Accepted Version IEEE SENSORS JOURNAL, VOL. XX, NO. X, JUNE 2016 1 Single-trial evoked potential estimation using iterative principal component analysis Md Rakibul Mowla, Student Member, IEEE, Siew-Cheok Ng, Muhammad S. A. Zilany, Raveendran Paramesran, Senior Member, IEEE, Abstract—In this paper, we have presented an iterative prin- cipal component analysis (PCA) method to obtain single trial evoked potential. The performance of the iterative PCA has been compared with the performance of other iterative component analysis methods such as, independent component analysis (ICA), canonical correlation analysis (CCA) and second order blind identification (SOBI), using simulated data at different SNRs as well as actual recordings of visual evoked potentials. In both the simulated and real cases, iterative PCA and CCA perform better than the other methods to estimate the amplitudes. In the estimated trials of the proposed method, the latency of the evoked potentials lies between ±4ms range of true latency in about 90% of the case but for other methods this percentage is around 60%. Index Terms—Evoked potentials, Single trials, Blind source separation, Principal Component analysis. I. I NTRODUCTION E VOKED potentials (EPs) are voltage fluctuations within the electroencephalogram (EEG) which are caused by an external sensory stimulus. EPs are routinely used for studying cognitive process [1], [2] and hence, neuroscientists have great interest on EPs. EPs have a wide range of uses from clinical to research such as clinical diagnosis [3], sleep research [4], monitoring of spinal cord motor function [5], observing the neonates early development [6], P3 component based controller [7], brain computer interfacing [8]–[10] and computer cursor control [11]. In research, EPs have immense use on EEG based brain computer interfacing(BCI) [7], [8], [11]–[14]. EPs can be readily measured by averaging a large number of single trial [15]. However, the averaging does not allow to analyze the changes in response from trial to trial. Single trial analysis allows to study trial to trial response changes in response to each stimulus. Moreover, the EEG lies between the frequency range of 0.1 to 100Hz with the magnitude range of 10 - 100µV. EPs also have overlapping spectra with the magnitude range of 1 - 10µV. EEG data are also comprised of different types of artifacts and thus all those artifacts with the stationary EEG activity make the parametric estimation of EP more M.R. Mowla was with the Department of Biomedical Engineering, Uni- versity of Malaya, 50603 Kuala Lumpur, Malaysia. Currently he is with the Department of Electrical & Computer Engineering, Kansas State University, KS, USA, 66506. E-mail: rakib.raju05@gmail.com S.C. Ng is with the Department of Biomedical Engineering, University of Malaya, 50603 Kuala Lumpur, Malaysia. e-mail: siewcng@um.edu.my M. S. A. Zilany is with the Department of Biomedical Engineering, Univer- sity of Malaya, 50603 Kuala Lumpur, Malaysia. e-mail: zilany@um.edu.my R. Paramesran is with the Department of Electrical Engineering, University of Malaya, 50603 Kuala Lumpur, Malaysia. e-mail: ravee@um.edu.my Manuscript received August 25, 2015; revised June 08, 2016. complex. Thus, different filtering methods were introduced to filter the background EEG activity [16]–[18]. A least mean square error filter was proposed by Yu and Mc Gillem for evoked potential estimation [16]. Later on, adaptive filters were introduced for minimizing more the mean square error and reducing the ensemble number for averaging [17], [18]. Weighted ensemble averaging was another attempt to improve the signal to noise ratio (SNR) of averaging methods [19]. Wavelets have also been introduced for decomposing and analyzing evoked potentials [20], [21]. All aforementioned methods were merely an improvement over ensemble averaging method and did not solve the problem of single-trial analysis. An early attempt has been made by using a linear minimum mean square error (MMSE) filtering to improve trial to trial SNR [22]. A recent study has been done on the use of spatial filter for single trial analysis [23]. However, filtering has a drawback as it considers EPs as a stationary signal whereas EPs and background EEG both are transient in nature and localized in different time and frequency location. Wavelets have been introduced to solve this problem of estimating single trial responses [24]–[27]. In all wavelet based methods, the thresholding coefficients have been estimated from the average of single trials and then the coefficients were applied to all the single trials. Though the idea is spectacular, it works well in estimating single trial responses at medium or high SNR. But for low SNR EP data, wavelets fail to recover single trials reliably. Later on, in 2007 Iyer and Zouridakis developed an ICA based iteration method which was highly efficient in estimating single trials [28] and found to be a better estimator than wavelets [29]. Although ICA is one of the most popular component separation method, various studies have shown that ICA may not work best in all conditions [30], [31]. In this study, various popular component separation methods such as CCA, SOBI and PCA are used in an iterative manner to extract visual evoked potentials and compared with the iterative ICA [28] method. The rest of the paper is arranged as follows: materials and methods are demonstrated in section II. Then results for simulated and actual data are presented in section III. Finally, the discussion and conclusion of this paper is provided in section IV. II. MATERIALS AND METHODS A. Subjects and EEG recordings 1) Synthetic data: In order to evaluate the performance of the proposed algorithm, a typical EP signal has to be simulated