deal with these issues in an efficient way, we have employed XGBoost learning with the decision tree classification approach. Historically, this algorithm has shown its prominent performance in epileptic signal classification 9, handling large dataset over graphics processing unit (GPU) 10, cloud computing 11 and many more research domains. Some other reasons for selection of the XGBoost algorithm are inspired by its ability to handle missing data, high speedup, efficient way to discriminate continuous data, robustness against noise, implicit variable selection ability and effective learning from non-linearly distributed data 12, 13. The paper is organized into six sections. Section II elaborates the related research work and their limitations. Section III presents the spatial-temporal properties of the BCI- 2008 dataset. In Section IV, we describe Signal to Noise Ratio SNR) optimization using function smoothing and its influence on Filter Bank Common Spatial Pattern (FBCSP) and eXtreme Gradient Boosting (XGBoost) algorithm. Section V demonstrates the performance of the XGBoost classifier with statistical considerations and the obtained results are compared with those obtained by the state of the art methods. In the ultimate section, we conclude the entire work and provide the scope of future work. Figure 1. Proposed model for closed-loop BCI implementation using a multichannel EEG spectrum II. RELATED WORK As discussed, BCI utilizes EEG signal properties to make communication between the brain and several intelligent devices. The EEG signal properties comprise features like high variance, spike presence, noise and many more. These signal properties have been exploited in several classification experiments but proper quantification of these properties is still an open question for researchers. Several noticeable researches focussed on these properties in terms of non- stationary features, effective pre-processing techniques and the scope of classification algorithms. The BCI system analyses EEG signals as a collection of observations and measures their non-stationary nature as a function of available noise levels. These noise levels are studied as signal to noise ratio (SNR) which is quite low in EEG signals 14. These studies claim that the position of electrodes over scalp decide how the SNR will map to different cortical regions. To meet these challenges, several advanced EEG BCI systems, e.g., neuronal action potentials, steady-state visual evoked potential (SSVEP) with stimulus frequencies ( 4 Hz), a slow cortical potential (SCP), mu and beta rhythms, P300 evoked potentials were developed 15. The main aim of these developments was to enhance the SNR value of specific EEG signal classes. These EEG signal classes geared up BCI performance by improving the information transfer rate (ITR) and training time. These high-performance systems approached highly precise results as compared to retrospective results 16. In addition to aforesaid improvements, some noise reduction approaches 17 have also been applied that have improved the performance of listed systems. Some common filter approaches such as Empirical Mode Decomposition (EMD) extract intrinsic oscillating components by defining energy levels of decomposed structures of noisy signals 18. It derives the decomposition basis function from the parent signal rather than a predefined basis vector. Based on experimental results, the EMD application is specifically limited to reduce Gaussian noise from the multi-channel spectrum. To decompose signals in an efficient manner, a new extension of the EMD algorithm is applied in the form of EEG signal translation invariant and wavelet thresholding 19. This novel approach reduces noise efficiently when SNR value is low and/or sampling frequency is significantly high. However, the performance of these algorithms significantly depreciates during the detection of different SNR levels within a single decomposition window. To investigate aforesaid challenges deeply, a noise level discrimination-based approach was applied by Cabrera et al. 20. The authors explored the power density of the signals by applying the Fast Fourier Transform (FFT) and separated variable background noises from parent signals. This study was only validated on EEG signals emitted from Broca regions so its robustness is still questionable. This problem was solved by Lotte et al. 21 by applying the Common Spatial Pattern (CSP) technique. In this technique after applying noise reduction methods, CSP was integrated with Tikhonov regularization to reduce overfitting. It gave approximately 10% enhanced results as compared to 4170 conventional CSP. Another conclusion of these results establishes that conventional CSP is sensitive towards outliers and overfitting. So, in the case of poor SNR value, its direct application is not advisable. Aforesaid results were enhanced again by Barachant et al. 22 by integrating topology of input multiclass spectra with Riemannian geometry and analyzed with CSP and linear discriminant analysis (LDA). This integrated framework demonstrated improved average classification accuracy of multiclass MI-BCI dataset from 65.1% to 70.2%. In further researches, Raza et al. 23 manifested the influence of the probabilistic distribution of input data over multiclass EEG classification. The authors exploited a covariant shift to measure the behavior of the process and integrated it with the feature extraction algorithm. The applied approach gave very competitive results with regularised CSP 21. III. BCI COMPETITION 2008 DATASET BCI Competition IV- 2008 Graz data set A, is an open source dataset and licensed under Creative Commons Attribution-No Derivatives license (CC BY-ND 4.0) 24. In this paper, we have used it for our demonstration and validation of proposed procedures. The dataset comprises signals spectra collected from 9 participants. This spectrum consists of 22 EEG channels and 3 EOG channels with left mastoid as a reference. It is four-class motor imagery (MI) task-based dataset where class 1 represents left-hand movement, class 2 represents right-hand movement, class 3 represents both feet movement and class 4 represents the tongue movement. The data were recorded from each participant in two different sessions having 6 runs per session (12 runs for individual participant). Each run consists of 48 trials (12 for each class), with 288 trials in a single session. The dataset was recorded with Ag/ Agcl electrodes implanted with an inter-electrode distance of 3 cm. All the extracted signals were sampled with a 250Hz frequency using a bandpass filter to collect frequency between 0.5 Hz to 100 Hz. Throughout the session, the amplifier frequency was fixed to 100V. Additionally, a notch filter was also applied to attenuate the band spectrum below to 20Hz. IV. PROPOSED METHODOLOGY In this paper, our proposed work consists of three important phases, viz. data preprocessing using appropriate mathematical optimization approach, feature analysis which deals with effective set of spatial-temporal feature extraction and then dimension handling and ultimate phase is classification using XgBoost algorithm. The overall block diagram is shown in Fig. 2. To explore the details regarding applied steps, all the three phases are discussed below. A. Data Preprocessing This phase comprises two sequential steps viz. (1) Noise level reduction and (2) Normalization. In the initial phase, noise level reduction means maximization of SNR value in given signal instances so that there is minimal distortion in signal shape. The mathematical background behind distortion minimization is motivated from objective function approximation by convoluting the adjacent data Figure 2. Block diagram of the proposed approach for BCI multiclass classification point in sample observation. We have considered SNR value in the context of abrupt changes in signal spectra and have estimated it as an approximated function of original EEG data without altering its functional properties. To adjust these abrupt changes, we have applied curve smoothing using Savitzky-Golay function 25 and then we have measured changes in terms of weighted dispersion metric. Fig. 3 shows different noise distributions in the context of different color noise and log weighted dispersion in training data samples. Further, we have scaled heterogenous EEG signal instances so that we can easily compare the performance of classification algorithm over each class of BCI data significantly. Further, a min-max normalization threshold (0:1) is fixed such that each instance value must fall in the defined data range. To do it, min and max values of instances from each EEG sample are extracted and then each value of the data sample is scaled down using the following formula (Equation 1): = min() max()min() (1) Where is real signal instance at time t and is scaled value of same instance at time t. B. Feature extraction and classification using XGBoost The feature extraction phase is a crucial step in multi-class EEG signal classification. In this paper, we have applied a spatial-temporal based subject specific algorithm termed as Filter Bank Common Spatial Pattern (FBCSP) (Figure 4) for multi-trial BCI EEG dataset classification. 4171 Figure 3. Performance enhancement in training data samples using Savitzky-Golay filter for subject A01T. All the images shown in the left side show noise level in raw data and the right-side images show a reduction of the noise level in corresponding raw data. Both the two windows (right column) show weighted dispersion after Savitzky-Golay smoothing of corresponding EEG signal band. A complete window of size 1800 data points is considered as a benchmark and is decomposed into 10 equal size sub-window. One of the main reasons to choose FBCSP is to maximize the relative variance between the pair of values in the Mutual Information States 26. Another criterion for the selection of FBCSP is its ability to discriminate subject specific EEG signal spectrum. In several demonstrations 26, 27 FBCSP has shown noticeable performance gain as compared to the state-of-the-art models. Structurally, the FBCSP algorithm (Fig. 4) comprises 4 steps to select spatial-temporal features from the EEG spectrum. In phase 1, the EEG channel is decomposed into equidistant signal chunks using the Chebyshev Type II filter bank. In the second phase, decomposed signal chunks are linearly transformed into the following feature vector (Equation 2). = 1,2, (2) where 2 represents m pairs of CSP features for bandpass filtered EEG measurements. In the third phase, features are selected based on Mutual Information of Best Individual Feature (MIBIF) 26,27. This algorithm sorts initial k features in decreasing order by considering mutual information of features. Finally, we have classified observations by merging weak learning-based features into a single strong feature in an iterative manner. This process provided a scalable and end to end decision tree based boosting approach i.e. called XGBoost approach 28. Figure 4. A generalized framework of the FBCSP approach. V. PERFORMANCE EVALUATION The XgBoost approach is implemented to resolve this four- class (left hand, right hand, tongue and feet) classification problem. This approach is also called an ensemble method because it employs a synergy that involves multiple varieties of classifiers to obtain better predictive performance than that which could be achieved by any of the single classifiers. These approaches are very popular and have been successfully used to reduce variance, bias and achieve improvement in prediction. In Python, we have used a scikit- learn 29 library to implement all the classification related tasks. In section (III), we have discussed the participants and channel details of the collected data. For classification with the XgBo