Authors

Claire Choi, Jessica Co, Nibha Akireddy

Abstract

Epilepsy affects around 65 million people around the world, with 30% of patients being unreceptive to antiepileptic drugs and in need of other methods of treatment such as neurostimulation or resective surgery — both of which require the localization of the seizure onset zone (SOZ). Electroencephalography (EEG), a noninvasive method of monitoring electrical brain activity, can sometimes be used to determine the location of the SOZ. However, EEGs are less sensitive and typically less effective for localization compared to invasive procedures such as Intracranial Electroencephalography (iEEG). Because neurologists rely on comparing and analyzing multiple montages to reach conclusions, we explored the most optimal montages that maximize the accuracy of SOZ localization using EEG recordings. The following montages were tested in this study: longitudinal, Laplacian, Cz referential, transverse, circumferential. The objective of our research was to build these montages and run sample data through them to identify their strengths and functionality by comparing raw EEG data and band power plots. While further trials are encouraged, the Cz referential montage had the greatest delta band power, and the longitudinal bipolar montage had the greatest theta, alpha, and beta band power.

Background

An electroencephalogram (EEG) is a type of brain test used to detect brain activity through electrodes that record synaptic potentials. EEGs can be utilized to detect abnormal brain activity correlated with brain disorders. A popular application of EEG tests is seizure detection, as it can not only identify abnormal brain activity, but also the readings it takes are also separated into channels that allow neurologists to locate the approximate location of the seizure onset zone (SOZ). An EEG functions by detecting electrical activity; however, by nature, it detects electrical activity from other areas besides the brain. As a result, EEGs tend to detect alternate noise on top of cerebral activity, known as artifact. Some relevant sources of artifact include muscle movement, physiological functioning, and even movement from the environment or the calibration of the electrodes. In order to limit this noise, the calibration of electrodes (montages) are manipulated, and the effectiveness of these montages can be measured in their band power.

Image: phase reversal indicates max. voltage.

As aforementioned, EEGs measure brain activity, and this is accomplished by calculating the voltage via the difference in electric potential between electrodes (see image above). In recording the relative electric potential, EEGs are able to eliminate unnecessary noise from the reading. A montage is an arrangement of raw EEG data that subtracts different electrical signals from others to decrease noise (Benbadis, S. R., 2006). One type of montage is a referential montage, where all raw EEG data is subtracted from one reference electrode. In referential montages, the center electrode, Cz, is typically used as the reference, as relatively, it is an electrically quiet spot on the brain and will not result in the omission of relevant activity. Another widely known montage is the bipolar montage, where electrodes are arranged such that the channels of electrodes that are placed in series are subtracted from each other (see image below for an example). Montages tested in this study, including their strengths and weaknesses will be further discussed in the following paragraphs.

Image: the “double banana” is a type of bipolar montage also referred to as the longitudinal bipolar montage — the two will be used interchangeably throughout this paper — it is a widely-used montage method for its versatility. Source: EEG Atlas. (See code.)

The distance between electrodes and which electrodes that are considered in the sample can impact the reading of the EEG. A typical placement of the electrodes is known as the 10-20 system, in which electrodes are placed evenly between the four anatomical landmarks: nasion (1), inion (1), and preauricular points (2). The distance between electrodes is 10%-20% of the maximum length (nasion to inion) or width (left preauricular point to the right preauricular point) (Benbadis, S. R., 2006). Certain arrangements result in incorrect source localization, as they do not capture activity in vital parts of the brain. The objective of changing the distance between electrodes and the number of electrodes utilized is to be able to effectively localize the SOZ. 

In seizure detection, EEGs are not only used to diagnose epilepsy, but also to reveal and locate epileptic brain activity and patterns for the identification of the type of epilepsy. Distinguished types and characteristics of epilepsy include: generalized or localized; idiopathic, cryptogenic, or symptomatic type; and localized-related, mesiotemporal, or extratemporal/neocortical epilepsy (Benbadis, S. R., 2006). These types of epilepsies are also characterized by different brainwaves. Previous research has revealed, for example, that activity in the theta band is correlated with tumor related epilepsy (Douw, L., 2010). For the purpose and scope of our study, the following brainwaves are most relevant: delta, theta, alpha, beta. In order to ensure precision and validity in EEG diagnostics, it is necessary that montages are calibrated to decrease noise and artifacts. And montages will be evaluated based off of their band power. 

In diagnostic and screening procedures, different montages are used, as they serve different purposes. The aforementioned bipolar montage is highly versatile and used in screening processes; however, identifying the type of epilepsy and locating the SOZ, other forms of montages must be implemented. The common reference montage, which uses one electrode as the reference, is best with broadly distributed abnormalities in the brain, and the average reference montage, which averages all of the other electrodes as a reference, is versatile as well but susceptible to referential contamination. Referential contamination is the result of an electrode whose electrical potential is an outlier, shifting the average. The Laplacian montage mitigates the issue of referential contamination by only taking the average electrical potential of the nearest electrodes as a reference, performing proficiently with focal abnormal brain activity (Moeller, 2015).

Image: Laplacian Montage – electrodes are compared with the average of the adjacent electrodes, as opposed to all of the other electrodes or a single electrode. The smaller, more specific reference demonstrates the focal power of the Laplacian montage. (See code.)

The third montage tested was the Cz referential montage (see image below). In referential montages, a single electrode is used as the second input for every channel. In other words, the electrical potential recorded by a single electrode is subtracted from the rest of the recordings from the other electrodes. In this study, the Cz electrode will be used as the reference. It is known for its applications for broad brain abnormalities, as it is standardized by the same electrode, giving less weight to outliers. That being said, it proves weak for locating and identifying focal brain activity (Moeller, 2015).

Image: Cz Reference Montage – the center electrode is the second input for all EEG channels. (See code.)

The transverse montage is a type of bipolar montage that localizes the maximum of the discharge in the left-to-right direction (J Clin Neurophysiol, 2006.). In this study, adjacent electrodes in a chain were linked starting at electrode F7. While transverse montages are effective for localization in certain areas, there are a couple disadvantages . For one, in-phase cancellation can occur when waveforms at two points are relatively synchronous, which ultimately leads to false localization. (Britton JW, Frey LC, Hopp JLet al., 2016.) Bipolar montages in general are susceptible to the end-of-chain phenomenon, in which the last electrode in a chain outputs a downward deflection and no phase reversal is seen. To combat this effect, three chains were grouped as opposed to five as seen in the image below.

Image: Transverse Montage – a bipolar montage that localizes the maximum of the discharge in the left-to-right direction. (See code.)

The last montage studied was the circumferential montage (see image below). When observing the channels formed from this montage, the only electrode pairings that are unique to circumferential compared with the longitudinal bipolar montage are the Fp1-Fp2 and O1-O2 pairs (Britton JW, Frey LC, Hopp JLet al., 2006.). However, these two pairs are included in the transverse bipolar montages. Since there are only ten channels recorded through the circumferential montage, this montage is not as effective for localization as the other bipolar montages. However, because its chain is continuous, it may be a good supplementary montage to overcome the end-of-chain phenomenon that bipolar montages may experience. 

Image: Circumferential Montage – a bipolar montage that paths around the circumference of the head, starting at Fp1 and proceeding in a counterclockwise direction. (See code.)

Methods and Materials

Data 

The sample data we analyzed was a 10 second EEG recording provided by our project supervisor via the Stanford Electrical Engineering Department. The file ‘Sleep1.edf’ was loaded into MATLAB where we built our code. 

Building Montages

The following images show the execution of methodology in code behind building each montage.

Longitudinal bipolar montage –

Laplacian montage

Cz referential montage

Transverse bipolar montage

Circumferential montage

Raw EEG Graphs

Raw EEG graphs were made by using the stackedplot function in order to graph the waves from distinct channels with respect to time (Stackedplot. (n.d.)).

Power Spectral Density (PSD) Graphs

Using Welch’s method, which is an approach for power spectral density estimations, signal from the time domain was converted to signal in the frequency domain. While PSD graphs were not used to draw conclusions about the effectiveness of montages, they were a necessary step to implement before analyzing band power graphs (Pwelch. (n.d.)).

Band Power Graph

The bandpower function returns the average power in each input signal. Band power plots provided a quantitative measure to compare the performance of different montages. This was a notably advantageous method, as it used the Fourier transform to provide distinct band powers for each brain wave, allowing the ability to draw specific conclusions on the functionality of the montages tested (Bandpower (n.d.)).

Results

EEG Data

Figure 1: channels created by the longitudinal bipolar montage

In the longitudinal bipolar montage, F7 and Cz are the only two electrodes that exhibit prominent deflections; at approximately 2.2, 5.1, and 8.2 seconds, the downward deflections indicate that both F7 and Cz are relatively negative. Additionally, the contemporaneous upward and downward deflections happening in opposing directions (see Figure 1, channels T5-O1 and O1-P3) reflect relatively positive readings from the O1 electrode at the same times as F7 and Cz. 

Figure 2: channels created by the Laplacian montage

In contrast, the Laplacian montage emphasizes different electrode potentials. Fp1, C3, Pz, and C4 all exhibit downward deflections at times 2.2 and 5.1 seconds, indicating that they are all relatively positive in comparison to the remaining electrodes. In addition to 2.2 and 5.1 seconds, C3 and C4 exhibit an additional downward deflection at 8.5 seconds, and Pz at 7.7 seconds. Notable electrodes that are relatively negative are Cz and F7, which demonstrate upward deflections at 2.2, 5.1, 7.7, and 8.5 seconds (Figure 2).

There are similarities between the longitudinal bipolar montage and the Laplacian montage that reinforce the accuracy of the two montages: in both montages, F7 and Cz have upward deflections at the same time, and thus, are relatively more negative than the surrounding electrodes (Figure 1, Figure 2). The Laplacian montage, however, captured more clearly prominent phase reversals in other electrodes. The Laplacian montage also found relevant phase reversals that the longitudinal bipolar montage did not in the following electrodes: Fp1, C3, P3, Pz, and C4 (Figure 2). The channels relevant to these electrodes in the longitudinal bipolar montage do not exhibit any indication of these differences in electric potential, where the Laplacian montage shows these deflections quite starkly. In general, the amplitude of the Laplacian channels is lower, emphasizing the points where the discharge is.

Figure 3: channels created by the Cz referential montage

The Cz referential montage showed significantly less variance among EEG channels. All channels have the same deflection at similar amplitudes at times 2.2, 5.1, and 8.5 seconds (Figure 3). The Cz referential montage is not as sensitive to the deflections picked up in other montages, and at surface level, it would be difficult to discern all abnormalities from this montage.

Figure 4: channels created by the transverse montage

The channels with the most notable downward deflections in the transverse montage were C4-T4, C3-Cz, and C4-T4, which occurred at 2.2s, 5.0s, and 8.3s. Additionally, at 8.3s, the channels F3-F7 and O1-O2 experienced a significant downward inflection. Upward deflections occurred at 2.2s 5.0s, and 8.3s for the channels F7-Fp1, F8-F4, and Cz-C4. Channels O2-T6 and Pz-P3 had a downward deflection at 8.3s. 

Deflections in the transverse montage occurred between similar electrodes compared to the longitudinal bipolar montage, with slight variance due to differences in electrode selection for certain channels. Similar to the longitudinal bipolar montage, there are places where deflections occur in opposite directions (see Figure 4: T3-C3 and C3-Cz, Cz-C4 and C4-T4). The transverse montage shares the most similarities with the longitudinal bipolar montage.

Figure 5: channels created by the circumferential montage

From the raw data of the circumferential montage, the most prevalent downward deflections occur at times 2.2s, 5.0s, and 8.3s in channel Fp1-F7. At 8.3s, downward deflections also occur in channels O1-O2, T6-T4, and T4-F8; upward deflections occur in channels F7-T3, T3-T5, T5-O1, and O2-T6. O2-T6 and T6-T4 have opposite wave patterns. 

Compared to the other four montages, the circumferential montage has fewer deflections at the earlier times, but more at the 8.3s instance. Overall, with fewer channels and less obvious deflections, the circumferential montage is less effective in demonstrating obvious abnormalities compared to the other montages.

PSD Welch

Figure 6: the power spectral density of each channel in a Bipolar Double Banana montage. The area under each line correlates with the respective power of each channel.

Figure 7: the power spectral density of each channel in a Laplacian montage. The area under each line correlates with the respective power of each channel.

Figure 8: the power spectral density of each channel in a Cz referential montage. The area under each line correlates with the respective power of each channel.

Figure 9: the power spectral density of each channel in a Transverse montage. The area under each line correlates with the respective power of each channel.

Figure 10: the power spectral density of each channel in a Circumferential montage. The area under each line correlates with the respective power of each channel.

Band Power

Figure 11: comparison of the band power distribution into different types of brainwaves between a longitudinal/double banana bipolar montage, Laplacian montage, Cz referential montage, transverse montage, and circumferential montage.

Among all montage methods tested, the Cz referential montage had the greatest band power for delta brainwaves, at 25.46 db/Hz, and the circumferential montage had the lowest band power for delta brainwaves, at 14.72 db/Hz. The longitudinal bipolar montage had the greatest theta band power, at 10.64 db/Hz, and the circumferential montage had the lowest band power for theta activity, at 4.676 db/Hz. The longitudinal bipolar montage also had the greatest alpha power, at 13.18 db/Hz, while the Cz referential montage had the lowest alpha band power, at 3.853 db/Hz. The highest beta band power was at 14.62 db/Hz, for the longitudinal bipolar montage, followed by the Laplacian montage, at13.63 db/Hz. The lowest beta band power was at 5.512 db/Hz, for the Cz referential montage. Because frequency is measured in decibels, the negative values on the logarithmic scale are very small and negligible. Thus, the negative gamma, ripple, and fast ripple brain activity were considered negligible in drawing conclusions from our data.

Conclusion

At the surface level, the Laplacian montage appeared to be the most sensitive in the raw EEG data, and picked up more, quantitatively and qualitatively, deflections and phase reversals than other montages (Figure 2). For this reason, when it comes to locating the SOZ, the Laplacian montage may be the most ideal. However, a central part to diagnosing epilepsy is understanding how different types of epilepsy present themselves through different types of brainwaves. According to the band power plots, the Cz referential montage would be optimal for epilepsies that are characterized by delta activity (Figure 11). The transverse montage had a higher delta band power than the longitudinal bipolar montage, but a lower theta, alpha, and beta band power (Figure 11). This implies that transverse montages may be more sensitive to slower wavelengths than longitudinal bipolar montage (Figure 11). However, because of their complementary channel paths, longitudinal and transverse montage should work well when analyzed together. The circumferential montage, however, had significantly lower band powers for all wave types (Figure 11). From the raw EEG data, it was also the least effective for finding deflections (Figure 5). Both of these observations indicate that the circumferential montage is the least effective montage out of the five. The longitudinal bipolar montage would be optimal for epilepsies that are characterized by alpha, theta, and beta activity, which supports the claim that it is versatile and ideal for early screenings.

Future Directions

One limitation of our research was that we only were able to analyze one data set, and without specific patient information from the data, reaching a solid conclusion or making generalizations about the most effective montage was not possible. In future research, more data sets — specifically that of epilepsy patients — should be considered in order to give less weight to outliers. We suggest that beyond this research, deep learning models that use patient data can assist in identifying more accurate band powers for each montage. Further, next steps that arose from this study would be to find ways to build and measure montages that can take into account the depth of brain activity and distribute that weight accordingly. 

Acknowledgements

This paper and research project as a whole would not have been possible without the support and mentorship of our project supervisor, Nibha Akireddy, who, from lecturing about the fundamentals of EEG to giving tutorials on MATLAB, had provided us with the knowledge and resources to conduct this research. Finding spaces to effectively learn and conduct a study in is extremely difficult under social distancing restrictions and COVID-19 in California; we are endlessly grateful to Cindy Nguyen, the STEM to SHTEM program director, who not only monitored the logistics of distance learning, but was also pivotal in our foundation of programming. We would also like to extend our gratitude towards the Stanford Compression forum and its founding director, Professor Tsachy Weissman, without whom this project and these individuals would not have been given a space to collaborate.

References

[1] Band power – MATLAB bandpower. (2020). Bandpower. https://www.mathworks.com/help/signal/ref/bandpower.html

[2] Benbadis, S. R. (2006). Introduction to sleep electroencephalography. Sleep: A comprehensive handbook, 989-1024.

[3] Douw, L., Dellen, E. V., Groot, M. D., Heimans, J. J., Klein, M., Stam, C. J., & Reijneveld, J. C. (2010). Epilepsy is related to theta band brain connectivity and network topology in brain tumor patients. BMC Neuroscience, 11(1). doi:10.1186/1471-2202-11-103

[4] Guideline 6: A Proposal for Standard Montages to Be Used in Clinical EEG. (2006). Journal of Clinical Neurophysiology, 23(2), 111–117. https://doi.org/10.1097/00004691-200604000-00007

[5] Moeller, J. (Director). (2015, June 08). EEG Montages [Video file]. Retrieved August, 2020, from https://www.bing.com/videos/search?q=eeg+montages+moeller

[6] Pwelch. (n.d.). Retrieved August 07, 2020, from https://www.mathworks.com/help/signal/ref/pwelch.html

[7] St, E. K., Frey, L. C., Britton, J. W., Frey, L. C., Hopp, J. L., Pearce Korb, Koubeissi, M. Z., Lievens, W. E., Pestana-Knight, E. M., & St, E. K. (2016). Appendix 2. Principles of Digital EEG. Nih.Gov; American Epilepsy Society. https://www.ncbi.nlm.nih.gov/books/NBK390353/

[8] Stacked plot of several variables with common x-axis – MATLAB stackedplot. (2020). Stackedplot. https://www.mathworks.com/help/matlab/ref/stackedplot.html