early detection of epileptic activity on eeg signals using phase … · 2019-02-06 · common...

4
Early Detection of Epileptic Activity on EEG Signals using Phase-Preserving Quantization Method Sylmarie Dávila-Montero, Ehsan Ashoori, and Andrew J. Mason Department of Electrical and Computer Engineering, Michigan State University, East Lansing, MI 48824 Abstract— This paper demonstrates the use of a data decimation method, called phase-preserving quantization (PPQ), for early seizure prediction. PPQ consists of a) amplifying and filtering the neural signals around the frequency band of interest, and b) compressing the filtered signal using a 1-bit quantizer with a 0-V single-threshold decision. The ability of PPQ to retain phase information and predict seizure events while compressing the signal resolution to a single bit is demonstrated using electroencephalography (EEG) recordings from the Children’s Hospital Boston-MIT (CHB-MIT) EEG database. Results show 97% accuracy when calculating synchrony values using PPQ, which is an improvement of 7% when compared to previously published results. The presented improved method enables the early detection of seizure events, resulting in a decrease in phase synchrony computation time while allowing an increase in the number of recording channels that can be screened when using EEG. Keywords—EEG, seizure prediction, data compression, phase synchrony, data decimation I. INTRODUCTION Advancements in neural recording technologies and neural signal processing techniques have provided significant insight in our understanding of brain function and ability to develop treatments for neural disorders. For example, patients with neurological conditions, such as epilepsy, have gained benefit from neural recording technologies. Affecting over 3.4 million people in the U. S., epilepsy is a neurological disorder caused by unusual neural activity in the brain. These alterations in brain activity cause episodes called seizures. Typically, seizure events can be recorded using electroencephalography (EEG), which is a well-known method in which electrodes are placed on the scalp to capture the aggregate electrical activity of a population of neurons. Recorded neural activity with millisecond resolution is then analyzed offline and seizure events are identified. In fact, identifying seizure events in real-time, even predicting them before they develop, is at the core of improving treatments for this neurological condition. Various methods have been developed to address the need for prediction and early detection of seizure events. These methods are divided in two main categories: 1) threshold-based algorithms such as phase synchrony thresholding [1], and coastline thresholding [2], and 2) machine-learning algorithms such as support vector machine (SVM) [3] and level learning set (LLS) classifiers [4]. The study in [1] shows that increasing the number of features demands more resources but does not necessarily contribute to better accuracy in classification. This study also revealed that threshold-based and machine-learning- based algorithms have comparable outcomes. Therefore, if the right features are selected and a simple thresholding method is used, fewer resources are demanded, which is necessary for real- time hardware implementations. Phase synchrony thresholding methods have been implemented efficiently in hardware with the end goal of predicting seizure events in real time using implantable devices [5]. However, phase synchrony results for performing real time prediction have only been shown for a limited number of channel pairs. Thus, compressing the recorded neural signals before calculating phase synchrony could decrease computation time and thus allow the number of simultaneously screened recording channels to be significantly increased. This paper presents new results for a data decimation method, called phase-preserving quantization (PPQ), with the capacity to retain instantaneous phase information within drastically compressed neural signals, allowing early detection of seizure events. EEG recordings from human subjects that suffered from seizures were used to verify our method. Results show that PPQ allows accurate early detection of seizure events while reducing the resolution of the signals to a single bit, permitting an extreme reduction in hardware resources and/or a significant increase in the number of channels that can be processed at a time. II. BACKGROUND A. Neural oscillatory activity Interactions among neurons produce oscillations in the delta (0 – 4 Hz), theta (5 – 8 Hz), alpha (9 – 12 Hz), beta (12 – 30 Hz), and gamma (> 30 Hz) frequency bands [6]. These frequency ranges have been defined based on their biological significance or the measured distribution over the scalp. For example, rhythms in the delta band are usually only perceived in adults in a deep sleep state. Thus, large amounts of activity in the delta band in awake adults is considered abnormal and related to neurological disorders [7]. Rhythms in the theta band are related to early stages of sleep, relaxation activities, meditative concentration, tasks that become automatic, and conscious awareness. Similar to delta band, it is believed that high theta activity in awake adults is related to neurological diseases [7]. Alpha rhythms are more prominent over occipital, parietal, and posterior temporal regions in the state of relaxed wakefulness with eyes closed. Activity in the alpha band is related to body relaxation, mental effort, and memory brain function, which makes them useful signals to measure mental effort [8]. Beta rhythms are more prominent in the frontal and central regions of the brain and are associated with engaged activities. Consequently, and similar to alpha rhythms, beta rhythms might be useful to measure mental effort and motor activities. Lastly, rhythms in the gamma band are the fastest 978-1-5386-3603-9/18/$31.00 ©2018 IEEE

Upload: others

Post on 08-Jul-2020

3 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Early Detection of Epileptic Activity on EEG Signals using Phase … · 2019-02-06 · common practice to filter the EEG signals into the frequency bands of interest to later be analyzed

Early Detection of Epileptic Activity on EEG Signals using Phase-Preserving Quantization Method

Sylmarie Dávila-Montero, Ehsan Ashoori, and Andrew J. Mason Department of Electrical and Computer Engineering, Michigan State University, East Lansing, MI 48824

Abstract— This paper demonstrates the use of a data

decimation method, called phase-preserving quantization (PPQ), for early seizure prediction. PPQ consists of a) amplifying and filtering the neural signals around the frequency band of interest, and b) compressing the filtered signal using a 1-bit quantizer with a 0-V single-threshold decision. The ability of PPQ to retain phase information and predict seizure events while compressing the signal resolution to a single bit is demonstrated using electroencephalography (EEG) recordings from the Children’s Hospital Boston-MIT (CHB-MIT) EEG database. Results show 97% accuracy when calculating synchrony values using PPQ, which is an improvement of 7% when compared to previously published results. The presented improved method enables the early detection of seizure events, resulting in a decrease in phase synchrony computation time while allowing an increase in the number of recording channels that can be screened when using EEG.

Keywords—EEG, seizure prediction, data compression, phase synchrony, data decimation

I. INTRODUCTION

Advancements in neural recording technologies and neural signal processing techniques have provided significant insight in our understanding of brain function and ability to develop treatments for neural disorders. For example, patients with neurological conditions, such as epilepsy, have gained benefit from neural recording technologies. Affecting over 3.4 million people in the U. S., epilepsy is a neurological disorder caused by unusual neural activity in the brain. These alterations in brain activity cause episodes called seizures. Typically, seizure events can be recorded using electroencephalography (EEG), which is a well-known method in which electrodes are placed on the scalp to capture the aggregate electrical activity of a population of neurons. Recorded neural activity with millisecond resolution is then analyzed offline and seizure events are identified. In fact, identifying seizure events in real-time, even predicting them before they develop, is at the core of improving treatments for this neurological condition.

Various methods have been developed to address the need for prediction and early detection of seizure events. These methods are divided in two main categories: 1) threshold-based algorithms such as phase synchrony thresholding [1], and coastline thresholding [2], and 2) machine-learning algorithms such as support vector machine (SVM) [3] and level learning set (LLS) classifiers [4]. The study in [1] shows that increasing the number of features demands more resources but does not necessarily contribute to better accuracy in classification. This study also revealed that threshold-based and machine-learning-based algorithms have comparable outcomes. Therefore, if the right features are selected and a simple thresholding method is

used, fewer resources are demanded, which is necessary for real-time hardware implementations.

Phase synchrony thresholding methods have been implemented efficiently in hardware with the end goal of predicting seizure events in real time using implantable devices [5]. However, phase synchrony results for performing real time prediction have only been shown for a limited number of channel pairs. Thus, compressing the recorded neural signals before calculating phase synchrony could decrease computation time and thus allow the number of simultaneously screened recording channels to be significantly increased. This paper presents new results for a data decimation method, called phase-preserving quantization (PPQ), with the capacity to retain instantaneous phase information within drastically compressed neural signals, allowing early detection of seizure events. EEG recordings from human subjects that suffered from seizures were used to verify our method. Results show that PPQ allows accurate early detection of seizure events while reducing the resolution of the signals to a single bit, permitting an extreme reduction in hardware resources and/or a significant increase in the number of channels that can be processed at a time.

II. BACKGROUND

A. Neural oscillatory activity

Interactions among neurons produce oscillations in the delta (0 – 4 Hz), theta (5 – 8 Hz), alpha (9 – 12 Hz), beta (12 – 30 Hz), and gamma (> 30 Hz) frequency bands [6]. These frequency ranges have been defined based on their biological significance or the measured distribution over the scalp. For example, rhythms in the delta band are usually only perceived in adults in a deep sleep state. Thus, large amounts of activity in the delta band in awake adults is considered abnormal and related to neurological disorders [7]. Rhythms in the theta band are related to early stages of sleep, relaxation activities, meditative concentration, tasks that become automatic, and conscious awareness. Similar to delta band, it is believed that high theta activity in awake adults is related to neurological diseases [7]. Alpha rhythms are more prominent over occipital, parietal, and posterior temporal regions in the state of relaxed wakefulness with eyes closed. Activity in the alpha band is related to body relaxation, mental effort, and memory brain function, which makes them useful signals to measure mental effort [8]. Beta rhythms are more prominent in the frontal and central regions of the brain and are associated with engaged activities. Consequently, and similar to alpha rhythms, beta rhythms might be useful to measure mental effort and motor activities. Lastly, rhythms in the gamma band are the fastest

978-1-5386-3603-9/18/$31.00 ©2018 IEEE

Page 2: Early Detection of Epileptic Activity on EEG Signals using Phase … · 2019-02-06 · common practice to filter the EEG signals into the frequency bands of interest to later be analyzed

oscillatory brain signals and are related to motor functions, perceptions, or consciousness, among others.

B. Information content of EEG signals

EEG electrodes are placed on the scalp to capture the summed electrical activity of groups of neurons. EEG has the capacity to capture neural activity up to 125 Hz at a millisecond temporal resolution [9]. Because activity in different frequency bands corresponds to different cognitive processes, it is common practice to filter the EEG signals into the frequency bands of interest to later be analyzed independently. Information about the distribution of activity across these frequency bands can provide insights related to brain damage, sleep disorders, cognitive functions, and epileptic activity [9]. Moreover, changes in amplitude due to an event and changes in the instantaneous phases of the signal are of interest when analyzing EEG recordings. Instantaneous phase information is used to determine synchronization and desynchronization events between two signals (or channels) existing in the same frequency band and recorded at the same period of time. In signals showing seizure activity, it has been proven that the amount of synchrony between multiple signal oscillations will change significantly, both before and during a seizure event [10], [11].

III. METHODS

A. Phase synchrony

Because significant changes in synchrony between signals from different channels are a strong indicator of a seizure event, phase locking value (PLV) can be used to calculate synchrony levels. PLV compares the phases between two signals and is often the preferred method to compute synchrony because of its tolerance to changes in magnitude [12].

To compute PLV between two neural signals, x1(t) and x2(t), a band-pass filter is applied to both signals to filter them around the desired frequency band. Then, the Hilbert transform of xn(t), where n=1,2, is used to compute the analytic form, Yn, of the signal:

= ( ) + ( ) (1)

( ) = ( ) (2)

where H denotes the Hilbert transform operator and the integral is evaluated as a Cauchy principal value.

Next, the instantaneous phases (over time), n(t), are computed by calculating:

( ) = tan ( )( ) (3)

After the instantaneous phases of each signal are determined, the difference between any two signals is calculated as ( ) = ( ) − ( ). This factor can then be used to calculate PLV as:

= ∑ ( ) (4)

where N is the number of data points in the signal. The resulting PLV will range between 0 and 1, where values closer to 0 show two signals are less synchronized and values closer to 1 are

more synchronized. PLV can be calculated across all channels in a neural recording, resulting in a triangular matrix, called connectivity matrix (CM), containing the PLVs of all combinations of two channels.

B. Phase-preserving quantization (PPQ)

We have developed a data decimation method called PPQ that saves hardware resources while maintaining high spatial and temporal resolution of instantaneous phase components [13], [14]. This efficient data decimation method, initially developed to enable wireless data transmission of thousands of electrocorticography (ECoG) channels, consists of (a) amplifying and filtering the neural signals around the frequency band of interest, and (b) compressing the filtered signal using a 1-bit quantizer with a 0-V single-threshold decision: every data point over 0 V is set to 1 and every data point below 0 V is set to 0. The 0-V threshold is set by analyzing the RMS signal value. After signal compression using PPQ, the PLV is calculated across pairs of channels over a window of time. Because the size of the window can influence the PLV results, we analyzed the effect of 15 different window sizes ranging between 10 and 3000 samples. To measure the effect of window size on the PLV calculation, we obtained EEG recordings from the Children’s Hospital Boston-MIT (CHB-MIT) EEG database [15], [16], which contains EEG recordings from pediatric subjects with seizures. The EEG data was recorded from a 23-channel array, pre-amplified and sampled at 256 Hz with a resolution of 16 bits. EEG recordings from one subject were filtered around 15-25 Hz, which has been demonstrated previously to contain sufficient PLV information to predict seizure events [5]. PLV was then calculated across channels for the resultant filtered signals, generating a 16-bit CM and an average PLV for the generated 16-bit CM. Moreover, to evaluate the effect that the window size has when using PPQ, the same filtered EEG signals were compressed to 1-bit using PPQ, and PLV was calculated across channels, generating a 1-bit PPQ CM and an average PLV for the 1-bit PPQ CM. Fig. 1 shows the results of the PLV calculation based on different window sizes. Fig. 1a and 1b show the 16-bit CMs for the same recording at the same particular period of time, but calculated using a window size of 10 samples and 512 samples, respectively. Fig. 1a shows high synchrony across all recording channels, which indicates that all channels are highly synchronized. However, this level of synchrony is unrealistic for EEGs, which suggests that a larger window size is required. Fig. 1b shows a CM with values ranging from 0.2 to 1, which is more common in these types of recordings. Fig. 1c shows the average PLVs for both the 16-bit CM and the 1-bit PPQ CM, given different window sizes. It can be observed that the smaller the window size, the more biased the results are to indicate high synchrony. On the other hand, a small window size is preferred in order to operate sufficiently fast to detect seizure events in real time. In Fig. 1c, it can be observed that the average PLV starts to converge to similar values after taking a window size of 512 samples. A window size of 512 samples at a sampling frequency of 256 Hz, represents a PLV calculation every 2 seconds of data. Because this is the smallest

Page 3: Early Detection of Epileptic Activity on EEG Signals using Phase … · 2019-02-06 · common practice to filter the EEG signals into the frequency bands of interest to later be analyzed

window size that provides the most stable PLVs, a window size of 512 samples was selected to obtain the results presented in this paper.

IV. RESULTS AND DISCUSSION

By applying the same analysis method presented in section III.B, two recording sections from a single subject were analyzed for early detection of seizure activity. Each recording section contains one seizure event. The CMs for both recordings around the seizure time were obtained to, based on the results, select pairs of channels with the potential to show a wide range of PLV variability. Fig. 2 shows the CMs (16-bit and 1-bit PPQ) for a single 23-channel recording section (chb01_04). It can be observed that the same physical areas of high PLV and low PLV are highlighted in both the 16-bit CM and the 1-bit PPQ. By evaluating the similarities between both CMs, the 1-bit PPQ provides a 97% accuracy when compared to the 16-bit ground truth. The accuracy was calculated as follows:

= ∑ − (5)

= (1 − ) ∗ 100 (6) where CMq is the 1-bit CM, CM16b is the 16-bit CM, and NM is the number of elements in the CM. The obtained accuracy shows an improvement of 7% in accuracy when compared to previously published results where the effect of the window size was not considered [13].

Based on the CM results, channels 16 and 21 were selected for the first recording section because of their average PLV of around 0.6. This mid-level PLV suggests that the signals do not maintain either a high or low level of synchrony, but a variation of it. Fig. 3 shows early seizure detection results for the selected pair of channels. Signals for each channel were filtered around

15-25 Hz (see Fig. 3a and 3b). The seizure event can be observed between 1467 secs and 1494 secs. To compare the performance of PLV over drastically decimated data using PPQ, PLV was calculated for the original 16-bit filtered signal and for the 1-bit PPQ filtered signal. Fig. 3c and 3d show the results of PLV over time for the 16-bit filtered signal and the 1-bit PPQ filtered signal, respectively. The similarity in the results can be observed visually for both cases, as well as the ability of PPQ to retain the necessary information to predict a seizure event 400 secs before it happens.

Fig.2. Comparison of 16-bit ground-truth CM (top) with 1-bit PPQ CM(bottom) of recording section chb01_04 at an instance of time in where seizureactivity was recorded.

Fig.3. Early seizure detection results based on a pair of channels from recording section chb01_04: (a) shows 15-25 Hz band-passed signal from channel 16; (b) shows 15-25 Hz band-passed signal from channel 21; (c) shows the original 16-bit average PLV result and a PLV threshold that detects a seizure event 400 secs before it happens, likewise (d) shows the quantized1-bit average PLV result obtained by applying PPQ and a threshold that detects a seizure event 400 secs before it happens.

Fig.1. Analysis of window size for PLV calculation: (a) shows a 16-bit CM obtained by calculating PLV with a window size of 10 samples and (b) showsa 16-bit CM obtained by calculating PLV with a window size of 512 samples,both from the same EEG recording; (c) shows the average PLV based on theselected window size for a 16-bit CM and for a 1-bit PPQ CM.

Page 4: Early Detection of Epileptic Activity on EEG Signals using Phase … · 2019-02-06 · common practice to filter the EEG signals into the frequency bands of interest to later be analyzed

Fig. 4 also shows early seizure detection results for the selected pair of channels at a second recording section (chb01_16). Signals for each channel were filtered around 15-25 Hz, and the seizure event can be observed in Fig. 4a and 4b between 1015 secs and 1066 secs. Similar to the analysis for the chb01_04 recording section above, PLV was calculated for the original 16-bit filtered signal and for the 1-bit PPQ filtered signal. Fig. 4c and 4d show the results of PLV over time. In this example, the similarity among the PLV results can also be visually observed. In addition, the results confirm the ability of PPQ to retain phase information needed to calculate PLV and perform early detection 150 secs before seizure events, even when the resolution of the neural signals have been reduced to a single bit. Compressing the neural signal to a single bit helps increase the computational speed of the PLV calculation and could help decrease power and area in implantable devices predicting seizure events.

V. CONCLUSION

This paper presented improvements and new results for a custom phase-preserving quantization (PPQ) data decimation method. PPQ has been shown to highly compress neural signals while retaining instantaneous phase information that allows early detection of seizure events. Based on an analysis performed to improve PPQ results, a window size of 512 for a frequency band of 15-25 Hz and a sampling rate of 256 Hz was chosen to obtain real time accurate phase synchrony results. The correct window size can provide a PPQ accuracy of up to 97%, 7% more than the previously published results. The results of the analysis with EEG signals containing seizure

events demonstrated that PPQ can retain the necessary phase information to obtain PLV results that are comparable to the 16-bit ground truth values, allowing signal compression to a single bit while maintaining the ability to predict seizure events.

ACKNOWLEDGEMENT

This work was supported by GRFP-NSF award number DGE-1424871.

REFERENCES [1] H. Kassiri, S. Tonekaboni, M. T. Salam, N. Soltani, K. Abdelhalim, J. L.

P. Velazquez, and R. Genov, “Closed-loop neurostimulators: A survey and a seizure-predicting design example for intractable epilepsy treatment,” IEEE Trans. Biomed. Circuits Syst., vol. 11, no. 5, pp. 1026–1040, 2017.

[2] M. Shoaran, C. Pollo, K. Schindler, and A. Schmid, “A fully integrated IC with 0.85-μw/channel consumption for epileptic iEEG detection,” IEEE Trans. Circuits Syst. II Express Briefs, vol. 62, no. 2, pp. 114–118, 2015.

[3] J. Yoo, L. Yan, D. El-Damak, M. A. Bin Altaf, A. H. Shoeb, and A. P. Chandrakasan, “An 8-channel scalable EEG acquisition SoC with patient-specific seizure classification and recording processor,” IEEE J. Solid-State Circuits, vol. 48, no. 1, pp. 214–228, 2013.

[4] W. Chen, H. Chiueh, T. Chen, C. Ho, C. Jeng, M. Ker, C. Lin, Y. Huang, C. Chou, T. Fan, M. Cheng, Y. Hsin, S. Liang, Y. Wang, F. Shaw, Y. Huang, C. Yang, and C. Wu, “A fully integrated 8-channel closed-loop epileptic seizure control,” IEEE J. Solid-State Circuits, vol. 49, no. 1, pp. 232–247, 2014.

[5] K. Abdelhalim, V. Smolyakov, and R. Genov, “Phase-synchronization early epileptic seizure detector VLSI architecture,” IEEE Trans. Biomed. Circuits Syst., vol. 5, no. 5, pp. 430–438, 2011.

[6] A. Schnitzler and J. Gross, “Normal and pathological oscillatory communication in the brain,” Nat. Rev. Neurosci., vol. 6, no. 4, pp. 285–96, 2005.

[7] A. Kübler, B. Kotchoubey, J. Kaiser, N. Birbaumer, and J. R. Wolpaw, “Brain-computer communication: Unlocking the locked in,” Psychol. Bull., vol. 127, no. 3, pp. 358–375, 2001.

[8] W. Klimesch, “EEG-alpha rhythms and memory processes,” Int. J. Psychophysiol., vol. 26, no. 1–3, pp. 319–340, 1997.

[9] A. M. Beres, “Time is of the essence: A review of electroencephalography (EEG) and event-related brain potentials (ERPs) in language research,” Appl. Psychophysiol. Biofeedback, vol. 42, no. 4, pp. 247–255, 2017.

[10] F. Mormann, K. Lehnertz, P. David, and C. E. Elger, “Mean phase coherence as a measure for phase synchronization and its application to the EEG of epilepsy patients,” Phys. D Nonlinear Phenom., vol. 144, no. 3, pp. 358–369, 2000.

[11] T. I. Netoff and S. J. Schiff, “Decreased neuronal synchronization during experimental seizures.,” J. Neurosci., vol. 22, no. 16, pp. 7297–7307, 2002.

[12] R. Q. Quiroga, A. Kraskov, T. Kreuz, and P. Grassberger, “Performance of different synchronization measures in real data: A case study on electroencephalographic signals,” Phys. Rev. E - Stat. Nonlinear, Soft Matter Phys., vol. 65, no. 4, pp. 1–14, 2002.

[13] S. Dávila-Montero and A. J. Mason, “An in-situ phase-preserving data decimation method for high-channel-count wireless μECoG arrays,” Biomed. Circuits Syst. Conf. (BioCAS), 2017 IEEE, pp. 30–33, 2017.

[14] E. Ashoori, S. Dávila-Montero, and A. J. Mason, “Compact and low power analog front end with in-situ data decimator for high-channel-count ECoG recording,” Circuits Syst. (ISCAS), 2018 IEEE Int. Symp., vol. 2, pp. 6–10, 2018.

[15] A. L. Goldberger, L. A. N. Amaral, L. Glass, J. M. Hausdorff, P. C. Ivanov, R. G. Mark, J. E. Mietus, G. B. Moody, C. Peng, and H. E. Stanley, “PhysioBank, PhysioToolkit, and PhysioNet: Components of a new research resource for complex physiologic signals,” Circulation, vol. 101, no. 23, pp. 215–220, 2000.

[16] A. H. Shoeb, “Application of machine learning to epileptic seizure onset detection and treatment,” pp. 157–162, 2009.

Fig.4. Early seizure detection results based on a pair of channels fromrecording section chb01_16: (a) shows 15-25 Hz band-passed signal fromchannel 3; (b) shows 15-25 Hz band-passed signal from channel 7; (c) showsthe 16-bit average PLV result and a PLV threshold that detects a seizure event150 secs before it happens, likewise (d) shows the 1-bit average PLV resultobtained by applying PPQ and a threshold that detects a seizure event 150 secsbefore it happens.