Comparison of EEG source reconstructed functional networks in healthy subjects elicited during visual oddball task

In this paper we have reconstructed electroencephalography (EEG) sources using weighted Minimum Norm Estimator (wMNE) for visual oddball experiment to estimate brain functional networks. Secondly we have evaluated the impact of time-frequency decomposition algorithms and scout functions on brain functional networks estimation using phase-locked value (PLV). Lastly, we compared the difference between target stimuli with response (TR) and non-target with no response (NTNR) cases in terms of brain functional connectivity (FC). We acquired the EEG data from 20 healthy participants using 129 channels EEG sensor array for visual oddball experiment. Three scout functions: i) MEAN, ii) MAX and iii) PCA were used to extract the regional time series signals. We transformed the regional time series signals into complex form using two methods: i) Wavelet Transform (WT) and ii) Hilbert Transform (HT). The instantaneous phases were extracted from the complex form of the regional time series signals. The FC was estimated using PLV. The joint capacity of the time-frequency decomposition algorithms/scout functions applied to reconstructed EEG sources was evaluated using two criteria: i) localization index (LI) and ii) R. The difference in FC between TR and NTNR cases was evaluated using these two criteria. Our results show that the WT has higher impact on LI values and it is better than HT in terms of consistency of the results as the standard deviation (SD) of WT is lower. In addition, WT/PCA pair is better than other pairs in terms of consistency as the SD of the pair is lower. This pair is able to estimate the connectivity within parietal region which corresponds to P300 response; although WT/MEAN is also able to do that, However, WT/PCA has lower SD than WT/MEAN. Lastly, the differences in connectivity between TR and NTNR cases over parietal, central, right temporal and limbic regions which correspond to target detection, P300 response and motor response were observed. Therefore, we conclude that the output of the connectivity estimation might be affected by time-frequency decomposition algorithms/scout functions pairs. Among the pairs, WT/PCA yields best results for the visual oddball task. Moreover, TR and NTNR cases are different in terms of estimated functional networks.

posed EEG inverse problem for source localization have to be implemented, followed by the 90 FC estimation in the source space. inverse problem for EEG source localization in recent researches [42][43][44][45]. As an outcome, the 97 spatial resolution of the EEG has been improved using EEG source localization. 98 Secondly the brain FC estimation methods are categorized into linear and non-linear unique source for that particular cortical region. Using MAX, the unique source selected from 124 the maximum sources across all the dipolar sources within that particular cortical region. And, 125 PCA takes the first mode of the PCA decomposition of all the sources within a cortical region 126 to form a unique source of that particular region. Different regional time series can be generated 127 using different scout functions. Thus, the scout functions can affect the extraction of 128 instantaneous phases from regional time series signals which may affect the FC estimation 129 using PLV. 130 Hence we believe that the network differences between two oddball cases are 131 significant. In addition we also assume that different time-frequency decomposition algorithms 132 and scout functions may have slightly different impact on FC estimation.

133
Based on our hypotheses, we have two main objectives. 1) To evaluate the differences 134 in terms of functional connectivity among oddball cases in the source space. Previously the 135 evaluation was done in the sensor space [50,51,64], 2) To assess the joint capacity of the time-136 frequency decomposition algorithms / scout functions applied to reconstructed EEG sources to 137 evaluate brain FC elicited by our oddball task; as the joint capacity has not been evaluated by 138 other studies. In this study, we also assess the impact of the scout functions and time-frequency 139 decomposition algorithms on FC estimation. The flow chart of the research is depicted in Fig 1.  In an oddball paradigm, the researchers asked the participants to distinguish the novel stimuli 156 (target) within a series of randomly displayed frequent stimuli (non-target). 157 We randomly presented the target stimuli (circle) and non-target (square) stimuli on the 158 computer monitor for 500ms during our visual oddball paradigm [64]. The fixation time was 159 set as 1000ms. During the fixation, an empty dark screen was presented. We requested the 160 subjects to pay attention towards the monitor. They have to make motor response by pressing 161 the keyboard button when the target stimuli appears on the computer monitor. When the the 162 non-target stimuli appears; the motor response is not required. Total 135 visual stimuli were 163 presented on the monitor. 40 out of 135 stimuli were the target stimuli, whereas 95 stimuli were 164 the non-target stimuli. The stimuli were projected on the monitor randomly.

165
Our oddball paradigm is categorized into 4 different oddball cases. The 'correct' cases 166 are target stimuli with response (TR) and non-target stimuli with no response (NTNR). While, 167 the 'incorrect' cases are target stimuli with no response (TNR) and non-target stimuli with 168 response (NTR). In TR case, the participants correctly respond to the target stimuli, whereas in 169 TNR case the subjects fail to respond to the target stimuli. In NTR case, the subjects respond 170 incorrectly to the non-target stimuli. In NTNR case, the participants did not provide the motor 171 response when non-target stimuli appeared. In this study, we used TR case for the evaluation of the scout functions and time-frequency decomposition algorithms. Moreover, we used TR 173 and NTNR to compare the differences between the two cases in term of connectivity as these 174 two cases are opposite to each other.

175
The 128-channel sensor array (HydroCel Geodesic Sensor Net) from EGI company 176 with a sampling frequency of 250 Hz was used to acquire EEG data. The EEG data was 177 acquired from 20 right-handed healthy participants with an age of around 19-23 years with 178 normal or corrected-to-normal vision. None of them had a history of substance abuse and a 179 personal or family history of psychiatric or neurological diseases.

180
The sampling frequency of the EEG acquisition system is 250 Hz. The maximum time 181 period of the epoch is about 500ms. The data acquisition for TR case will be stopped once the 182 subjects provided the motor response, hence the period of the epoch for this case could be less 183 than 500 ms. The raw data was converted into Matlab format by netstation. The high frequency 184 artifacts and DC components were removed using a finite impulse response (FIR) digital filter 185 with a band-pass frequency range from 0.5 Hz to 70 Hz. The EEGLAB function called eegplot() 186 was used to plot the filtered EEG data. The EEG samples that consist of artefacts induced by 187 muscles contraction and eye blinking were manually rejected. For TR case, the clean EEG 188 data were shifted to the right to align with the event when the button is pressed by the 189 participants. For NTNR case, shifting is not required. The data of 2 participants was rejected.

190
For TR and NTNR cases, three trials have been randomly selected from each subject (3 × 18 = 191 54 trials) to perform the analysis. We performed the analysis 3 times by selecting different where W E is the diagonal weighting matrix which consists of weighted factor for depth . After that, the dipolar sources were projected on the 3D cortical surface. The Time-frequency decomposition 235 We used HT and WT to decompose the regional time series into time-frequency representation.

236
Hilbert Transform 237 The regional time series signals were decomposed into gamma band (30 -60 Hz) using FIR 238 band pass filter. After filtering, HT was applied. HT of a function f(t) can be defined as HT of a regional time series signal x(t) can be mathematically 1 240 represented as: where PV denotes the Cauchy principle value [69]. ). This process is to ensure the Morlet Wavelet has unit energy. extract the instantaneous phases of the regional time series signals in radian form.

254
The phase differences between two regional time series signals a and b at time ( , ) 255 bins t and trial n were computed as follows:

315
We applied WT and HT to decompose the regional time series signals into time-     By combining the analysis from LI and R, we can say that WT is more consistent than

431
HT. Moreover, we also realized that WT/PCA and WT/MEAN pairs have high performance.

432
Hence the results of Wavelet/PCA pair is more consistent. Besides that, PLV also was able to 433 localize more than 80% of networks (LI=0.831) using WT/PCA pair. The efficiency of this 434 pair is acceptable and the consistency of this pair is better. Therefore, we conclude that 435 WT/PCA is adequate and a good choice for visual Oddball task. 436 We performed the comparisons between TR and NTNR cases in order to validate the 437 performance of the WT/PCA pair based on LI and R criterions. Moreover, we would like to 438 analyse the differences of TR versus NTNR in terms of FC using R criterion. surface for TR and NTNR cases. As illustrated, the difference in terms of estimated 441 connectivity among both cases is obvious. Consequently, target and non-target stimuli elicited 442 different brain networks during oddball task. As depicted in Fig 7,  In this study our aim is to observe the FC for visual oddball task using the source space. We 498 used wMNE method to reconstruct the sources for data acquired by EEG for visual oddball 499 task. We applied three different scout functions (MEAN, MAX and PCA) to generate the 500 regional time series signals. We applied two time-frequency decomposition algorithms (HT 501 and WT) to represent the regional time series signals into complex functions. We extracted 502 instantaneous phases from the complex form of regional time series signals and estimated the 503 FC using PLV. The connectivity graphs were proportionally thresholded to retain 10% of 504 strongest networks. We evaluated the performance of the scout functions/time-frequency 505 decomposition algorithms pairs based on LI and R. Lastly, we compared the differences 506 between TR and NTNR cases based on the LI and R.

507
Our results demonstrate that time-frequency decomposition algorithms have higher 508 impact on the LI values than the scout functions. In addition, WT is better than HT in terms of 509 the consistency of LI. All pairs show good efficiency in connectivity estimation as all pairs 510 yield more than 80% of LI. However, WT/PCA pair is more consistent than others. Moreover,

511
WT/PCA is capable to estimate the connectivity within parietal region which corresponds to 512 P300 response. Lastly, we observe the differences in connectivity between TR and NTNR cases 513 over parietal, central, right temporal and limbic regions which correspond to target detection, 514 P300 response and motor response.

515
In conclusion, the outcome of the connectivity estimation might be affected by scout 516 functions/time-frequency algorithm pairs. Consequently, WT/PCA is the best choice for visual 517 oddball task as the scout function to generate regional time series signals and time-frequency 518 decomposition algorithms to transform the signals into gamma band for instantaneous phase 519 extraction. The TR and NTNR cases are different in terms of FC. Greater R values are observed 520 over the regions which correspond to P300 and motor response.

521
The performance of the combinations of scout functions/time-frequency decomposition 522 algorithms have not been evaluated so far in the literature. We found that the WT/PCA is best 523 for visual oddball task. We believe that PCA is superior for generation of regional time series