2017 29th International Conference on Microelectronics (ICM)

Low-Power High-Accuracy Seizure Detection

Algorithms for Neural Implantable Platforms
Khaled A. Helal∗ , Ahmed Yasser Abo Elmkarem∗ , Al-Moataz Bellah Refaat∗ , Taha Shawky Kamel∗ ,
Kareem Ayman Mohamed∗ , Mohamed Mahmoud Kamal∗ , Mohamed Mostafa Abdelrahman∗ ,
Hassan Mostafa∗† and Yehea Ismail†
∗ Department of Electronics and Communications Engineering, Cairo University, Egypt
† Center for Nanoelectronics and Devices, AUC and Zewail City of Science and Technology, Egypt

Abstract—Neural interfaces are systems operating at the in- The paper is organized as follows: the details of each feature
tersection of the nervous system and an internal or external are discussed in Section II. Section III highlights the classifiers
device. Neuro-stimulator is one of the most important neural that precede the feature extraction block. The simulation
interfaces used to help those who experience epileptic seizures.
To use this stimulator efficiently, seizure should be detected at the results are illustrated in Section IV and the proposed algorithm
right time. Seizure detection is basically founded on digital signal and the conclusions are summarized in Section V.
processing by monitoring certain features of the intracranial elec-
troencephalogram. Many of the previous researches are directed II. FEATURE EXTRACTION
to study the detection efficacies using different systems, however,
a few of them study the feasibility of implementing these systems Feature extraction employs mathematical operations on the
over a computationally limited power implantable platforms. In data recorded to differentiate them into a seizure and non-
this paper, five time-domain features and three wavelet-domain
features are investigated. Following that, a high accuracy seizure seizure states. These mathematical operations are performed
detection algorithm is presented with efficient power consumption on the data in the time domain or the wavelet domain. The
which makes it suitable for implantable neural systems. The features that comprise complex operations are excluded to
experiment results show that the presented method achieves avoid excess power consumption, then, three wavelet-domain
a sensitivity, specificity, and accuracy of 92.64%, 99.29%, and features and three time-domain features are selected to inves-
99.16% respectively for long-term iEEG seizure detection. The
area and power results are obtained from implementing the tigate them.
algorithms on Xilinx Spartan-6 XC6SLX45T FPGA.
Index Terms—Intracranial Electroencephalogram (iEEG), A. Time Domain Features
Seizure Detection, Feature Extraction, Time domain, Wavelet
domain, Support Vector Machine (SVM), Threshold, Matlab, 1) Accumulated Energy [2]:
Algorithm.
N
1 
I. I NTRODUCTION Eacc (k) = E(i + (k − 1)N ) (1)
N i=1
Epilepsy is a neurological disorder of the brain that more
than 50 million people globally suffer from. According to the whereE(i) = x2 (i), x(i) refers to the ith sample of data
World Health Organization (WHO), epilepsy is characterized given by the vector x, and k represents the window number
by recurrent seizures, which are physical reactions to sudden, of size N .
usually brief, excessive electrical discharges in a group of brain 2) Coastline [3]:
cells. Uncontrolled attacks result in placing patients at risk of
suffering oro-facial trauma [1]. 
N

CL(k) = abs [x(i + (k − 1)N ) − x(i − 1 + (k − 1)N )] (2)

The block diagram of seizure detection system consists of i=1
several blocks as shown in Fig. 1. The data acquisition block
passes the input iEEG signal with a fixed rate to the following 3) Hjorth Variance Parameter [2]:
blocks. The feature extraction block calculates specific features N
for the iEEG input signal on different domains. The classifier 1  2
V ar(k) = x (i + (k − 1)N ) − μ2k (3)
block monitors the features to detect the seizure occurrence. N i=1
When seizure occurs, the decision making block takes the final
decision and send to neural stimulator for treatment. N
1 
μk = x(i + (k − 1)N ) (4)
Data acquisition
N i=1
Feature Decision
and Classifier
Extraction Making
preprocessing In equations (3) and (4), the variance of a window of
N samples is calculated and averaged to obtain the mean
Fig. 1. Block diagram of the seizure detection system. variance.

978-1-5386-4049-4/17/$31.00 ©2017 IEEE

 2017 29th International Conference on Microelectronics (ICM)

B. Wavelet Domain Features &ĞĂƚƵƌĞdžƚƌĂĐƚŝŽŶ

Wavelet transform utilizes long time windows for more EŽ
ŝсϬ
precise low frequency information and short time intervals for
high frequency information. The wavelet transform has better EŽ &ĞĂƚƵƌĞх ŝх
ŝсŝнϭ ^ĞŝnjƵƌĞ
resolution and higher performance for the representation and dŚƌĞƐŚŽůĚ dŚƌĞƐŚŽůĚ zĞƐ
͍ zĞƐ ͍
the visualization of the epilepsy activity than the short time
Fourier transform. Discrete Wavelet Transformations (DWT)
Fig. 2. Flowchart shows the modified thresholding algorithm.
is used in tests as it analyzes the signal at different frequency
bands with different resolutions.
1) Relative Energy [4]: IV. PROPOSED ALGORITHM
N
 To make the process of simulation smooth and domain-
E(l) = Di2 τ /N (5) based, time-domain features are tested alone with SVM clas-
i=1 sifier as well as wavelet-domain features are tested alone with
where τ is the sampling interval and N is the number of SVM classifier to compare between both types of features.
DWT coefficients Di presented at scale l. The relative energy The time-domain algorithm uses two features, accumulated
of the scale is computed as energy and coastline features, to represent time-domain fea-
E(l) tures extraction, followed by SVM classifier as shown in Fig.
Er (l) = S (6) 3. Each feature alone has less detection efficacy than Hjorth
i=1 E(i)
feature, but the two feature combined have higher efficacy and
where S is the number of the wavelet scales. lower power consumption than Hjorth feature as the results
2) Coefficient of variation [4]: section will show.
σ(l)2
Vc = (7)
μ(l)2
1
N ĐĐƵŵƵůĂƚĞĚŶĞƌŐLJн ^sDǁŝƚŚ
Where μ(l)
 = N i=1 Di and
ŽĂƐƚůŝŶĞ ĐŽƵŶƚĞƌ
1
N
σ(l) = N i=1 (Di − μ(l))2 .
Fig. 3. The time-domain algorithm which represents time-domain features
3) Fluctuation index [4]: with modified SVM classifier.
N
1 
F I(l) = |Di+1 − Di | (8) Compared to time-domain features, wavelet-domain fea-
N i=1 ture extraction requires the conversion from time-domain
III. CLASSIFIERS to wavelet-domain first which expected to consume a large
A. Support Vector Machine (SVM) portion of power. The wavelet-domain algorithm uses a feature
extraction unit with of three features; average energy, fluctu-
SVM is a learning model with associated learning algo-
ation index, and coefficient of variation feature followed by
rithms that analyze data and is used for classification and
SVM classifier as shown in Fig. 4.
regression analysis. Given a set of training features, each
marked for belonging to one of two categories. In the proposed
case, it is either a seizure or non-seizure. The SVM training ǀĞƌĂŐĞŶĞƌŐLJн
^sDǁŝƚŚ
&ůƵĐƚƵĂƚŝŽŶŝŶĚĞdžн
algorithm builds a model that assigns new samples into one ŽĞĨĨŝĐŝĞŶƚŽĨǀĂƌŝĂƚŝŽŶ
ĐŽƵŶƚĞƌ
of the two categories.
The training of the model creates the threshold line which
Fig. 4. The wavelet-domain algorithm which represents wavelet-domain
divides the space into two areas, seizure area and non-seizure features with modified SVM classifier.
area. To detect the seizures after training, SVM compares the
new data with the trained data and assigns the new data to its Finally, as SVM requires complicated and high power
best related area. hardware, a new algorithm is derived. The proposed algorithm
consists of the selected time-domain features, accumulated
B. Multi-window Event-Based Counting Threshold
energy, and coastline, followed by multi-window event-based
Instead of using high power consuming classifier, multilevel counting threshold classifier as shown below in Fig. 5. This
thresholds [5], [6] are used as standalone detection algo- combination provides a high detection efficacy with the lowest
rithms.This classifier works as follows, if the feature extrac- power consumption possible.
tion’s output is larger than a preset patient-specific threshold, The internal block diagram of the proposed algorithm is
then it is decided that a seizure is detected. A counter is illustrated in Fig. 6.
triggered by the result of the previous comparison and it is
decided that a seizure is detected if, and only if, this counter V. SIMULATIONS AND RESULTS
exceeds a certain value as portrayed in the flowchart in Fig. The previous algorithms are implemented in the hardware
2. description Language (HDL) and applied to the FPGA plat-
 2017 29th International Conference on Microelectronics (ICM)

One hour is taken from each patient’s as a training data and

DƵůƚŝǁŝŶĚŽǁ
ĐĐƵŵƵůĂƚĞĚŶĞƌŐLJн several hours is used as a testing data. The classification of
ĞǀĞŶƚͲďĂƐĞĚ
ŽĂƐƚůŝŶĞ
ĐŽƵŶƚŝŶŐƚŚƌĞƐŚŽůĚ testing data and training data is shown in Table I.

Fig. 5. The proposed algorithm which represents time-domain features with TABLE I
multi-window event-based counting threshold without SVM block. C LASSIFICATION OF SEIZURE AND NON - SEIZURE EPOCHS IN THE
TRAINING DATA AND THE TESTING DATA FOR EACH PATIENT.

The training data The testing data

ŶĞƌŐLJ dŚƌĞƐŚŽůĚ Non- Non-
&ĞĂƚƵƌĞ ůŽĐŬ ^ĞŝnjƵƌĞŽƌ Seizure Num. of Seizure
EŽŶͲƐĞŝnjƵƌĞ Patient# Seizure Seizure
epochs hours epochs
WƌĞƉƌŽĐĞƐƐŝŶŐ epochs epochs
ŽĂƐƚůŝŶĞ dŚƌĞƐŚŽůĚ 1 10 890 6 101 5299
&ĞĂƚƵƌĞ ůŽĐŬ 2 21 879 1 3 897
3 17 883 5 72 4428
&ĞĂƚƵƌĞ ĞĐŝƐŝŽŶ
džƚƌĂĐƚŝŽŶ
ůĂƐƐŝĨŝĞƌůŽĐŬ
DĂŬŝŶŐ
4 13 887 2 54 1746
5 29 871 4 111 3489
6 3 897 3 11 2689
Fig. 6. Block diagram of the proposed seizure detection algorithm. 7 22 878 2 60 1740
8 48 852 3 140 2560
9 16 884 2 36 1764
form. The FPGA resources utilization and the power dis- 10 17 883 4 70 3530
sipation targeting Xilinx Spartan-6 XC6SLX45T FPGA are
calculated. The results are obtained by Xilinx ISE 14.5 tool. C. Selected Features Results
A. Terminology As shown in Table II , a comparison of all time-domain
In order to do a fair comparison between all the algorithms, features is stated [2].
several performance metrics are calculated for all the algo-
TABLE II
rithms. These metrics are sensitivity, specificity and accuracy. H ARDWARE IMPLEMENTATION RESULTS ( AREA AND POWER ) AND
These metrics are defined as follows: DETECTION EFFICACIES ( ACCURACY ) OF ALL THE COMPARED
TIME - DOMAIN FEATURES .
TP
Sensitivity = ×100 (9) Area (LUTs) Power (μW) Accuracy (%)
TP + FN
Accumulated Energy 284 2990 85.3
TN Coastline 64 600 74.7
Specif icity = ×100 (10) Hjorth 500 6520 91.1
TN + FP
TN + TP Based on Table II, the coastline feature has the most hard-
Accuracy = ×100 (11) ware efficiency, lowest power and area, and Hjorth variance
TN + TP + FN + FP
parameter has the best detection efficiency. However, Hjorth
They are calculated using four other parameters; True variance has a complicated hardware and consumes a power
Positives (TP), True Negatives (TN), False Positives (FP) and and occupy an area larger than accumulated energy and coast-
False Negatives (FN). These four metrics are obtained after line features combined. Therefore, the proposed algorithm uses
running the detection algorithm on the input iEEG data and the accumulated energy and the coastline features.
defined as follows: Table III summarize the area and power of the wavelet
1) TP: the number of seizure segments identified by both conversion and the three wavelet-domain feature extraction
the automated system and the experts of EGG. hardware.
2) TN: the number of non-seizure segments identified by
both the automated system and the experts of EGG. TABLE III
H ARDWARE IMPLEMENTATION RESULTS ( AREA AND POWER ) OF THE
3) FP: the number of seizure segments identified by the WAVELET CONVERSION ALL THE COMPARED WAVELET- DOMAIN
automated system but not by the experts of EGG. FEATURES .
4) FN: the number of non-seizure segments identified by
Area (LUTs) Power (μW)
the automated system, but not by the experts of EGG.
Wavelet Transform Unit 5509 8340
B. Input Data Relative Energy Feature 2281 81480
Coefficient of Variation Feature 1987 27770
This database, collected at the Children’s Hospital in Fluctuation Index Feature 66 210
Boston, consists of iEEG recordings from pediatric subjects
with intractable seizures [7]. Subjects were monitored for up The wavelet conversion hardware requires an area of 5509
to several days following withdrawal of anti-seizure medica- LUTs and a power of 8340 μW when implemented using
tion in order to characterize their seizures and assess their Daubechies filter [8]. The conversion requires area and power
candidacy for surgical intervention. greater than any of the three time-domain features. Hence, the
 2017 29th International Conference on Microelectronics (ICM)

results of wavelet-domain are added for the sake of comparison TABLE VI

with the time-domain features. R ESULTS OF THE PROPOSED ALGORITHM .

Patient Sensitivity Specificity Accuracy

D. System Simulation Results
1 100 99.55 99.56
Using MATLAB, the simulations of the three algorithms 2 86.36 96.31 95.4
using the input data stated above is obtained and summarized 3 88.24 99.43 99.22
4 64.91 99.8 99.25
in the three Tables V, IV and VI. 5 96.67 93.79 93.89
Table IV shows the results of the time-domain algorithm. 6 100 98.43 98.44
The accuracy varies from 43% to 97%. Although the results 7 91.67 98.49 98.42
of detection are sensitive to the selected training set, using the 8 92.68 98.6 98.33
9 92.31 96.63 96.58
same training set for all algorithm gives a good grasp of the 10 100 94.35 94.4
algorithms efficacy relative to other algorithms. Avg. 91.28 97.54 97.35
Table V shows the results of the wavelet-domain algorithm.
The accuracy varies from 83% to 98%, but for the most
patients it is above 90%. removal of all complex mathematical processes such as SVM
Table VI shows the results of the proposed algorithm. The and wavelet-domain features.
accuracy here varies from 94% to 100%. SVM is a powerful classification algorithm, nevertheless,
the Multi-window classifier outperforms it in this application.
TABLE IV This because of two reasons. First, Multi-window classifier has
R ESULTS OF THE TIME - DOMAIN ALGORITHM . two thresholds, this requires some offline design effort, how-
ever, it adds a degree of freedom that can be tuned to get both
Patient Sensitivity Specificity Accuracy
1 83.18 97.01 96.96
high sensitivity and high specificity. Second, the advantages
2 66.67 94.39 94.39 of SVM are tangible for the systems with high dimensions,
3 42.11 93.56 93.44 a high number of features, however, the proposed algorithm
4 49.12 76.6 76.59 uses only two features for seizure detection that reduces the
5 66.96 91.56 91.46
6 92.31 43.05 43.08 power consumption and yet provides high detection efficacy.
7 77.05 90.93 90.89
8 78.17 94.96 94.79 R EFERENCES
9 82.14 97.2 97.18 [1] L. Orosco, A. G. Correa, and E. L. Leber, “Epileptic seizures detection
10 84.21 93.88 93.78 based on empirical mode decomposition of eeg signals,” Management of
Avg. 72.19 87.31 87.26 Epilepsy-Research, Results and Treatment, pp. 3–20, 2011.
[2] S. Raghunathan, S. K. Gupta, H. S. Markandeya, K. Roy, and P. P.
Irazoqui, “A hardware-algorithm co-design approach to optimize seizure
detection algorithms for implantable applications,” Journal of neuro-
TABLE V science methods, vol. 193, no. 1, pp. 106–117, 2010.
R ESULTS OF THE WAVELET- DOMAIN ALGORITHM . [3] H. S. Markandeya, P. P. Irazoqui, and K. Roy, “Low-energy two-stage
algorithm for high efficacy epileptic seizure detection,” IEEE Transactions
Patient Sensitivity Specificity Accuracy on Very Large Scale Integration (VLSI) Systems, vol. 23, no. 1, pp. 208–
1 93.24 97.67 97.65 212, 2015.
2 66.67 97.43 97.43 [4] Y. Liu, W. Zhou, Q. Yuan, and S. Chen, “Automatic seizure detection
3 91.49 89.85 89.85 using wavelet transform and svm in long-term intracranial eeg,” IEEE
4 39.53 90.16 90.13 transactions on neural systems and rehabilitation engineering, vol. 20,
5 89.53 99.46 99.43 no. 6, pp. 749–755, 2012.
6 30.77 83.22 83.2 [5] M. Safi-Harb, M. T. Salam, D. K. Nguyen, and M. Sawan, “An im-
7 75 94.77 94.73 plantable seizure-onset detector based on a dual-path single-window
count-based technique for closed-loop applications,” IEEE Journal on
8 88.03 86.14 86.16
Emerging and Selected Topics in Circuits and Systems, vol. 1, no. 4, pp.
9 82.35 86.71 86.71
603–612, 2011.
10 90.41 92.5 92.48 [6] S. Raghunathan, S. K. Gupta, M. P. Ward, R. M. Worth, K. Roy, and P. P.
Avg. 74.7 91.79 91.78 Irazoqui, “The design and hardware implementation of a low-power real-
time seizure detection algorithm,” Journal of neural engineering, vol. 6,
no. 5, p. 056005, 2009.
VI. DISCUSSION AND CONCLUSION [7] A. L. Goldberger et al., “Components of a new research resource for
complex physiologic signals, physiobank, physiotoolkit, and physionet,
The results show the time-domain and the wavelet-domain american heart association journals,” Circulation, vol. 101, no. 23, pp.
algorithms have approximately the same detection efficacy 1–9, 2000.
[8] P. Sangeetha, M. Karthik, and T. KalavathiDevi, “Vlsi architectures
with the wavelet-domain algorithm shows a slightly higher for the 4-tap and 6-tap 2-d daubechies wavelet filters using pipelined
accuracy. However, based on the previous results, it is obvious direct mapping method,” in Innovations in Information, Embedded and
that the proposed algorithm exhibits the best performance Communication Systems (ICIIECS), 2015 International Conference on.
IEEE, 2015, pp. 1–6.
metrics. Also, it requires no domain transformation in contrary
to the wavelet-domain algorithm. Besides the simple classifi-
cation unit that requires significantly less power and area than
SVM. This algorithm shows better performance due to the