Neurophysiology, Vol. 44, No.

2, June, 2012

Seizure Detection Algorithms Based on Analysis

of EEG and ECG Signals: a Survey
S. Nasehi1 and H. Pourghassem1
Received July 8, 2011.

Epilepsy is a chronic disorder of the CNS that predisposes individuals to recurrent seizures.
Computerized seizure detection algorithms will enable alerting systems that may decrease the harm
of the seizures. This paper attempts to provide a comprehensive survey of different types of seizure
detection algorithms and their potential role in diagnostic and therapeutic applications. Major recent
algorithms use electroencephalogram (EEG) and electrocardiogram (ECG) signals to detect the seizure
onset and seizure event. In these algorithms, various features are extracted from the EEG signal alone or
in concert with the ECG signal until the patients are classified into two classes, seizure and non-seizure.
We identify three major categories for seizure detectors; EEG-based seizure-event detectors, EEG-
based seizure-onset detectors, and EEG/ECG-based seizure-onset detectors. In addition, some other
related issues, such as dataset and evaluation measures, are also discussed. Finally, the performance of
algorithms is evaluated, and their capabilities and limitations are described.

Keywords: epilepsy, seizure detection algorithm, EEG, ECG, feature extraction, classification.

INTRODUCTION These types of seizures are known as medically

intractable seizures [4]. The latter severely limit
Epilepsy is a chronic disorder of the CNS that the independence and mobility of an individual
predisposes individuals to recurrent seizures. and can result in social isolation and economic
Electrographically, a seizure is a sudden transient hardship. Most disturbing is that refractory seizures
aberration in the brain electrical activity that significantly increase an individual’s chance of
produces disruptive sensory and motor behavioral experiencing burns, lacerations, skull fractures, and
symptoms. These symptoms range between a lapse in even sudden unexpected death [5]. The negative
attention, a sensory hallucination, or a whole-body influence of uncontrolled seizures extends beyond
convulsion [1]. Epilepsy is not a “single” disease the individual to affect their family members,
but a family of syndromes that share features of the friends, and the whole of society. The families and
recurrent seizures. Epilepsy may develop as a result friends of people with epilepsy experience chronic
of inheriting mutations in molecular mechanisms anxiety and adapt their lives to ensure the safety
that regulate neuronal behavior, migration, and/or of their loved one [6]. Society incurs an annual
organization. Alternatively, it may develop as a loss of 12.5 billion dollars in health care costs and
result of brain (craniocerebral) trauma, stroke, losses in productivity [7]. There is a need for novel
some cerebral infections, or brain malignancy [2]. therapies that better control seizures, as well as for
Fifty million people worldwide have epilepsy. technologies that help both the individual and their
In the United States, epilepsy affects 3 million family to cope with the consequences of seizures.
people and is the third most common neurologic Among novel therapeutic systems, local electrical
disorder after Alzheimer’s disease and stroke [3]. stimulators, such as vagus nerve stimulators (VNS)
In an unfortunate subset of 1.2 million individuals, [8] and trigeminal nerve stimulators (TNS) [9], are
frequent and unpredictable seizures persist despite employed to stop the headway of a seizure prior to
treatment by one or multiple anti-epileptic drugs. the extension of clinical symptoms. In these systems,
the mass brain electrical activity (EEG) alone or in
concert with other physiological signals (ECG) of a
1
Department of Electrical Engineering, Najafabad Branch, Islamic Azad
patient are recorded and monitored by portable [10]
University, Isfahan, Iran.
Correspondence should be addressed to H. Pourghassem or wearable [11] devices in order to recognize the
(e-mail: [email protected]). seizure. Computerized seizure detection algorithms

174 0090-2977/12/4402-0174 © 2012 Springer Science+Business Media, Inc.

Seizure Detection Algorithms Based on Analysis of EEG and ECG Signals 175

will enable novel therapeutic and alerting systems to affect the progression of a seizure [19]. Within
that may decrease the harm of the seizures. A the realm of alerting, seizure-onset detectors can
therapeutic system capable of detecting and reacting prompt a patient or a care provider to ensure safety
to the onset of a seizure may administer a local or administer a fast-acting anticonvulsant. Seizure-
electrical [12], thermal [13], or neurochemical [14] event detectors can enable physicians to better titrate
stimulus that halts the progression of a seizure prior therapy (pharmacological or otherwise) over time
to the development of the symptoms. Moreover, [20]. A seizure-event detector within the ambulatory
just-in-time local therapy can relieve patients of setting can provide physicians with a summary
the toxic side effects that accompany systemic of the number, frequency, duration, and time of
administration of multiple anti-epileptic drugs. An individual seizure experiences. By correlating this
alerting system (Fig. 1) equipped with seizure onset information with different medication regimens, a
detection can warn the patient of the seizure prior physician can more quickly decide on a treatment
to the development of debilitating symptoms or can plan that maximally benefits the individual.
notify a family member so that the consequences of Research on seizure detection methods began
a seizure are limited [15]. Knowledge that a reliable with the development of seizure-event detectors
warning will be issued rapidly following seizure [21]. The developed detectors were meant to detect
onset may restore within individuals the confidence seizures of any individual with epilepsy, i.e., they
to overcome the limits on life that accompany were patient-nonspecific. The variability within
seizures. EEGs severely limited the detection accuracy of
A seizure detector can be classified either as a these patient-nonspecific detectors. To improve
seizure-onset detector or as a seizure-event detector. the performance, investigators developed patient-
The purpose of the seizure-onset detector is to specific event detectors, i.e., detectors that could be
recognize that a seizure has started with the shortest “trained” to the EEG of an individual [22]. These
possible delay but not necessarily with the highest detectors exhibited improved performance because
possible accuracy [16]. In contrast, the purpose of seizure and non-seizure EEGs recorded from an
the seizure-event detector is to identify seizure with individual exhibit less variability [23].
the greatest possible accuracy but not necessarily Seizure-onset and seizure-event detectors are
with the shortest delay [17]. Figure 2 shows often based on analysis of EEG signals. The EEG
categorization of the seizure detection algorithms. is a multichannel recording of the field electrical
Seizure-onset detectors can facilitate the initiation activity generated by enormous numbers of neurons
of delay-sensitive diagnostic, therapeutic, and within different brain regions [24]. The physics
alerting procedures. Within the realm of diagnosis, of EEG generation constrains both the origin and
seizure-onset detectors can be used to quickly characteristics of neural activity visible within the
initiate functional neuroimaging studies designed scalp EEG. In particular, the neurons that contribute
to localize the cerebral origin of a seizure [18]. the most to the scalp EEG are those localized in
Within the realm of therapy, seizure-onset detectors closest vicinity to the scalp surface. In contrast,
can be used to trigger neurostimulators designed the activity of neurons buried within deep brain

Electrods
Analysis of EEG and
ECG signals by a seizure
detector to diagnose of
EEG signals the seizure onset

Monitoring system

Heart EСG signal

Fig. 1. Block diagram of an alerting system.

176 S. Nasehi and H. Pourghassem

Seizure detection algorithms

Seizure -onset detector Seizure -event detector

patient-specific or non -patient specific patient-specific or non-patient specific

based on based on based on based on

analysis of EEG analysis of EEG analysis of EEG analysis of EEG
signals and ECG signals signals and ECG signals

Fig. 2. Categorization of the seizure detection algorithms.

structures is almost not observable. Furthermore, the hypersynchronous neural activity. Seizures of this
cerebrospinal fluid and skull surrounding the brain act type are difficult to detect with a high specificity and
as attenuators that greatly diminish the amplitude of short latency, since the activities, such as eye flutter
high-frequency neural oscillations [25]. An important and muscle contractions, are routinely observed as
consequence of these physical limitations is that the individual partakes in the activities of daily life
certain types of seizures, namely those involving [29].
a small deep region within the brain, practically In order to detect these types of seizures, a
cannot be observed using the scalp EEG. The EEG detector requires information beyond that within
composition is usually classified as having a delta the scalp EEG to ascertain whether a seizure is
component with the dominant frequency f < 4 Hz, taking place or not. The additional information can
a theta component 4 < f < 8 Hz, an alpha component be derived using some other physiological signal
8 < f < 12 Hz, a beta component 12 < f < 30 Hz, and whose dynamics are influenced by the seizure. The
a gamma component f > 30 Hz [26]. second physiological signal and the scalp EEG will
Typically, following the onset of a seizure, the complement each other and improve seizure onset
set of EEG channels, as well as the spectral content detection. The changes in each of these signals
of the rhythmic activity, varies across individuals suggesting the onset of a seizure rarely coincide
[27]. Furthermore, the EEG signature of one during non-seizure states and often coincide at
patient’s seizure may closely resemble the signature the time of an actual seizure. The patient-related
of abnormal non-seizure EEG gathered from the specificity remains essential to the success of this
same patient or from a different patient [28]. For approach since the manner with which the scalp
example, two seizure events within the EEG of a EEG and the secondary signal change during
patient are shown in Fig. 3A, B. Both seizures seizures and non-seizures varies across patients.
appear on the same channel and have the similar For example, seizures resulting in repetitive motor
rhythmic characteristics. Figure 3C, D shows activity may become readily detectable if scalp
the seizure within EEG for another patient. The EEG data are supplemented with accelerometer
seizure event occurs on different channels and has sensor data [30]. For other types of seizures,
dissimilar rhythmic activity in comparison with the especially those originating within or spreading
first patient. So, the detector must be individually to the temporal lobes, seizures are associated with
designed for each patient in order to achieve the electrocardiographic (ECG) changes [31]. The most
best performance, i.e., the classifier must be trained common ECG change associated with seizures is
based on the extracted features from seizure and a heart rate (HR) acceleration (tachycardia) [32].
non-seizure EEG signals for each patient. When Figure 3E shows an example of an electrographic
the epileptic neural network is deep within the seizure, which begins at the 56th sec. It involves a
brain, the scalp EEG may reflect physical sequelae 12-sec-long period of low-amplitude EEG activity
of the seizure, such as repetitive eye-blinks (eye across most EEG channels and, at the same time, the
flutter) or muscle contractions, before reflecting patient’s HR increases.
Seizure Detection Algorithms Based on Analysis of EEG and ECG Signals 177

sec

sec

C D

sec
sec

E F

sec

Fig. 3. Seizure events within the EEG of different patients. A) The seizure starts at sec 1723 and involves the F3-C3 and C3-P3
channels. B) The seizure starts at sec 6210 and involves the F3-C3 and C3-P3 channels. C) The seizure starts at sec 6313 and involves
the F7-T7 and T7-P7 channels. D) The seizure starts at sec 2381 and involves the F7-T7 and T7-P7 channels. E) The seizure starts at
sec 56 and has a low-amplitude EEG activity across most EEG channels for 12 sec; at the same time, the patient’s heart rate accelerates
at sec 56. F) The channels arrayed symmetrically across the scalp EEG.
178 S. Nasehi and H. Pourghassem

STRUCTURE OF SEIZURE essentially the proportion of signal energy contained

DETECTION ALGORITHMS in that scale. This feature is directly calculated
from the wavelet coefficients and not from the
In this section, we present the general architecture amplitudes of the segment decomposition. It serves
of seizure detection algorithms based on the analysis as a measure of arhythmicity, as a sustained elevated
of EEG alone or in concert with ECG signals. The value in one scale indicates a relatively constant
general structure of these algorithms is shown in frequency in the signal. The coefficient of amplitude
Fig. 4; it includes two basic steps, feature extraction variation is defined as the square of the ratio of the
and feature classification. standard deviation to the mean of the peak-to-peak
Classification is the process of assigning amplitudes. The waveform decomposition method is
an observation to one of predefined classes or used to calculate it as well. This feature serves as
categories in a manner that minimizes the error of a measure of variability of the signal amplitude. A
classification. In the seizure detection problem, the low value indicates little variation.
observation is an EEG epoch, and the two classes are There are different algorithms to diagnose the
seizure activity (class one) and non-seizure activity seizure with the above structures, and these can be
(class two). Determining the class label of an categorized into three groups, EEG-based seizure-
observation involves two steps. First, salient features event detection algorithms, EEG-based seizure-
that are most discriminative between instances of onset detection algorithms, and EEG-ECG-based
each class are extracted from the observation and seizure-onset detection algorithms.
assembled into a vector of the features. The success EEG-Based Seizure-Event Detectors. One of the
of a classification task depends strongly on which earliest patient-nonspecific seizure-event detectors
features are extracted from an observation and the was the one developed by Gotman et al. [33] in
classifier used to determine the class label. Features 1982. This algorithm searches EEG channels for
that provide the most diversity between observations the presence of rhythmic activity with a dominant
of two classes and the most resemblance between frequency between 3 and 20 Hz. It uses the frequency
observations of one class improve the detection spectrum analysis to detect periodic discharges. The
performance. Table 1 shows some effective features EEG is divided into 10-sec-long windows moving
that can be extracted from the frequency and time in 2.5-sec increments. The frequency spectrum of
domains of EEG and ECG signals. These features each 10-sec-long epoch is calculated, and a number
include a large number of the single-channel EEG of features, such as the frequency and width of
features, such as power estimations in several the dominant spectral peak, as well as the relative
frequency ranges, as well as several measures of power of the frequency band, are extracted to
the HR variability described in time and frequency detect seizures. The Gotman algorithm successfully
domains of the R-R time series extracted from the detects seizures whose evolution includes sustained
ECG features. The R-R interval is defined as the rhythmic activity with a fundamental frequency
time (sec) between the adjacent R-wave maximum below 20 Hz. It is not successful in detecting seizures
points. The relative scale energy is defined as the where EEG contains a mixture of frequencies or
ratio of the energy in the coefficients of a given those with low-amplitude high-frequency activity.
scale to the energy of the coefficients in all scales. Since the scalp EEGs of individuals with epilepsy
The relative energy of a given wavelet scale is contain pathological, normal, and artifact-induced

Seizure and non-seizure EEG signal or EEG and ECG signals

Pre-processing Divide to L-sec-long epochs

Classification Feature extraction

Fig. 4. General structure of the seizure detection algorithms.

Seizure Detection Algorithms Based on Analysis of EEG and ECG Signals 179

TABLE 1. Effective Features Extracted from the Frequency and Time Domains of EEG and ECG Signals
Frequency/time domains Description
EEG signals
Mean-squared error of estimated AR models (model order 10) [12], [26], [41], [56]
Relative power of spectral band delta (0.1-4 Hz) [42], [29], [34], [53]
Relative power of spectral band theta (4-8 Hz) [42], [29], [37], [53]
Relative power of spectral band alpha (8-15 Hz) [42], [13], [30], [53]
Relative power of spectral band beta (15-30 Hz) [42], [53], [30], [54]
Frequency domain
Relative power of spectral band gamma (30-200 Hz) [55], [19], [30], [54]
Spectral edge frequency [25], [56], [16], [34], [46]
Spectral edge power [11], [17], [53], [57]
Decorrelation time [27]

First statistical moment of EEG amplitudes (mean) [10], [23], [41], [58]
Second statistical moment of EEG amplitudes (variance) [10], [23], [41], [58]
Time domain Third statistical moment of EEG amplitudes (skewness) [10], [23], [41], [58]
Fourth statistical moment of EEG amplitudes (kurtosis) [10], [23], [41], [58]
Long-term energy [15] [24], [44], [59]
Time and frequency
Energy of the wavelet coefficients [11], [25], [43], [60], [37], [26], [42], [56], [63]
domain
ECG features
Very low frequency (VLF): <0.04 Hz [33], [39], [20]
Frequency domain Low frequency (LF): 0.04-0.15 Hz [33], [39], [20]

Mean R-R intervals (msec) [61], [20], [28]

Variance of R-R intervals (msec) [33], [46], [28]
Maximum R-R interval (msec) [33], [46], [28]
Minimum R-R interval (msec) [33], [46], [28]
Time domain Mean heart rate (min–1) [33], [46], [20]
Variance of the heart rate (min–1) [62]
Maximum heart rate (min–1) [62], [64]
Minimum heart rate (min–1) [62]
Approximate entropy (ApEn) describing the complexity and irregularity of the RR intervals

bursts of rhythmic activity, a significant fraction of with a seizure taking place on that channel. The
detections produced by the Gotman algorithm are not thresholds for some of the neural network rules are
associated with seizures [34]. So, investigators have determined using both archetypal seizures from
developed seizure-event detectors that utilize more individuals with epilepsy and background EEG
sophisticated signal processing to characterize the from individuals without epilepsy.
rhythmicity associated with seizures. Liu’s algorithm Hassanpour et al. [37] investigated a time-
[35] also relies on the periodicity as the dominant frequency domain (TFD)-based seizure-event
characteristic of seizures in the EEG signals. The detection algorithm. The EEG signal was
degree of periodicity in the autocorrelation function segmented into 30-sec-long epochs. A singular
of 30-sec-long epochs of EEG data is scored and value decomposition (SVD) was performed on the
used to classify the epoch as seizure or non-seizure. TFD representation of each epoch. To discriminate
Wilson’s algorithm [36] decomposes 2-sec-long between seizure and non-seizure activities in each
EEG epochs from each input channel into time- EEG epoch using the TFD, this method uses two left
frequency “atoms” using the matching pursuit and two right singular values (SVs). The left and
algorithm. Wilson then employs hand-coded and right SVs correspond to the time- and frequency-
neural network rules to determine whether features domain components of the signal. The extracted
derived from the “atoms” of a channel are consistent features through the histograms of the four SVs are
180 S. Nasehi and H. Pourghassem

organized into a feature vector and fed into a trained Instead of extracting and then classifying feature
neural network to classify each feature vector as vectors from one channel to another, Meier
seizure or non-seizure. extracted a single feature vector that includes the
EEG-Based Seizure-Onset Detectors. Saab et al. average (across-channels) of signal properties, such
[38] designed an automatic seizure onset detection as the number of zero crossings, wavelet coefficient
that was used on-line within a long-term monitoring power, and cross-correlation.
facility. It employs a Bayesian formulation to output Celka’s algorithm [41] is a time-domain method
a variable based on the probability that an EEG employing patient-specific pre-processing. The pre-
section contains seizure activity. Saab’s algorithm processing involves estimating an autoregressive
uses features derived from a wavelet decomposition moving average model of the pre-recorded normal
of each EEG channel to estimate the probability of EEG. The corresponding inverse is applied to
a seizure. Whenever the probability exceeds a user- the signal being analyzed, leaving only seizure
defined threshold for a given period of time, the components and Gaussian white noise. A SVD-
algorithm declares the onset of a seizure. Qu et al. based algorithm is then used to extract seizure
developed the first patient-specific seizure-onset features from noise.
detection algorithm [39]. Qu’s algorithm relies on a Shoeb presented a patient-specific seizure-onset
nearest-neighbor classifier to assign a list of features detection algorithm [42]. It extracts eight features
to the seizure or non-seizure classes. The classifier from a 0-25 Hz frequency band by means of a
is trained on seizure and non-seizure feature vectors 3 Hz bandwidth filter, and a support vector machine
derived from the available EEG channels and (SVM) classifier is used to classify the feature
declares a seizure if the set of positively classified vectors. The detector passes L-sec-long epochs
channels matches half of those chosen by an expert. from each N EEG channels through a filter bank. In
Meier et al. [40] grouped seizures in a database turn, the filter bank computes M features for each
into six categories based on the frequency of the channel, which correspond to the energies within
dominant rhythm that appears following the seizure M frequency bands. The M extracted features from
onset. He then trained a set of support-vector each of the N channels are then concatenated to form
machines, one for each seizure type, to determine an M×N element vector that automatically captures
whether an extracted feature vector from an EEG the spectral and spatial relations between channels.
epoch is consistent with one of the seizure types. Finally, the feature vector is assigned to the

Fig. 5. Structure of Barry’s algorithm [46].

Seizure Detection Algorithms Based on Analysis of EEG and ECG Signals 181

seizure or non-seizure class using a two-class SVM spectral peak, power ratio, bandwidth of the
classifier. Since seizure and non-seizure activities dominant spectral peak, nonlinear energy, spectral
are generally stereotypical for a patient and highly entropy, and line length) from EEG signals and six
variable across patients, the SVM is trained on features (mean R-R interval, standard deviation of
discriminating seizure and non-seizure vectors from these intervals, mean R-R interval spectral entropy,
a single patient. The training non-seizure vectors are mean change in the R-R interval, interval coefficient
extracted from H hours of the continuously recorded of variation, and interval power spectral density)
scalp EEG. The seizure vectors are derived from the from ECG signals. The R-R interval is defined as
first S sec following the onset of K training seizures. the time (sec) between adjacent R-wave maximum
EEG and ECG-Based Seizure-Onset Detectors. points described by Benitez [47]. Figure 5 shows
Several authors have published data on the the general structure of the Barry’s algorithm.
relationship of changes in the ECG signal with adult Nasehi et al. developed a seizure onset detection
epileptic onset and its utility for epileptic seizure algorithm based on the analysis of EEG and
detection. As was mentioned, seizures involve the ECG signals to detect seizure onsets that are not
hyperactivity and hypersynchrony of a population of associated with rhythmic EEG activity [48]. In
neurons. At the level of the scalp EEG, this coherent this algorithm, L-sec-long epochs from seizure and
neural firing gives rise to rhythmic activity with a non-seizure EEG signals are decomposed by Gabor
dominant frequency between 0 and 25 Hz. However, functions and represented in the spatial, spectral,
when the underlying neural hypersynchrony and temporal domains. Then, five features, such as
involves a neural network deep within the brain, number of zero coefficients, smallest and largest
the earliest scalp EEG changes may not reflect the coefficients, and mean and standard deviation
above neuronal hypersynchrony. In order to detect of the coefficients, are extracted from each sub-
seizures, additional information can be obtained representation. Synchronously, four features, such
from a second physiological signal, such as the ECG as the mean HR, instantaneous HR, power ratio,
signal. The most common ECG change associated and spectral entropy, are extracted from L-sec-long
with seizures is an HR increase. ECG epochs. Finally, a probabilistic neural network
To design a seizure-onset detector based on the classifier is employed to train on the extracted
analysis of EEG and ECG signals, the features are features from seizure and non-seizure EEG-ECG
extracted from the seizure and non-seizure EEG signals of each patient for determination of optimal
signals of the patient and simultaneously combined nonlinear decision boundaries.
with the extracted features from the ECG signals. The
spectral and spatial features can be obtained from
the energy spectrum of each EEG channel. Also, the EEG AND ECG DATASET AND
HR mean and the instantaneous HR can be extracted EVALUATION MEASURES.
from ECG signals. For example, Quint et al. [43]
studied changes in the ECG during epileptic seizure The seizure detection algorithms require a
onset in adults and concluded that characteristic training dataset of EEG or ECG signals. Since these
changes in the mean heart period are frequently, databases were constructed to empirically evaluate
if not always, present upon detection of seizure recognition algorithms in certain domains, we first
onset in adult EEG. Zijlmans et al. [44] attempted review the characteristics of these databases and
to rigorously document cardiac behavior during their applicability to seizure detection. Then, the
epileptic seizure onset in adults. These authors evaluation measures are introduced.
found that there was an increase in the HR of at least
10 min –1 in 73% of seizures (93% of patients) around
the point of seizure onset from a point 30 sec prior Dataset
to the moment of clinical or electrographic seizure
onset. Kerem et al. [45] used the R-R interval time SMC Dataset. This dataset is collected from
series to forecast epileptic seizure in adults using the Epilepsy Telemetry Unit at the Montreal
successive HR timing intervals in an unsupervised Neurological Institute and Hospital, using the
fuzzy clustering algorithm. Barry et al. [46] presented Stellate Harmonic system for EEG monitoring
a seizure onset detection algorithm based on EEG (Stellate, Montreal, Canada). Data are sampled at
and ECG signals. It extracted six features (dominant 200 sec –1 after filtering between 0.5 and 70 Hz,
182 S. Nasehi and H. Pourghassem

and bipolar electrode montages of either 24 or refers to the onset of scalp EEG changes associated
32 channels are used in the analysis. This dataset with a seizure. The clinical onset of a seizure refers
consists of EEG recording from 28 patients that to the onset of its physical or cognitive symptoms.
contained 652 h of EEGs recorded and 126 seizures. In scalp EEGs, the electrographic onset of a seizure
Further information about this data is available in may or may not precede its clinical onset. The
[49]. latency refers to a delay between the electrographic
CHB Dataset. This dataset consists of continuous onset and detector recognition of seizure activity.
scalp EEG recording from 23 pediatric patients Sensitivity. The sensitivity refers to the proportion
undergoing medication withdrawal for epilepsy (percentage) of test seizures identified by a detector.
surgery evaluation at the Children’s Hospital A high sensitivity increases the capability of a
(Boston, USA). The EEG was sampled at 256 sec –1 detector to recognize seizures in order to initiate
and recorded using an 18-channel 10-20 bipolar timely therapy procedures.
montage. Overall, this 23-patient dataset contained Specificity. The specificity refers to the number
844 h of continuously recorded EEGs and 163 of times, over the course of an hour, that a detector
seizures. The scalp EEG dataset is segmented into declares the onset of seizure activity in the absence
records. Typically, a record is 1 h long. Records of an actual seizure.
that do not contain a seizure are called non-seizure FDR. The false detection rate is defined as
records, and those that contain one or more seizures the percentage of non-seizure epochs incorrectly
are called seizure records. Further information identified as seizure epochs.
about this data is available in [50]. GDR. The seizure sensitivity, or good detection
KCH Dataset. A dataset of 12 records from 10 term rate (GDR), is defined as the percentage of
neonates containing 633 labeled seizure events with electrographic seizure events as defined by an expert
a mean seizure duration of 4.60 min were recorded in EEG-ECGs correctly identified by the detector.
and analyzed. The records had a mean duration of
12.84 h. Each record contained 7 to 12 channels of
EEG and one channel of simultaneously acquired PERFORMANCE EVALUATION OF
ECG. Ten records sampled at 256 sec –1 were made ALGORITHMS AND DISCUSSION
in the neonatal intensive care units of the Unified
Maternity Hospitals in Cork (Ireland) using the Table 2 illustrates the comparison between the
Viasys NicOne video-EEG system. The remaining best seizure-detection algorithms that have been
recording, sampled at 200 sec –1, was recorded at proposed in the literature. Gotman et al. [33]
Kings College Hospital, London (Great Britain) on reported a GDR (or sensitivity) of 71% with 1.7
a Tele factor Beehive video-EEG system. A total of false detections per hour. The results of the Gotman
154.1 h of EEGs and ECGs were analyzed. Further method were validated in a subsequent paper [34]
information abut this data is available in [51]. by analyzing a separate dataset containing 281 h of
Freiburg Dataset. This dataset consists of EEG data from 54 patients (again in three medical
continuous scalp EEG recording from 57 pediatric centers). The sensitivity of this set was 69% with
patients undergoing medication withdrawal for a specificity of 2.3 per hour. Gotman’s algorithm
epilepsy surgery evaluation at University Hospital was not successful in detecting seizures with EEG
Freiburg (Germany). The EEG was sampled at 256 containing a mixture of frequencies or those with
sec –1 and recorded using a 6-channel 10-20 bipolar low-amplitude high-frequency activity. Liu et al.
montage. Overall, this 57-patient dataset contained [36] reported a sensitivity of 84% and a FDR of
1400 h of continuously recorded EEGs and 91 98% for their proposed method. Wilson et al. [36]
seizures. Further information about this data is reported that their proposed algorithm detected
available in [52]. 76% of 672 seizures gathered from 426 individuals
with a sensitivity of 0.11 false detections per hour.
It demonstrated poor specificity when the scalp
Evaluation Measures EEG signal of patients was abnormal (non-seizure
rhythmic activity). Hassanpour’s algorithm [37]
Five metrics are usually used to characterize the was evaluated on 8 patients. It showed an average
performance of the seizure-detection algorithm. sensitivity of 92.5% and a FDR of 3.7%.
Latency. The electrographic onset of a seizure Saab’s algorithm [37] was evaluated on SMC
Seizure Detection Algorithms Based on Analysis of EEG and ECG Signals 183

TABLE 2. Comparison of Performance of the Best Seizure Detection Algorithms

Algorithms Sensitivity (%) Specificity FDR GDR Latency Dataset description
(false per h) (%) (%) (sec)
Gotman [34] 69 2.3 – – – 281 h from 54 patients
Liu [35] 84 – – 98 – 281 h from 54 patients
Wilson [36] 76 0.11 – – – 672 seizures from 426 patients
Hassanpour [37] 92 – 3.7 – – 64 seizures from 8 patients
Saab [38] 78 0.86 – – 9.8 SMC dataset
Qu [39] 100 0.03 – – 9.35 29 h and 47 seizures from 12
patients
Meier [40] 96 0.45 – – 1.6 Freiburg dataset
Celka [41] 93 – 4 – – 53 seizures from 4 patients
Shoeb [42] 96 0.07 – – 4.6 CHB dataset
Kerem [45] 86 – – – – 95 seizures from 8 patients
Barry [46] 74 – 13.18 97.52 – KCH dataset
Nasehi [48] 79 – 12.47 98.25 4.7 KCH dataset

Footnote. FDR and GDR are, respectively, false and good detection rates.

dataset. This algorithm detected 78% of seizures used to record the training seizures. Seizures whose
with a median detection latency of 9.8 sec and a onsets lack the development of rhythmic activity
specificity of 0.86 false detections per hour. Saab and instead reflect physical sequelae of the seizure
reported that missed seizures included those with (such as eye flutter) do not fall within the defined
onsets characterized by focal activity, mixed categories. Consequently, such seizures will be
frequencies, or short duration, and those false detected later or not at all. Celka’s algorithm [40]
detections were mainly caused by short bursts of obtained a sensitivity of 93% and a false detection
rhythmic activity, rapid eye blinking, and chewing. rate (FDR) of 4% for four neonatal subjects. In
Qu’s algorithm [38] was evaluated on 29.7 h and the reported results by Celka, channels known
47 seizures from 12 patients. This method detected to contain seizures were chosen for processing.
100% of seizures with an average delay of 9.35 Although this does not bias the results on a per-
sec and a specificity of 0.03 false detections per channel basis, a real-time seizure detection system
hour. The non-seizure EEG signals that Qu used to would require processing and polling of all channels
calculate the FDR of his algorithm were formed by in parallel, as the spatial location of the seizure is
concatenating segments of EEG extracted at regular a priori unknown. Shoeb’s algorithm [41] is trained
intervals from several days of the dataset. When on the CHB dataset. The algorithm detected 96%
compared to Saab’s work, Qu’s work illustrates that of 163 test seizures with a median detection delay
a patient-specific approach can result in improved of 4.6 sec and specificity of 0.07 false per hour.
sensitivity and specificity, but not necessarily in The latency, however, was large for some patients,
an improvement in the detection latency. Meier’s which can arise from great similarity of seizure and
algorithm [39] is evaluated on the Freiburg dataset. non-seizure signals.
It detected 96% of the test seizures with an average Zijlmans et al. [43] found that there was an increase
detection delay of 1.6 sec and specificity of 0.45 in the HR of at least 10 min –1 in 73% of seizures
false detections per hour. Meier’s approach depends (93% of patients) around the point of seizure onset
on the test seizure being a member of one of the from a point 30 sec prior to the moment of clinical
six defined categories, as well as it being recorded or electrographic seizure onset. These results were
using the same number and position of channels estimated for 281 seizures in 81 epileptic patients.
184 S. Nasehi and H. Pourghassem

The large size of this dataset provides high weight epileptic seizure onset detection and treatment,” Ph.D.
and statistical significance to the results presented. dissertation, Dept. Electrical Eng., MIT, USA (2009).
4. B. S. Chang and D. H. Lowenstein, “Epilepsy,” Eng. J.
Kerem et al. [44] reported a prediction sensitivity of
Medicine, 13, No. 349, 1257-1266 (2005).
86% for eight epileptic patients. Barry’s algorithm 5. S. Nasehi and H. Pourghassem, “Automatic prediction
[45] was evaluated on the KCH dataset. It could of epileptic seizure using kernel fisher discriminant
recognize 617 of 633 expert-labeled seizures with classifiers,” in: IEEE International Conference
an FDR of 13.18%. Nasehi’s algorithm [47] was on Intelligent Computation and Bio-Medical
evaluated on the KCH dataset. It could reach a mean Instrumentation (ICBMI2011), Wuhan, China (2011),
pp. 200-203.
GDR of 98.25% with an FDR of 12.47%. It also
6. N. Verma, A. Shoeb, J. Bohorquez, et al., “A micro-
obtained a classification sensitivity of 79.66% and a power EEG acquisition SoC with integrated feature
mean latency of 4.7 sec. extraction processor for a chronic seizure detection
system,” IEEE J. Solid-State Circuits, 45, No. 4, 804-
816 (2010).
CONCLUSION 7. M. C. Alexander, T. S. Felice, H. E. Boto, and M. W.
Brett, “Automated seizure onset detection for accurate
onset time determination in intracranial EEG,” Clin.
This paper proposes an overview of the seizure Neurophysiol., 119, 2687-2696 (2008).
detection algorithms based on the analysis of EEG 8. E. Ben-Menachem, “Vagus nerve stimulation for the
and ECG signals. The basic structure of these treatment of epilepsy,” Lancet Neurol., 1, No. 8, 477-
algorithms is constructed in two main components, 482 (2002).
feature extraction and feature classification. The 9. M. DeGiorgio, A. Shewmon, D. Murray, and T.
Whitehurst, “Pilot study of trigeminal nerve stimulation
extracted features from EEG signals can be defined
(TNS) for epilepsy: a proof-of concept trial,” Epilepsia,
based on the power spectrum or representation forms 47, No. 7, 1213-1215 (2006).
of the signal that are provided by using wavelet 10. S. Nasehi and H. Pourghassem, “EEG signal compression
transforms or Gabor filter banks of L-sec-long epochs based on adaptive arithmetic coding and first-order
of the EEG signals. Also, the HR acceleration can markov model for an mbulatory monitoring system,” in:
be considered as a feature extracted through ECG IEEE 3rd International Conference on Signal Processing
Systems (ICSPS), Yantai, China, vol. 1, (2011), pp. 185-
signals. All of these features and classifiers, such
189.
as SVM, artificial neural networks, and statistical 11. D. Casson, S. Yates, J. Smith, et al., “Wearable
classifiers, detect the seizure and non-seizure EEG Electroencephalography,” IEEE Eng. Med. Biol. Mag.,
signals. Five standard measures (latency, sensitivity, 29, No. 3, 44 (2010).
specificity, FDR, and GDR) are used to evaluate and 12. W. Theodore and R. Fisher, “Brain stimulation for
compare the seizure detection algorithms. For the epilepsy,” Lancet Neurol., 3, No. 2, 111-118 (2004).
13. S. Rothman and X. F. Yang, “Local cooling: A therapy
seizure-onset detection algorithms, the ability of
for intractable neocortical epilepsy,” Epilepsy Curr., 3,
the proposed algorithm to obtain smaller values of No. 5, 153-156 (2003).
the latency is very important. For the seizure-event 14. F. Zhang, L. P. Wang, and M. Brauner, “Multimodel fast
detection algorithms, the sensitivity measure is an optical interrogation of neural circuity,” Nature, 446,
important parameter in the evaluation of the ability 6633-6639 (2007).
of the presented algorithm. 15. S. Nasehi and H. Pourghassem, “Patient-specific
epileptic seizure onset detection algorithm based on
spectral features and IPSONN classifier,” in: IEEE 3rd
International Conference on Signal Processing Systems
REFERENCES (ICSPS), Yantai, China, Vol. 1 (2011), pp. 217-222.
16. H. Shoeb, J. Edwards, B. Connolly, et al., “Patient-
1. C. E. Begley, M. Famulari, J. F. Annegers, et al., “The specific seizure onset detection,” Epilepsy Behav., 5,
cost of epilepsy in the United States: An estimate from No. 4, 483-498 (2004).
population-based and clinical survey,” Epilepsia, 41, 17. G. I. Varghese, M. J. Purcaro, and J. E. Motelaw,
No. 3, 342-351 (2007). “Clinical use of ictal SPECT in secondarily generalized
2. S. Nasehi and H. Pourghassem, “Epileptic seizure tonic-clonic seizures,” Brain, 132, 2102-2113 (2007).
onset detection algorithm using dynamic cascade feed- 18. Z. I. Nanobashvili, T. R. Chachua, I. G. Bilanishvili,
forward neural networks,” in: IEEE International et al., “Peculiarities of the effects of stimulation of
Conference on Intelligent Computation and Bio-Medical emotiogenic central structures under conditions of a
Instrumentation (ICBMI2011), Wuhan, China (2011), kindling model of epilepsy,” Neurophysiology, 43,
pp. 196-195. No. 4, 337-344 (2011).
3. H. Shoeb, “Application of machine learning to 19. S. Nasehi and H. Pourghassem, “A new feature
Seizure Detection Algorithms Based on Analysis of EEG and ECG Signals 185

dimensionally reduction approach based on general tensor for the newborn EEG,” Electroencephalogr. Clin.
discriminant analysis in EEG signal classification,” Neurophysiol., 103, No. 3, 363-369 (1997).
in: IEEE International Conference on Intelligent 35. Liu, J. S. Hahn, G. P. Heldt, and R. W. Coen, “Detection
Computation and Bio-Medical Instrumentation of neonatal seizures through computerized EEG
(ICBMI2011), Wuhan, China (2011), pp. 188-191. analysis,” Electroencephalogr. Clin. Neurophysiol., 82,
20. Kharbouch, A. Shoeb, J. Guttag, and S. Cash, “An 30-37 (1992).
algorithm for seizure onset detection using intracranial 36. S. B. Wilson, “Algorithm architectures for patient
EEG,” Epilepsy Behav., 22, Suppl. 1, S29-S35 (2011). dependent seizure detection,” Clin. Neurophysiol., 117,
21. J. Gotman, “Neurophysiology, seizure detection No. 6, 1204-1216 (2006).
and prediction,” Encycl. Basic Epilepsy Res., 37. H. Hassanpour, M. Mesbah, and B. Boashash, “Time-
Neurophysiology, 3, 1606-1613 (2009). frequency feature extraction of newborn EEG seizure
22. P. Williams, A. White, D. Ferraro, et al., “The use of using SVD-based techniques,” EURASIP J. Appl. Signal
radio telemetry to evaluate electrographic seizures Process., 16, 2544-2554 (2004).
in rats with kainate-induced epilepsy,” J. Neurosci. 38. M. E. Saab and J. Gotman, “A system to detect the onset
Methods, 155, No. 1, 39-48 (2006). of epileptic seizures in scalp EEG,” Clin. Neurophysiol.,
23. Litt and J. Echauz, “Prediction of epileptic seizures,” 116, No. 2, 427-442 (2005).
Lancet Neurol., 1, No. 1, 22-30 (2002). 39. H. Qu and J. Gotman, “Improvement in seizure detection
24. J. Overweg, “Withdrawal of antiepileptic drugs (AEDs) performance by automatic adaptation to the EEG of each
in seizure-free patients, risk factors for relapse with patient,”  Electroencephalogr. Clin. Neurophysiol., 86,
special attention for the EEG,” Seizure, 4, No. 1, 19-36 No. 2, 79-87 (1993).
(1995). 40. R. Meier, U. Häussler, A. Aertsen, et al., “Short-term
25. Aarabi, R. Grebe, and F. Wallois, “A multistage changes in bilateral hippocampal coherence precede
knowledge-based system for EEG seizure detection in epileptiform events,” NeuroImage, 38, No. 1, 138-149
newborn infants,” Clin. Neurophysiol., 118, No. 12, (2007).
2781-2797 (2007). 41. P. Celka and P. Colditz, “A computer-aided detection of
26. Alkan, E. Koklukaya, and A. Subasi, “Automatic seizure EEG seizures in infants: a singular-spectrum approach
detection in EEG using logistic regression and artificial and performance comparison,” IEEE Trans. Biomed.
neural network,” J. Neurosci. Methods, 148, No. 2, 167- Eng., 49, No. 5, 455-462 (2002).
176 (2005). 42. Shoeb, A. Kharbouch, J. Soegaard, et al., “A machine-
27. Wang, W. Yuan, J. Liu, et al., “A novel pattern learning algorithm for detecting seizure termination in
recognition algorithm: Combining ART network with scalp EEG,” Epilepsy Behav., 22, Suppl. 1, S36-S43
SVM to reconstruct a multi-class classifier,” Comput. (2011).
Math. Appl., 57, Nos. 11/12, 1908-1914 (2009). 43. S. R. Quint, J. A. Messenheimer, M. B. Tennison, and H.
28. S. Nasehi and H. Pourghassem, “A novel effective T. Nagle, “Assessing autonomic activity from the EKG
feature selection algorithm based on S-PCA and wavelet related to seizure onset detection and localization,” in:
transform features in EEG signal classification,” in: 2nd Ann. IEEE Symp. on Computer-Based Med. Syst.,
IEEE 3rd International Conference on Communication Minneapolis, pp. 2-9 (1989).
Software and Networks (ICCSN), Xi’an, China (2011), 44. M. Zijlmans, D. Flanagan, and J. Gotman, “Heart
pp. 114-117. rate changes and ECG abnormalities during epileptic
29. P. C. Petrantonakis and L. J. Hadjileontiadis, “Emotion seizures: prevalence and definition of an objective
recognition from EEG using higher order crossings,” clinical sign,” Epilepsia, 43, 847-854 (2002).
IEEE Trans. Inform. Technol. Biomed., 14, No. 2, 186- 45. H. Kerem and A. B. Geva, “Forecasting epilepsy from
197 (2010). the heart rate signal,” Med. Biol. Eng. Comput., 43, 230-
30. S. B. Wilson, M. L. Scheuer, R. G. Emerson, and A. 239 (2005).
J. Gabor, “Seizure detection: evaluation of the Reveal 46. R. G. Barry, G. B. Boylan, R.B. Reilly, et al., “Neonatal
algorithm,” Clin. Neurophysiol., 115, No. 10, 2280-2291 seizure detection using a combination of EEG and ECG
(2004). analysis,” Clin. Neurophysiol., 117, Suppl. 1, 200
31. S. Rothman and X. F. Yang, “Local cooling: A therapy (2006).
for intractable neocortical epilepsy,” Epilepsy Curr., 3, 47. Benitez, P. A. Gaydecki, A. Zaidi, and A. P. Fitzpatrick,
No. 5, 153-156 (2003). “The use of the Hilbert transform in ECG signal
32. M. Valderrama, C. Alvarado, S. Nikolopoulos, et al., analysis,” Comput. Biol. Med., 31, No. 5, 399-406
“Identifying an increased risk of epileptic seizures using (2001).
a multi-feature EEG-ECG classification,” Biomed. Signal 48. S. Nasehi and H. Pourghassem, “A novel epileptic
Process. Control, (2011). Available online 2 October seizure detection algorithm based on analysis of EEG
2011 (http: dx.doi.org/10.1016/jobspc.2011.05.005). and ECG signals using probabilistic neural network,”
33. J. Gotman, “Automatic seizure detection: improvements Aust. J. Basic Appl. Sci., 5, No. 12, 308-315 (2011).
and evaluation,” Electroencephalogr. Clin. 49. SMC Dataset, http://www.csie.ntu.edu.tw/cjlin/liblinear
Neurophysiol., 76, No. 4, 317-324 (1990). 50. CHB Dataset, http://www.physionet.org/pn6/chbmit
34. J. Gotman, D. Flanagan, B. Rosenblatt, et al., 51. KCH Dataset, http://www.vis.caltech.edu/~rodri/data.
“Evaluation of an automatic seizure detection method htm
186 S. Nasehi and H. Pourghassem

52. Freiburg dataset, http://epilepsy.uni-freiburg.de/data­b ase 60. J. Altenburg, R. J. Vermeulen, R. L. M. Strijers, et

53. P. de Chazal, C. Heneghan, E. Sheridan, et al., “Automated al., “Seizure detection in the neonatal EEG with
processing of the single-lead electrocardiogram for the synchronization likelihood,” Clin. Neurophysiol., 114,
detection of obstructive sleep apnoea,” IEEE Trans. 50-55 (2003).
Biomed. Eng., 50, No. 6, 686-696 (2003). 61. S. Nasehi and H. Pourghassem, “Real-time seizure
54. S. Nasehi and H. Pourghassem, “Online mental detection based on EEG and ECG fused features using
task classification based on DWT-PCA features and gabor functions,” in: IEEE International Conference on
probabilistic neural network,” Int. J. Imaging Robotics, Intelligent Computation and Bio-Medical Instrumen­
7, No. 1, 110-118 (2012). tation (ICBMI2011), Wuhan, China (2011), pp. 204-
55. M. LeVanQuyen, J. Martinerie, V. Navarro, et al., 207.
“Anticipation of epileptic seizures from standard EEG 62. N. B. Karayiannis, G. Tao, J. Frost, et al., “Automated
recordings,” Lancet, 357, 183-188 (2001). detection of videotaped neonatal seizures based on
56. L. D. Iasemidis, D. S. Shiau, W. Chaovalitwongse, et al., motion segmentation methods,” Clin. Neurophysiol.,
“Adaptive epileptic seizure prediction system,” IEEE 117, 1585-1594 (2006).
Trans. Biomed. Eng., 50, No. 5, 616-627 (2003). 63. S. Nasehi, H. Pourghassem, and A. Etesami “Online
57. R. Acharya, U. A. Kumar, P. S. Bhat, et al., “Classification epilepsy diagnosis based on analysis of EEG signals by
of cardiac abnormalities using heart rate signals,” Med. hybrid adaptive filtering and higher-order crossings,”
Biol. Eng. Comput., 42, 288-293 (2004). in: IEEE International conference on Intelligent
58. S. Nasehi and H. Pourghassem, “Mental task Computation and Bio-Medical Instrumentation
classification based on HMM and BPNN,” in: IEEE 3rd (ICBMI2011), Wuhan, China (2011), pp. 192-197.
International Conference on Signal Processing Systems 64. P. de Chazal, M. O’Dwyer, and R. B. Reilly, “Automatic
(ICSPS), Yantai, China, Vol. 2 (2011), pp. 52-56. classification of heartbeats using ECG morphology and
59. S. Faul, G. Boylan, S. Connolly, et al., “An evaluation heartbeat interval features,” IEEE Trans. Biomed. Eng.,
of automated neonatal seizure detection methods,” Clin. 51, No. 7, 1196-1206 (2004).
Neurophysiol., 116, 1533-1541 (2005).