Improving the quality of a collective signal in a consumer EEG headset

Abstract

This work focuses on the experimental data analysis of electroencephalography (EEG) data, in which multiple sensors are recording oscillatory voltage time series. The EEG data analyzed in this manuscript has been acquired using a low-cost commercial headset, the Emotiv EPOC+. Our goal is to compare different techniques for the optimal estimation of collective rhythms from EEG data. To this end, a traditional method such as the principal component analysis (PCA) is compared to more recent approaches to extract a collective rhythm from phase-synchronized data. Here, we extend the work by Schwabedal and Kantz (PRL 116, 104101 (2016)) evaluating the performance of the Kosambi-Hilbert torsion (KHT) method to extract a collective rhythm from multivariate oscillatory time series and compare it to results obtained from PCA. The KHT method takes advantage of the singular value decomposition algorithm and accounts for possible phase lags among different time series and allows to focus the analysis on a specific spectral band, optimally amplifying the signal-to-noise ratio of a common rhythm. We evaluate the performance of these methods for two particular sets of data: EEG data recorded with closed eyes and EEG data recorded while observing a screen flickering at 15 Hz. We found an improvement in the signal-to-noise ratio of the collective signal for the KHT over the PCA, particularly when random temporal shifts are added to the channels.

Fig 1. Channel locations and labels for the 14 electrodes on the Emotiv device.

There are two additional reference sensors, which are 30 degrees above the ears.

Fig 2. Segment of 3 seconds of normalized EEG time series for two different experiments measured using the Emotiv device.

(Left) Resting state with eyes closed. (Right) Watching a screen displaying a 15 Hz flicker. (Red) Channel locations are shown in Fig 1.

Fig 3. Mean SNR of the data corresponding to the two different experimental conditions for a subject in the study.

(Left) Average over 6 realizations with eyes closed. (Right) Average over 6 realizations looking at a 15 Hz flickering screen.

Fig 4. Evaluation of the global phase for the eyes closed experiment.

(Left) Instantaneous phases and linear fits. (Right) Residual phases as functions of the corresponding cycles. (Top) Only 5 channels have been used in the analysis. (Bottom) All available channels have been used.

Fig 5. Evaluation of the global phase for the 15 Hz flickering experiment.

(Left) Instantaneous phases and linear fits. (Right) Residual phases as functions of the corresponding cycles. (Top) Only 2 channels have been used in the analysis. (Bottom) All available channels have been used.

Fig 6. Single subject mean SNR and SNR enhancement, and their standard deviations of the PCA (blue) and KHT (red) collective rhythms.

The mean SNR over channels and experiments and its standard deviation is also shown in black/grey. PCA results correspond to the highest SNR eigensignal, which in this case is the second principal component. The results are an average over the six realizations for both experiments: eyes closed and flicker at 15 Hz, and we analyze 10 seconds for each realization. The parameters are: 14 channels, 128 samples per second, 1 Hz bandwidth for each center frequency (KHT) and 20 oscillations per window.

Fig 7. SNR for the eyes closed and flickering experiments.

The parameters are the same as in Fig 6, changing the number of channels and evaluating the SNR of the estimated collective rhythms only at the peaks of interest. These collective rhythms have been estimated using PCA (using first and second largest variance projections) and KHT methods. For this subject, the peaks are located at 9 Hz in the case of eyes closed and 15 Hz for the flickering.

Fig 8. Several examples of estimations of collective rhythms from a data set with the eyes closed, which are shifted from each other for clarity.

For each signal, the corresponding SNR for the frequency band 8.5–9.5 Hz is indicated in the legend. The PCA signals are computed by parts of 20 oscillations at 9 Hz and KHT signals are also computed at this frequency band (8.5–9.5 Hz). (Grey) Reference channel for the KHT estimations, this channel is the one with highest SNR. (Dark green) PCA estimation using the first principal component computed from the raw data. (Blue) PCA estimation using the second principal component computed from the raw data. (Red) KHT estimation. (Dark red) Example of a KHT estimation computed from time shifted data. We used 14 channels for the computation of all the collective rhythms, except for the top time series, which is the raw data of the reference channel.

Fig 9. Mean SNR for the eyes closed experiments at the corresponding peaks in the alpha band computed from KHT (left) and PCA (right) collective rhythm estimations.

The horizontal axis indicates the maximum time shift. This time shift is random uniform among time series and we obtain the SNR averaged over 30 random realizations and the corresponding experimental realizations. The vertical axis indicates the number of channels added in decreasing order of SNR used for the computation of both quantities.

Fig 10. Mean SNR for the flickering experiments at the 15 Hz peak computed from KHT (left) and PCA (right) collective rhythm estimations.

The horizontal axis indicates the maximum time shift. This time shift is random uniform among time series and we obtain the SNR averaged over 30 random realizations and the corresponding experimental realizations. The vertical axis indicates the number of channels added in decreasing order of SNR used for the computation of both quantities.