Professional Documents
Culture Documents
information that allows for the delineation of the presence of interictal HFOs upon initial review.
epileptogenic zone with the highest possible Second, a group of three consecutive patients with
precision. intractable epilepsy was considered in the current
The duration of HFOs is 10-100 ms, with an study. Finally, channels containing different
average occurrence rate of 12 ± 17 Imin for Ripples background power levels were taken into
and 7 ± 18 I min for Fast Ripples [16]. Therefore, consideration.
associated with large amounts of EEG data is an rules where the decision is based on successive tests
extremely tedious process and requires a great deal associated with features, values and classes
of mental concentration and focus. Fatigue or organized in a tree structure based on information
distraction during visual marking can easily lead to theory. In our case, the EEG data is composed of
errors (detection of false positives and false two types or classes (background activity and HFO
negatives) [16,17]. Visual scoring of HFOs is activity). The implementation of the decision tree is
highly time-consuming: processing 10 channels of done using the following six features (Fl to F6)
10-min recordings could take up to 10 h of hard computed from the filtered signal (x) in the 80-500
work even for an experienced reviewer. Besides, Hz band and the time-frequency map (C) in the 80-
visual methods apply subjective interpretations. By 500 Hz band.
contrast, the automatic detection of HFOs relies on • F1 is the mean + standard deviation defined as
the application of a set of mathematical rules follows:
implemented in a software environment that aimes
Fl = x + G (x) (1)
to detect and identify these events of interest not
• F2 is the difference between maximum and
only faster but also more objectively. Broadly
minimum coefficients of time frequency map
speaking, automatic detection of HFOs may have
computed using the CMOR2-1.114 wavelet
several important applications and implications,
family.
such as: i) Propelling the clinical use of HFOs as a
reliable indicator of epileptogenic zone [18, 19] ii) F2 = max I C(t, f) - min I C (t, f) (2)
t,f t,f
The development of new clinical diagnostic
procedures and novel therapeutic interventions for
• F3 is the mean + standard deviation of TEO
patients with pharmaco-resistant epilepsy, and iii) (Teager Energy Operator) :
Providing a robust tool with which neurologists and
researchers can systematically analyze large
F3 = TEO + G (TEO) (3)
I abS(diff C + s�gn(x» ) /N
sensitivity, False Discovery Rate (FOR) and Area
(4)
Under the Curve (AUC). The organization of the F4 =
Entropy Gain [21] which is defined as follows: baselines segments selected manually. � is the input
training parameter to be optimized. Learning and
automatic recognition flowcharts of HFOs by the
" IXai=)
- HeX)
. ( 7)
Gam ex,aj ) -
method based on the decision tree is shown in
1 Ix - (Xai=J
valeurs (ai) --
- L.. v E
H
Figurel .
Where Xaj=v is the set of examples, on which a IV. VISUAL MARKING OF HFO EVENTS AND
PERFORMANCE METRICS
considered feature aj takes the value v and the
notation, IXI indicates the cardinality of the set X. Visual marking of HFOs was performed by two
Given a set X of examples of which a portion p+ is reviewers trained in electrophysiology and HFO
positive and portion p_ is negative. (with p+ + p_ =
analysis. Each event was marked as a relevant
1). The entropy HeX) is defined as follows: HFO; if it is seen as an oscillation with at least
three consecutive cycles in the frequency band 80-
500 Hz, and could be clearly distinguished from the
average of the background in the filtered signal (80-
Where 0 <;:: HeX) <;::1. 500Hz). In addition, to be confirmed, each event is
visually verified also in the unfiltered EEG signal.
Once the learning phase is completed, the automatic
Every event detected by the two reviewers was
recognition of HFOs using the optimized decision
considered as a relevant HFO burst. In contrast, the
tree (Tmax) can be done.
remaining segments were considered "background
Based on the six features FI-F6 and Tmax, each
activity". Note that although these segments may
associated window positioned at every time sample
contain various other oscillatory or transient events,
is tested, to verify whether it matches the HFO class
we use the term "background activity" here as an
or the background class. First, a sliding window of
indicator for the absence of an HFO event.
length 50 ms with predefined overlap of 1 sample
The performance measures we used are sensitivity,
scans all the samples of the original EEG signal.
False Discovery Rate (FOR) and Area Under a
The length of window (50 ms) is approximately
Curve (AUC), which are respectively defined as:
equal to the average duration of all HFOs events
used in the learning phase. Once scanning the entire 0p
. . . (10)
signal is completed, the samples corresponding to SenSltivlty = 100 -
pos
the HFO class are set to 1 and the other segments
FP (II)
are set to O. Next, the segments with value 1 are FOR= 100 --
TP + FP
delimited in time. Each portion of the filtered and
rectified signal with level 1 above the threshold (12)
ei\ectified) is marked as a probable HFO if it has at
least six peaks (equivalent to 3 cycles). The
threshold C\ectified is defined as follows:
Temporal filtering and time frequency filtering in 80-500 Hz Temnoral filterin!:! and time freauencv filterin!:! in 80-500 Hz
*
100 HFO events,100 Background events Sliding window (SOms) with I sample overlap
*
Feature Calculation Feature Calculation
t
Thresholding by minimal cycle number (3 cycles)
Learning
t
Automatic recognition
Fig. I. Learning and automatic recognition of HFO using the proposed descion tree framework.
IEEE IPAS' 14: INTERNATIONAL IMAGE PROCESSING APPLICATIONS AND SYSTEMS CONFERENCE 2014 4
The HFO detection algorithm produces a list of previously implemented, namely: RMS [3 ], CMOR
probable HFOs with their location and duration. [22], Bumps [23], Matching Pursuit [24] and HHT
Probable HFOs may correspond to positives or [25].
negatives. Positives: are the HFOs segments The RMS detector proposed by Staba [3 ] is
visually identified. Negatives: are different based on linear finite impulse response (FIR) filter
background segments that do not contain HFOs and the moving average of the root mean square
(We use Pas to indicate the number of positives, (RMS) feature. The description regarding the
and Dpas (Detected Positives) as the number of implementation and the varying input parameters
positives which overlap with at least one probable for this method are provided in [3 ]. The second
algorithm for HFOs detection is based on the
HFO. TP (True Positives): is the number of
complex Morlet wavelet (CMOR). The description
probable HFOs which overlap with at least one
of this method is provided in [22]. The bumps
positive event. FP (False Positives): is the number
detector was presented by Doshi (2011). This
of probable HFOs which do not overlap with any
method is based on bumps modeling technique. The
positive event. AUC is the area under the ROC
description of the implementation and various input
curve.
parameters for this method are provided in [23 ].
The method based on matching pursuit technique
V. RESULTS & DISCUSSION
for HFOs detection is based on classical matching
th
pursuit as described in [24]. Finally, the 5 method
The calculation of the performance (sensitivity to which we compared our approach is the HHT
and FOR) of the proposed algorithm is done by detector, was presented by Chaibi et al.(2013 ) and
changing the input parameter (P). Figure 2 shows that is based on Hilbert Huang Transform (HHT)
the ROC curve which characterizes the variation the and RMS features (see [25] for a formal description
sensitivity as a function of the FOR. The optimal and input parameter details). The results of
threshold is indicated by the blue arrow that performance comparison between different methods
corresponds to P = 2.66. Sensitivity and FDR are shown in Figure 3 and summarized in Table 1.
corresponding to this point are 66.96% and 8.62%
As shown in Figure 3 and Table 1, the most
respectively. The AUC of the ROC curve is 0.8208.
robust method -with the data used here- in terms of
best compromise between sensitivity and FOR
appears to be the HHT-based method yielding the
largest difference between the sensitivity and the
FOR. On the other hand, the method based on the
complex Morlet wavelet (CMOR) is the one that
provides the highest AUC. It is also important to
note that the bumps method can achieve a
sensitivity of 98% with an FOR that is
approximately equal to 23 % (Figure 3 ). One
advantage of the proposed decision-tree method
100
compared to other techniques, is that it is the
FDR method that provides the lowest false detections
Fig. 2. ROC curve of the algorithm based on the descion tree (FDR = 8.62%) when considering the ROC point of
obtained by varying the input parameter (�). best compromise (see arrows on Fig.3). However,
the main limitation of the current implementation of
The comparison of performances between the decision-tree method is the lack of sensitivity
several automated detectors of HFOs is not an easy (66.96%). This sensitivity is not particularly poor
task. There are several challenges that account for but it is lower than what we report with the other
this difficulty: There are different techniques of methods. Indeed, the choice of the method to be
recording of EEG signals. Different types and used for detecting HFOs depends on the purpose,
locations of electrodes have been used in the the intended user and the minimal required
recordings of EEG signals. The defmitions of HFOs sensitivity, FOR and AUC performances (the latter
(e.g. frequency bands) vary across the literature. In may depend on whether the method is used in a
addition, there is also a lack of a "Gold Standard" fully or semi-automatic framework). In a semi
and a common database for researchers working on automatic approach, an excessively sensItlve
the detection of HFOs. So, caution should be taken method can be later corrected by visual inspection
into consideration when comparing the performance of the results of the automatic procedure. In a fully
among distinct detection algorithms. automatic method (or a pipeline that automatically
Therefore, it is crucial to test different methods uses the output of the detector for further analysis
using the same database and the same "Gold and decisions), a minimal false detection rate might
Standard". For this reason, we propose in this paper be a priority, even if it comes at the cost of missing
a comparative study between our proposed some events. As shown in Table 1, the FOR that
algorithm and five other methods that have been was obtained with the other 5 reference methods is
IEEE IPAS' 14: INTERNATIONAL IMAGE PROCESSING APPLICATIONS AND SYSTEMS CONFERENCE 2014
probably too high for practical and clinical positives that arise from spurious high-frequency
implementations. It is important to keep in mind bursts that are not true physiological HFOs and
here that, particularly in the quest for identifying which can be merely caused by the filtering of
HFOs, several phenomena can lead to false spikes and sharp waves.
100
90
80
70
60
Sensitivity
Bumps
50 MP
CMOR
40
Decision Tree
30 HHT
RMS
20
10
0
0 10 20 30 40 50 60 70 80 90 100
FDR
Fig. 3. Different ROC curves associated with different tested methods of HFOs detection.
Acknowledgment [16] Naeini RY. [Doctoral thesis] Automatic detection of high frequency
oscillations ofneural signals in epileptic patients [Doctoral thesis].
Authors wish to thank the Montreal Neurological Institute Montreal,Quebec,Canada:Concordia University; 2012.
and Hospital (MNI, Canada) for kindly providing the database [17] Gardner AB, Worrel GA, Marsh E, Dlugos D, Litt B. Human and
automated detec-tion of high-frequency oscillations in clinical
which we used to test and compare the performance of various
intracranial EEG recordings. ClinNeurophysiol 2007; 118:1134-43.
methods. We would like to thank Dr. Mohamed Dogui
[18] Zelmann R, Mari F, Jacobs J, Zijilmans M, Dubeau F, Gotman J. A
(Service of Functional Exploration of the Nervous System, comparisonbetween detectors of high frequency oscillations. Clin
CHU Sahloul of Sousse, Tunisia) for helping us doing visual NeurophysioI 2012;123:106-16.
identification process of HFOs in intracerbral EEG signals. [19] G.A. Worrell, K. Jerbi, K. Kobayashi, J.M. Lina, R. Zelmann, M. Le
Van Quyen. Recording and analysis techniques for high-frequency
References oscillations. Progress in lITeurobiology 2012- 98: 265-278.
[20] E. Kvedalen, Signal processing using the Teager Energy Operator and
othernonlinear operators, University of Oslo, 2003 (Cand. Scient
[I] Anatol Bragin, Jerome Engel, Charles L. Wilson, Ttzhak Fried , Gary Thesis).
W. Mathern. Hippocampal and entorhinal cortex high-frequency [21] Lachiche N., Apprentissage automatique : Arbres de decision. 2008.
oscillations (100-500 Hz) in human epileptic brain and in kainic
[22] Khalilov, I., Le Van Quyen, M., Gozlan, H., Ben-Ari, Y., 2005.
acid-treated rats with chronic seizures. Epilepsia 1999; 40(2): 127-
Epileptogenic actions of GABA and fast oscillations in the
137. developing hippocampus. lITeuron 2005; 48:787-796.
[2] Anatol Bragin, Charles L. Wilson, Richard J. Staba, Mark Reddick,
[23] Chiran Dilip Doshi. Methods for detecting high frequency oscillations
Itzhak Fried, Jerome Engel. Interictal High-Frequency Oscillations in ongoing brain signals: Applications to the determination of epileptic
(80-500Hz) in the Human Epileptic Brain: Entorhinal Cortex. Ann seizure zones,Marquette University,20II.
lITeuroI 2002; 52: 407-415.
[24] Sahbi Chaibi, Tarek Lajnef, Zied Sakka, Mounir Samet, Abdennaceur
[3] Staba, RJ.,Wilson,C.L.,Bragin,A., Fried, I., Engel, J.J. Quantitative Kachouri. A Comparison of Methods for Detection of High Frequency
analysis of high-frequency oscillations (80-500 Hz) recorded in
Oscillations (HFOs) in Human Intacerberal EEG Recordings.
human epileptic hippocampus and entorhinal cortex. J. lITeurophysiol American Journal of Signal Processing 2013,3(2): 25-34.
2002 ; 88 :1743-2152.
[25] Sahbi Chaibi, Zied Sakka, Tarek Lajnef, Mounir Samet, Abdennaceur
[4] MacReady N. Radiotherapy and localization of seizures cited as
Kachouri. Automated detection and classification of high frequency
promising therapies. Epilepsy Neurol Today 2008;8:12-4. oscillations (HFOs) in human intracereberal EEG. Biomedical Signal
[5] Elena Urrestarazu, Rahul Chander, Francois Dubeau, Jean Gotman. Processing and Control 8 (2013) 927- 934.
Interictal high-frequency oscillations (100-500Hz) in the intracerebral
EEG of epileptic patients. Brain 2007; 130: 2354-2366.
[6] Julia Jacobs, Pierre LeVan. Rahul Chander, Jeffery Hall, Francois
Dubeau, Jean Gotman. Interictal high-frequency oscillations (80-500
Hz) are an indicator of seizure onset areas independent of spikes in the
human epileptic brain. Epilepsia 2008; 49(11) :1893-1907.
[7] Jacobs J, LeVan P, Chatillon CE, Olivier A, Dubeau F, Gotman J.
High frequency oscil-Iations in intracranial EEGs mark
epileptogenicity rather than lesion type. Brain2009a;132:I022-37.
[8] Jacobs J, Zijlmans M, Zelmann R, Chatillon CE,Hall J, Olivier A, et
al. High-frequency electroencephalographic oscillations correlate with
outcome of epilepsy surgery. Ann lITeurol2010; 67:209-20.
[9] Wu JY, Sankar R, Lerner JT, Matsumoto JH, Vinters HV, Mathern
GW. Removing interictal fast ripples on electrocorticography linked
with seizure freedom in children. lITeurology 2010;75:1686-1694.
[10] Cuevas LA, Castillo-Toledo B, Medina-Ceja L, Ventura-Mejia C,
PardoPen K. An algo-rithm for on-line detection of high frequency
oscillations related to epilepsy. Comput Methods Prog Biomed
2013;110:354-60.
[II] Khosravani H,Mehrotra N,Rigby M,J. Hader W,Pinnegar CR,Pillay
N, et al.Spatial localization and time-dependant changes of
electrographic high fre-quency oscillations in human temporal lobe
epilepsy. Epilepsia 2009;50:605-16.
[12] Kalitzin S, Zijlmans M, Petkov G, Velis D, Claus S, Visser G, et al.
Quantification ofspontaneous and evoked HFO's in SEEG recordings
and prospective for pre-surgical diagnostics. Case study. In: 34th
Annual International Conference ofthe IEEE EMBS; 2012.
[13] Zijlmans M, Jacobs J, Zelmann R, Dubeau F, Gotman J. High
frequency oscillations and seizure frequency in patients with focal
epilepsy. Epilepsy Res 2009;85:287-92.
[14] Salami P, Levesque M, Gotman J, Avoli M. A comparison between
automated detec-tion methods of high-frequency oscillations (80-500
Hz) during seizures. JNeurosci Methods 2012;211:265-71.
[15] Gotman J. The epileptic focus: a zone or a network? The interictal
focus: spikes andHFOs. In: 30th International Epilepsy Congress
Montreal. Montreal Neurologicallnstitute McGill University; 2013.