CLINPH 2007856

No. of Pages 12, Model 5G

14 June 2016
Clinical Neurophysiology xxx (2016) xxxxxx
1

Contents lists available at ScienceDirect

Clinical Neurophysiology
journal homepage: www.elsevier.com/locate/clinph
5
6
3
4
7
8
9
10
11
12
13
14
2 9
15

Real-time multi-channel monitoring of burst-suppression using neural

network technology during pediatric status epilepticus treatment
Christos Papadelis a,b,, Seyedeh Fatemeh Salimi Ashkezari a,b, Chiran Doshi a,b, Sigride Thome c,
Phillip L. Pearl c, Patricia Ellen Grant a,b,d, Robert C. Tasker c,e, Tobias Loddenkemper c
a

Center for Fetal-Neonatal Neuroimaging and Developmental Science, Boston Childrens Hospital, Harvard Medical School, 1 Autumn St, Boston, MA 02215, USA
Division of Newborn Medicine, Department of Medicine, Boston Childrens Hospital, Harvard Medical School, 300 Longwood Ave, Boston, MA 02115, USA
Department of Neurology, Boston Childrens Hospital, Harvard Medical School, 300 Longwood Ave, Boston, MA 02115, USA
d
Department of Radiology, Boston Childrens Hospital, Harvard Medical School, 300 Longwood Ave, Boston, MA 02115, USA
e
Department of Anesthesia, Perioperative and Pain Medicine, Division of Critical Care Medicine, Boston Childrens Hospital, Harvard Medical School, 300 Longwood Ave, Boston,
MA 02115, USA
b
c

a r t i c l e

i n f o

h i g h l i g h t s

1 7
18
19
20

Article history:
Accepted 27 May 2016
Available online xxxx

21
22
23
24
25
26
27
28

Keywords:
Status epilepticus
Intensive care unit
Burst-suppression
Monitoring
Electroencephalography
Neural network

54
55

30
31
 Multichannel real time system that automatically identifies burst-suppression.

32

 Estimates burst-suppression index using neural network technology.

33
34
35

 Excellent agreement between automated and manual classification of burst-suppression.

36

a b s t r a c t

3 7

Objective: To develop a real-time monitoring system that has the potential to guide the titration of anesthetic agents in the treatment of pediatric status epilepticus (SE).
Methods: We analyzed stored multichannel electroencephalographic (EEG) data collected from 12 pediatric patients with generalized SE. EEG recordings were initially segmented in 500 ms time-windows.
Features characterizing the power, frequency, and entropy of the signal were extracted from each segment. The segments were annotated as bursts (B), suppressions (S), or artifacts (A) by two electroencephalographers. The EEG features together with the annotations were inputted in a three-layer feed
forward neural network (NN). The sensitivity and specificity of NNs with different architectures and
training algorithms to classify segments into B, S, or A were estimated.
Results: The maximum sensitivity (95.96% for B, 89.25% for S, and 75% for A) and specificity (89.36 for B,
96.26% for S, and 99.8% for A) was observed for the NN with 10 nodes in the hidden layer. By using this
NN, we designed a real-time system that estimates the burst-suppression index (BSI).
Conclusions: Our system provides a reliable real-line estimate of multichannel BSI requiring minimal
memory and computation time.
Significance: The system has the potential to assist intensive care unit attendants in the continuous EEG
monitoring.
2016 International Federation of Clinical Neurophysiology. Published by Elsevier Ireland Ltd. All rights
reserved.

38

56
57
58
59

Abbreviations: A, artifact; B, burst; BCH, Boston Childrens Hospital; BFGS, BroydenFletcherGoldfarbShanno quasi-Newton backpropagation; BR, Bayesian
regularization; BS, burst-suppression; BSD, burst-suppression dynamic; BSI, burst-suppression index; BSP, burst-suppression probability; CM, confusion matrix; EDF,
European data format; EEG, electroencephalography; FN, false negative; FP, false positive; FPR, false positive rate; ICU, intensive care unit; LM, LevenbergMarquardt; MSE,
mean square error; NN, neural network; NPV, negative predicted value; PatternNet, pattern recognition NN; PPV, positive predicted value; ROC, receiver operating
characteristic; RP, Resilient backpropagation; S, suppression; SCG, scaled conjugate gradient; SE, status epilepticus; TN, true negative; TP, true positive; TPR, true positive rate.
Corresponding author at: Boston Childrens Hospital, Harvard Medical School, 300 Longwood Ave, Boston, MA 02115, USA. Tel.: +1 781 216 1128; fax: +1 781 216 1172.
E-mail address: [email protected] (C. Papadelis).
http://dx.doi.org/10.1016/j.clinph.2016.05.358
1388-2457/ 2016 International Federation of Clinical Neurophysiology. Published by Elsevier Ireland Ltd. All rights reserved.

Please cite this article in press as: Papadelis C et al. Real-time multi-channel monitoring of burst-suppression using neural network technology during
pediatric status epilepticus treatment. Clin Neurophysiol (2016), http://dx.doi.org/10.1016/j.clinph.2016.05.358

39
40
41
42
43
44
45
46
47
48
49
50
51
52
53

CLINPH 2007856

No. of Pages 12, Model 5G

14 June 2016
2

C. Papadelis et al. / Clinical Neurophysiology xxx (2016) xxxxxx

60

1. Introduction

61

Status epilepticus is a condition resulting either from the failure

of the mechanisms responsible for seizure termination or from the
initiation of mechanisms, which lead to abnormally, prolonged seizures (Trinka et al., 2015). Convulsive SE is the most common
childhood neurological emergency in developed countries with
1820 episodes per 100,000 cases per year in children younger
than 16 years old (Chin et al., 2004; 2006). Because of its significant
morbidity and mortality (Logroscino et al., 2001; Claassen et al.,
2002; DeLorenzo et al., 1996), SE requires prompt diagnosis and
treatment. After initial first- and second-line therapies, anesthetic
agents are sometimes used if the episode fails to come under
control.
In the setting of using anesthesia for seizure control, patients
are invariably supported by mechanical ventilation and managed
on an intensive care unit (ICU). The goal or target now is to induce
and maintain a depth of anesthesia that will control and prevent
seizures, yet avoid the possibility of overdosing the anesthetic
agent, which leads to a cessation of all EEG activity as well as
impairing cardiac function to the point that vasopressors are
required to support life (Wilkes and Tasker, 2013, 2014; Tasker
and Vitali, 2014). The dose of the anesthetic agent is frequently
adjusted with the aim of achieving a characteristic EEG pattern,
termed burst-suppression (BS), which consist of alternate periods
of slow waves of high-amplitude (the bursts, B) and periods of
low-amplitude EEG (the suppressions, S) that represent the interaction between neuronal excitatory dynamics and metabolism.
The duration of the B and S periods varies with the concentration
of the anesthetic agent. Higher concentrations lead to progressively longer duration S periods eventually causing continuous
EEG suppression (Vijn and Sneyd, 1998; Young, 2000; Brenner,
2005; Ching et al., 2012a,b). These changes in the relative duration
of B and S periods can be quantified by the burst-suppression index
(BSI), which is a measure of the percentage of time within an interval spent in the suppressed state (DeGiorgio, 1993). When anesthesia is used in the treatment of SE, the standard approach is to
control the suppression intensity of the EEG signal by continuously
monitoring the BSI and manually adjust drug dosage. Current
treatment protocols of SE typically target a BS pattern of 515 s
suppressions alternating with 15 s bursts or equivalently BSI values of 5095% (Cottenceau et al., 2008; Riker et al., 2003; Rossetti
et al., 2011; Musialowicz et al., 2010; Wilkes and Tasker 2013).
In clinical practice, the identification of BS patterns and the differentiation between B periods is currently performed visually by
the ICU nursing staff who estimate the approximate timing
between one or more interburst intervals. The BSI is thus most
often assessed as a guesstimate rather than using a quantitative
reproducible approach. Equally important is the practicality of
such an approach. During this process, the ICU staff should continuously watch for subtle changes in the BS pattern interburst interval, and titrate anesthetic agents based on the calculated interval.
Since these patients often require therapy for days (Wilkes and
Tasker, 2014), it is unrealistic to expect the attendant staff to maintain minute-to-minute control of BS state by manually changing
the rate of anesthetic administration in response to changes in
the EEG signal. Thus, a closed loop feedback between the monitoring and the treating team members is required.
Many methods of varying complexity have been previously
described for the detection of BS patterns (Kaiser, 1990; Thomsen
et al., 1991; Arnold et al., 1996; Akrawi et al., 1996; Sherman
et al., 1997; Griessbach et al., 1997; Leistritz et al., 1999;
Muthuswamy et al., 1999; Bruhn et al., 2000a,b; Srkel et al.,
2002; Lfhede et al., 2007; Yoon et al., 2012; Chemali et al.,
2013; Westover et al., 2013; Liang et al., 2014). All these methods

62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123

have dealt with the EEG as a single channel (usually recorded from
central areas) assuming that its activity represents the activity of
the entire brain. This is because BS has classically been viewed as
a global state with synchronous activity throughout the cortex,
a perspective that was derived from previous EEG studies in which
periods of B and S have been shown to occur concurrently across
the scalp (Clark and Rosner, 1973). However, recent evidence of
intracranial electrocortigrams from patients who entered BS indicated that in contrast to previous characterizations, periods of B
could be substantially asynchronous across the cortex (Lewis
et al., 2013). Furthermore, it was shown that the state of BS itself
could occur in a limited cortical region while other areas exhibit
ongoing continuous activity.
Here, we describe the development of a multichannel EEGbased system that automatically identifies the BS patterns and estimates the BSI. The system was built using NN technology combining information from the amplitude, the frequency content, and
the entropy of the EEG signal. The NN was trained using multichannel EEG data from pediatric patients hospitalized in the ICU
and was validated against manual segmentation of the data by
two board certified electroencephalographers. Different architectures and training algorithms were tested. The NN showing the
best performance in terms of sensitivity and specificity for detecting the BS patterns was used for the design of the on-line system.
The on-line system was incorporated into a commercial EEG device
(eegoTM mylab, ANT Neuro, Netherlands). Our system has the potential to assist ICU attendants with continuous monitoring of the EEG
signal by providing reliable automatic estimates of the BSI in the
treatment of pediatric SE.

124

2. Methods

153

2.1. Patients

154

We analyzed stored EEG data from 12 pediatric patients (7

males and 5 females) hospitalized in the ICU of Boston Childrens
Hospital (BCH) with generalized SE. These patients were a representative retrospective sample of BS pattern EEG recordings from
all pediatric patients who underwent continuous EEG monitoring
in the ICU. Demographic data, primary admission diagnosis, final
diagnosis/discharge, the history of epilepsy and/or SE, and the
anesthetic agents administered at the time of EEG recording are
presented in Table 1. Review of clinical and EEG data was carried
out with the approval of the local institutional review board.

155

2.2. EEG recordings and expert data segmentation

165

Given the retrospective nature of the study, we did not control

the length of the EEG recordings. All EEGs were recorded using 19
silver/silver electrodes, affixed to the scalp according to the international 1020 system. EEG data were recorded by using the
XLTEK clinical EEG equipment (Natus Medical Inc., Canada) with
a sampling rate of 200 Hz. The signals were referred to mastoids.
Two board certified electroencephalographers independently
reviewed the EEG recordings using inhouse viewing and annotation software written in Matlab (MathWorks Inc., MA, USA). The
software segments the continuous EEG recording into epochs of
500 ms duration, and allows the reviewer to classify each segment
as B, S, or A. The review was based on the bipolar recordings of
frontal-central electrodes (i.e., F3-C3 and F4-C4), but data from
the complete array of recording channels was also available to
the reviewer during the marking-up process. Only segments for
which there was an agreement between the two electroencephalographers were used in further analysis. If a segment was annotated
and classified by both reviewers as B, S, and A, then the segment

166

Please cite this article in press as: Papadelis C et al. Real-time multi-channel monitoring of burst-suppression using neural network technology during
pediatric status epilepticus treatment. Clin Neurophysiol (2016), http://dx.doi.org/10.1016/j.clinph.2016.05.358

125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152

156
157
158
159
160
161
162
163
164

167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183

CLINPH 2007856

No. of Pages 12, Model 5G

14 June 2016
3

C. Papadelis et al. / Clinical Neurophysiology xxx (2016) xxxxxx

Table 1
Patients demographic data (gender, age in years); primary admission diagnosis; final diagnosis/discharge; history of epilepsy; previous SE; sedation drug in ICU; number of EEG
datasets used in NN training, validation, and testing. M: male; F: female; FIRES: Febrile Infection-Related Epilepsy Syndrome.
ID

Gender

Age

Primary admission diagnosis

Final diagnosis/Discharge

History of
Epilepsy

Previous
SE

Sedation drug

Number of
Datasets

1
2
3

M
F
M

10
9
11

Status epilepticus
FIRES
Tonic clonic seizure

Yes
No
No

Yes
No
No

Isoflurane
Propofol
Pentrobarbital

10
1
7

18

Yes

Yes

Pentobarbital

Seizures with profound

encephalopathy
FIRES

Metabolic (partial Trisomy 13)

Arteriovenous malformation
Left thalamic arteriovenous
malformation
Unknown etiology

Yes

No

Pentobarbital

Refractory generalized epilepsy

Yes

Yes

Isoflurane

7
8

M
M

13
18

Refractory intracranial hypertension/

Deceased
Pyridoxal 5-phosphate-dependent
epilepsy
Unknown etiology
Epilepsy; Grand Mal Status, Epileptic

Yes
Yes

No
No

Pentobarbital
Isoflurane

4
4

9
10
11
12

M
M
F
M

Yes
No
No
No

Yes
No
No
No

Pentrobarbital
Pentobarbital
Pentobarbital
Pentobarbital

3
3
1
1

9
3.5
7
5

Catastrophic refractory epilepsy

Primary generalized idiopathic
epilepsy
Status epilepticus
Traumatic brain injury
Traumatic brain injury
Acute respiratory and renal failure

Seizure Disorder
Bilateral subdurmal hematomas
Subdural Hematoma/Deceased
Tetralogy of Fallot

203

was further used in the analysis. For the cases that there was not an
agreement between the two classifications, the segments were not
used in the analysis.
According to the literature, a common approach to quantitatively distinguish B and S is to define suppressions as EEG segments of less than a threshold voltage value, typically set at
between 0.5 and 20 lV, which last at least 500 ms (Rampli and
Laster, 1992; Niedermeyer et al., 1999). Here, we instructed
reviewers to mark as B, all segments with an activity having
approximately an amplitude of >10 lV and duration more than
half the segment duration >250 ms. Since the theoretical lower
amplitude during busts has been proposed to be at 20 lV (or
even higher) (Hirsch et al., 2013), reviewers also considered the
relevant amplitude change between the suppression and burst patterns in their assessment. All remaining segments were classified
as S, except segments classified empirically as artifacts (A). EEG
patterns having alternate segments of burst and suppression activity in one minute of continuous EEG recording were regarded as
distinct BS. EEG patterns having continuous burst activity in one
minute of continuous recording were regarded as indistinct BS.

204

2.3. EEG feature extraction

205

218

The data from two bipolar EEG channels (i.e. F3-C3 and F4-C4)
were used as inputs in the system. These two channels were
selected because they provide information from the frontal and
central areas of the brain. The data were initially chopped into segments of 500 ms duration. A line filter with cutoff frequencies of
0.6 and 30 Hz was applied to each segment. Then, the absolute
and the relative power for four different frequency bands of the
signal were estimated (i.e., delta (14 Hz), theta (47 Hz), alpha
(712 Hz), and beta (1230 Hz)), as well as the Shannon entropy
(Shannon, 1948). The boundaries of the frequency bands were
selected to include the entire spectrum of the EEG signal from 1
to 30 Hz. Fig. 1 displays 30 s of EEG recording (two bipolar channels) with the corresponding estimated power and frequency features of the signal.

219

2.4. NN Design

220

Our classification system was based on artificial NN technology

(Patterson, 1996) and used the pattern recognition NN (PatternNet) approach. PatternNet is a feed forward NN that can be
trained to classify inputs based on the desired outputs/targets. In

184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202

206
207
208
209
210
211
212
213
214
215
216
217

221
222
223

our study, the NN was supposed to classify each segment of the

recorded EEG signal as B, S, or A. Our NN consisted of an input
layer, a hidden layer, and an output layer with sigmoid transfer
function (Fig. 2a). As inputs of the NN we used the nine features,
which were extracted from the EEG signal; these features include
the absolute and relative power of four frequency bands (delta,
theta, alpha and beta) and the signal entropy of the EEG signal.
Also, as targets the results from the manually annotated EEG files
were used. NNs with different numbers of nodes in the hidden
layer (ranging from 5 to 35 with steps of 5) were tested to estimate
the optimum number of hidden layer nodes. The output layer had
three nodes which indicate the three classes B, S, or A; each node
had an output ranging from 0 to 1. For each segment the algorithm
selects the value that is closest to 1 as the class for the segment (for
each segment, from the three classes, the closest value to 1 is converted to 1 and the other two classes are converted to 0 to make a
binary output). The design of the NN was performed by using the
pattern recognition tool in the NN toolbox of Matlab version
R2014a (Mathworks, MA, USA).

224

2.5. Training, validation, and testing of the NN

243

The data were originally divided into segments with 500 ms

duration. The segments were randomly selected to be used as
training, validation, or testing segments. The first subset of segments was the training set; a set of examples used for learning that
is to fit the parameters (i.e., weights and biases) of the NN. The second subset was the validation set; a set of examples used to tune
the parameters (i.e., architecture) of the NN. The third subset
was the testing set; a set of examples used to assess the performance of the fully trained NN. None of the segments were used
in two different subsets.
For the training set of the NN, we randomly selected 70% of all
segments. Five different optimization training algorithms were
used to individually train the network: the scaled conjugate gradient (SCG), the LevenbergMarquardt (LM), the Bayesian regularization (BR), the BroydenFletcherGoldfarbShanno quasi-Newton
backpropagation (BFGS), and the Resilient backpropagation (RP)
algorithm. Technical details about the different algorithms are provided in the Appendix A. During the training process, the performance of each algorithm was estimated by comparing the
predicted output of the NN with the desired output based on the
manual annotations provided by the electroencephalographers.
The closer the result, the better the performance of the network

244

Please cite this article in press as: Papadelis C et al. Real-time multi-channel monitoring of burst-suppression using neural network technology during
pediatric status epilepticus treatment. Clin Neurophysiol (2016), http://dx.doi.org/10.1016/j.clinph.2016.05.358

225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242

245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265

CLINPH 2007856

No. of Pages 12, Model 5G

14 June 2016
4

C. Papadelis et al. / Clinical Neurophysiology xxx (2016) xxxxxx

Fig. 1. Display of EEG recordings (bipolar channels: F3-C3 and F4-C4) and the estimated power and frequency features. Features are estimated for segments of 500 ms
duration: 1st and 5th rows: filtered EEG recordings, 2nd and 6th rows: absolute power for the different frequency bands represented with different colors (delta: blue, theta:
cyan, alpha: yellow, and beta: red), 3rd and 7th rows: relative power for the same frequency bands, and 4th and 8th rows: Shannon entropy measures. The user can adjust: (i)
the duration of the displayed EEG section, (ii) the selection of two representatives EEG channels (i.e. bipolar, monopolar montages), and (iii) the amplitude of the EEG signals.
(For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 2. Schematic overview of (a) the off-line NN training process, and (b) the on-line system.
266
267
268
269
270
271

was in classifying different segments. The training process continued until: (i) the difference between the predicted output and the
desired output/target became zero, (ii) the training reached a maximum number of epochs (1000)1, (iii) the number of validation
checks2 reached 6, or (iv) the magnitude of the minimum performance gradient3 became less than <10 5.

An epoch is a measure of the number of times all of the training vectors are used
once to update the weights.
2
Validation checks are used to stop the training process early if the network
performance on the validation vectors fails to improve or remains the same.
3
If training reaches a minimum of the performance, the gradient will become very
small.

For the validation set of the NN, we randomly selected 15% of all
segments. The goal of validation was to minimize the NN overfitting (see Geman et al., 1992). Overfitting occurs when the accuracy
over the training set increases, but the accuracy over the validation
set stays the same or decreases. In this case, the outputs perfectly
match the training set, but when the NN is used to make predictions for a new dataset it provides poor accuracy. Each algorithm
was validated using the following estimated measures, where: TP
is true positive, TN is true negative, FP is false positive, and FN is
false negative: (a) sensitivity = TP/(TP + FN)*100, (b) specificity = TN/(FP + TN)*100, (c) PPV (positive predicted value) = TP/
(TP + FP)*100, (d) NPV (negative predicted value) = TN/(TN + FN)
*100, and (e) the mean square error (MSE), which presents the dif-

Please cite this article in press as: Papadelis C et al. Real-time multi-channel monitoring of burst-suppression using neural network technology during
pediatric status epilepticus treatment. Clin Neurophysiol (2016), http://dx.doi.org/10.1016/j.clinph.2016.05.358

272
273
274
275
276
277
278
279
280
281
282
283
284

CLINPH 2007856

No. of Pages 12, Model 5G

14 June 2016
C. Papadelis et al. / Clinical Neurophysiology xxx (2016) xxxxxx
285
286

287
289

ference between the predicted output and the desired NN model as

shown in Eq. (1):

between the two different montages. Statistical significant threshold was set to p < 0.01 (corrected for multiple comparisons).

345

n
n
X
X
MSE 1=n errori 2 1=n Ti

2.8. Fixed duration segments vs. manual annotations

347

In order to test whether our approach to use a fixed duration for

the EEG segments (i.e. 500 ms) -instead of an adaptive segmentation approach- was affecting significantly the reliable estimation
of BSI values, we performed the following test. The filtered EEG
data from two bipolar channels (i.e. F3-C3 and F4-C4) were annotated by the two electroencephalographers independently but now
using a different approach. The exact duration of burst and suppression was estimated for each dataset (see Supplementary
Fig. S1), and then for each dataset the BSI was calculated from
these manual annotations. Only segments for which there was an
agreement between the two electroencephalographers were used
in the calculation of BSI. The BSI values that were estimated based
on the manual annotations were then compared to the BSI values
estimated by using the NN for each dataset using t-test. Statistical
significant threshold was set to p < 0.01.

348

2.9. On-line classification system

363

EEG signals were recorded using the eegoTM mylab EEG system
(ANT Neuro, Netherlands). On-line reading of the data was performed through the TMSI plugin of the MNE-X software (Dinh
et al., 2013). EEG data were saved as individual European data format (edf) files, each one having duration of one minute. The NN
showing the best performance was used for the development of
the on-line system (see Fig. 2b). The same procedure was followed
as in the off-line version of the algorithm regarding segmentation,
filtering, and the nine EEG features estimation (see Section 2.3.).
Afterwards, the features were estimated for each channel and feed
into the already trained NN. The on-line NN gives as outputs values
that range from 0 to 1 for each node. Each segment is then classified as B, S, or A according to the output value of each node and the
BSI is estimated for each minute of recording (see Fig. 1).

364

3. Results

378

3.1. EEG patterns

379

Different EEG patterns varying both in amplitude and spectral

characteristics were observed during the visual inspection of the
data. Fig. 3 summarizes the four most common patterns which
were observed: (i) patterns with periodic discharges, (ii) patterns
with distinct BS, (iii) patterns with indistinct BS, and (iv) patterns
contaminated with muscular artifacts. The total duration of the
recordings was 159 h and 18 min (55 datasets; 1,146,940 segments
in total for all channels). Patterns with epileptiform spikes were
observed within 7 datasets; patterns with distinct BS were
observed in 27 datasets; patterns with indistinct BS were observed
in 40 datasets and patterns with artifacts were observed within 18
datasets. The annotation process revealed in total 883,182 (77%)
segments that were classified as periods of B, 259,684 (22.6%)
segments that were classified as periods of S, and 4,074 (0.4%)
segments that were classified as A.

380

3.2. NN performance in detecting BS patterns

395

Table 2 presents the MSE values for NNs with different architecture being trained with different training algorithms. Overall, MSE
was <0.1 for all NN architectures and all training algorithms except
for two NNs (BFGS with 5 and 15 nodes in the hidden layer). The
LM and BR training algorithms showed the lowest MSE regardless

396

i1

Oi 2

307

308

2.6. NNs performance

309

As the NN with the best performance, we identified the network

which was satisfying the following specific criteria: (i) highest sensitivity and specificity without overfitting (see Geman et al., 1992);
(ii) MSE close to zero; (iii) reasonable training time (<5 h); and (iv)
high accuracy on the testing set.
For each algorithm, the training duration, the confusion matrix
(CM) (also known as an error matrix), and the receiver operating
characteristic (ROC) curve were estimated. The CM visualizes the
algorithms performance; it presents the correctly classified and
misclassified cases for each of the defined classes (B, S, and A),
and the total percentage of correctly classified and misclassified
cases. The ROC curve plots the sensitivity versus the 1-specificity
for different thresholds of the classifier output. It thus represents
the balance between sensitivity and specificity: an increase in sensitivity results in a decrease in specificity. The closer the ROC curve
is to the left-top corner of the plot, the more precise the classifier is
[100% specificity (no false positive) and 100% sensitivity (no false
negative)]. The diagonal divides the ROC space. Points below the
diagonal indicate bad classification (worse than random) and
points above the diagonal indicate good classification (better than
random).

291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306

310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329

330

2.7. Spatiotemporal analysis

331

Based on observations from previous electrocorticography studies (Lewis et al., 2013), we hypothesized that multichannel EEG
recordings contain significantly different information compared
to single (or dual-channel) recordings in terms of BS content. To
test our hypothesis, we explored our multichannel EEG data from
representative cases of our datasets showing patterns of distinct
BS, estimated the BSI for each bipolar channel and displayed them
on a topoplot color-coded map. We further estimated the BSI for all
of our datasets using: (i) two bipolar channels (i.e., F3-C3 and F4C4), and (iii) the full EEG sensor array (14 bipolar channels). The
first configuration was used because it represents the traditional
approach assuming that BS patterns from the central areas represent the activity of the entire brain. Repeated measures analysisof-variance (ANOVA) was used to assess differences in the BSI

332
333
334
335
336
337
338
339
340
341
342
343
344

346

i1

where, n is the number of predictions, Ti is the target/desired output, and Oi is the predicted output. Target/desired output is the binary representation of the manual annotations; the output will be
zero when the segment is either A or S, and it will be 1 when the
segment is B. These values are inputted into the NN during training.
On the other hand, predicted output is the final result of the NN
after it has been trained. Unlike the target output, the predicted output is a number between 0 and 1 and not a binary number. The MSE
is basically a network performance function and it assesses the
quality of the predictor/NN. The MSE for NNs with different number
of hidden layer neurons and different training algorithms was
calculated.
For the testing set of the NN, we randomly selected 15% of all
segments. The testing set was used to estimate the performance
of the NN on data which were not used before for validation or
training. The main reason of using the testing set was to check
whether the training was sufficient and whether the validation
prevented overfitting.

290

Please cite this article in press as: Papadelis C et al. Real-time multi-channel monitoring of burst-suppression using neural network technology during
pediatric status epilepticus treatment. Clin Neurophysiol (2016), http://dx.doi.org/10.1016/j.clinph.2016.05.358

349
350
351
352
353
354
355
356
357
358
359
360
361
362

365
366
367
368
369
370
371
372
373
374
375
376
377

381
382
383
384
385
386
387
388
389
390
391
392
393
394

397
398
399
400

CLINPH 2007856

No. of Pages 12, Model 5G

14 June 2016
6

C. Papadelis et al. / Clinical Neurophysiology xxx (2016) xxxxxx

Fig. 3. Summary of EEG BS patterns. The four common patterns are shown: (a) patterns containing periodic discharges, (b) patterns with distinct BS, (c) patterns with
indistinct BS and (d) signals contaminated with artifacts.

Table 2
MSE for NNs with different number of nodes and trained with different algorithms.
Algorithm

BFG
BR
SCG
LM
RP

Number of nodes in the hidden layer

5

10

15

20

25

30

35

0.1053
0.0327
0.0765
0.0328
0.0942

0.0877
0.0311
0.0747
0.0314
0.0884

0.1061
0.0315
0.0811
0.0311
0.0836

0.0418
0.0019
0.0650
0.0314
0.0791

0.0565
0.0308
0.0649
0.0308
0.0762

0.0422
0.0312
0.0631
0.0306
0.0839

0.0485
0.0311
0.0682
0.0310
0.0781

410

of the number of nodes in the hidden layer. No remarkable

improvement was observed by adding more nodes in the hidden
layer. The lowest MSE was achieved by the NN with 20 nodes in
the hidden layer that was trained with the BR (see Table 2). However, this value occurred due to over-fitting of the NN. Over-fitting
is one of the problems that occur during NN training. It assigns a
very small value to the error on the training set. However, when
new data are inputted into the NN the error is large. In that case,
the network has memorized the training examples, but it has not
learned to generalize to new situations.

411

3.3. Training time

412

417

Table 3 presents the training time for NNs with different architectures being trained with different training algorithms. The fastest algorithm was the RP with mean training time 16 min. The
slowest algorithm was the BR with mean training time
1186 min. Remarkable increase of training timing (in most cases)
was observed by the addition of more nodes in the hidden layer.

418

3.4. Sensitivity and specificity

419

Fig. 4 presents the sensitivity and specificity of the different NN

configurations for detecting the B segments. The overall sensitivity
and specificity ranged from 72.9% to 95.96% (91.89% 5.5) and
from 63.66% to 89.36% (80.28% 8.7), respectively. The LM and

401
402
403
404
405
406
407
408
409

413
414
415
416

420
421
422

BR training algorithms presented the highest sensitivity values

for the detection of B (95.86% 0.07 for LM, and 95.84% 0.13 for
BR). The maximum sensitivity and specificity was observed for
the NN with 10 nodes in the hidden layer that was trained with
the LM algorithm (95.96% and 89.36% respectively). The sensitivity
was steady (95%) for NNs with >10 nodes for the BR and LM training algorithms. NNs trained with other algorithms having >10
number of nodes started to indicate an unstable behavior (see
Fig. 4).
Fig. 5 presents the sensitivity and specificity of different NN
configurations to detect the S segments. The overall sensitivity
and specificity of the NN to detect S segments ranged from 61.9%
to 89.91% (79.23% 16.4) and from 73.09% to 96.26%
(92.19% 5.5), respectively. The maximum sensitivity was
observed for the NN with 20 nodes that was trained with the BR
algorithm (89.91%). Also, the maximum specificity was observed
for the NN with 10 nodes that was trained with the LM algorithm
(96.26%). NNs trained with other algorithms with a number of
nodes >10 started to indicate an unstable behavior (see Fig. 5).

423

3.5. Artifact

442

SCG and BFGS showed zero sensitivity and specificity of 99% for
the detection of A segments regardless of the training configuration. The mean sensitivity and specificity for the rest of algorithms
(LM, BR and RP) was 74.29% 28 and 99.76% 0.11 respectively.
The sensitivity and specificity of the NN with 10 nodes in the hidden layer that was trained with the LM algorithm were 75% and
99.8% respectively.

443

3.6. PPV, NPV, CM, and ROC Curve

450

The overall PPV and NPV ranged between 69.8% and 99.3%, and
between 68.8% and 99.03% respectively. For the NN presenting the
best performance, the PPV for detecting B, S, and A segments were
95.69%, 90.22% and 60% respectively. For this NN configuration, the
corresponding NPV values were 89.68%, 95.86% and 99.9%.

451

Please cite this article in press as: Papadelis C et al. Real-time multi-channel monitoring of burst-suppression using neural network technology during
pediatric status epilepticus treatment. Clin Neurophysiol (2016), http://dx.doi.org/10.1016/j.clinph.2016.05.358

424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441

444
445
446
447
448
449

452
453
454
455

CLINPH 2007856

No. of Pages 12, Model 5G

14 June 2016
7

C. Papadelis et al. / Clinical Neurophysiology xxx (2016) xxxxxx

Table 3
Computational time (h:m:s) for training NNs with different number of nodes using different training algorithms.
Algorithm

Number of nodes in the hidden layer

5

10

15

20

25

30

35

BFG
BR
SCG
LM
RP

02:26
1:47:49
06:01
28:32
07:41

07:24
8:14:16
09:36
1:44:45
10:18

3:59
17:19:26
06:06
5:49:02
12:43

33:26
25:05
19:20
12:16:15
15:20

23:30
44:40:45
22:22
25:00:48
18:27

41:38
19:59:56
24:03
38:05:53
23:16

1:04:51
45:54:18
26:32
37:24:42
23:24

Fig. 4. Sensitivity and specificity (%) of NNs with different configurations (number of nodes in the hidden layer and different algorithms) for detecting the B segments.

Fig. 5. Sensitivity and specificity (%) of NNs with different configurations (number of nodes in the hidden layer and different algorithms) for detecting the S segments.

456
457
458
459

Fig. 6 shows the CM and ROC curve for the NN configuration

with the optimal performance. By using the CM, we visualized
the performance of the algorithm. Correctly classified cases are
shown in green diagonal cells. Off-diagonal cells (in red) present

the cases that were misclassified. The total percent of correctly

classified cases and the total percent of misclassified cases are
shown in the blue cell (94% and 6% respectively). 0.2% of the A segments and 2.7% of the S segments were misclassified as burst. Also,

Please cite this article in press as: Papadelis C et al. Real-time multi-channel monitoring of burst-suppression using neural network technology during
pediatric status epilepticus treatment. Clin Neurophysiol (2016), http://dx.doi.org/10.1016/j.clinph.2016.05.358

460
461
462
463

CLINPH 2007856

No. of Pages 12, Model 5G

14 June 2016
8

C. Papadelis et al. / Clinical Neurophysiology xxx (2016) xxxxxx

Fig. 6. Confusion matrix (a) and ROC curve (b) for the NN configuration showing the best performance. In the CM, the total percent of correctly classified was 94% and the
total percent of the misclassified cases was 6%. Also ROC showed perfect classification, since all the cases were above the diagonal and close to the left-top corner.

464
465
466
467
468

3% of the B segments were misclassified as Suppression. The number of B and S segments that were misclassified as S were minimal
0.1%. In our NN, all the cases were above the diagonal in the ROC
curve and close to the top left corner; this finding indicates an
excellent classification of segments by our NN.

469

3.7. Spatiotemporal analysis

470

By using a topoplot display, we plotted the BSI values estimated

for each bipolar EEG channel. Fig. 7 presents the multichannel EEG
recordings from two patients (patients 2 and 8) with distinct BS
patterns. Note remarkable differences in the BS patterns between
channels highlighted with green color. The inset figure displays
the distribution of BSI across the scalp using a color-coded mapping: red colors indicate higher suppression activity vs. green colors indicating higher burst activity. We observed prominent
evidence of asynchronous patterns across scalp in 6 patients (22
datasets). Significant differences were observed between the BSI
values estimated using the two different montage configurations:
two bipolar channels (i.e., F3-C3 and F4-C4), and the full bipolar
EEG sensor array (df = 1, F = 7.92, p = 0.007).

471
472
473
474
475
476
477
478
479
480
481
482

483

3.8. Fixed duration segments vs. manual annotations

484
486

Analysis between the BSI values estimated by manual annotations (using variable segment lengths) and the NN revealed no statistically significant differences (two tail; p = 0.535).

487

4. Discussion

488

The treatment of pediatric SE patients in the ICU who require

continuous titration of anesthetic agents uses BSI monitoring,
which is a measure of the relative duration of suppression in the
BS patterns of the EEG signal. Current practice requires the visual
identification of these patterns by the ICU staff who manually
assesses the rough timing between one or more interburst intervals and gives a guesstimate of the BSI. Thus, the titration of anesthetic drugs in the treatment of SE patients is not based on a truly
quantitative reproducible estimation of the BSI, rather on a
guesstimate that is prone to errors. Moreover, the estimation of
BSI is often based on single channel recordings assuming that the

485

489
490
491
492
493
494
495
496
497
498

activity it records represents the activity from the entire brain. This
is in contrast to most recent evidence showing that episodes of B
could be substantially asynchronous across the cortex (Lewis
et al., 2013), and to our results showing significant differences in
the BSI values when the entire EEG sensor array is used in the estimates compared to when two channels are used. Here, we demonstrate the development of an on-line system that can automatically
detect the BS patterns in multichannel EEG recordings, and estimate the BSI considering information from all the channels. The
system was designed using NN technology, and was tested and validated on long recordings (159 h and 18 min) from 12 pediatric
patients with SE. It shows excellent agreement between the automated and manual classification of BS segments and requires minimal memory and computation time.
Compared to previously proposed solutions, our system adds
new value in the following aspects: (i) it considers the local
dynamics of BS, (ii) it uses NN technology allowing tailoring the
system to detect BS patterns from different patient populations,
anesthetic drugs, montages, and medical conditions having potentially different features, (iii) it can be tested using very large datasets, (iv) it detects and classifies automatically segments
contaminated by artifacts, and (v) it does not require any parameter to optimize its performance.

499

4.1. System development and performance

522

Many methods of varying complexity have been previously

described for the detection of BS patterns. Early attempts were
based on power or spectral analysis (Thomsen et al., 1991;
Akrawi et al., 1996; Yoon et al., 2012), while later more complicated methods were based on a variety of advanced signal processing techniques including the bispectral method (Muthuswamy
et al., 1999; Bruhn et al., 2000a), information theory, non-linear
time series analysis and non-linear energy operator (Bruhn et al.,
2000b; Srkel et al., 2002; Sherman et al., 1997; Kaiser, 1990;
Westover et al., 2013), adaptive Hilbert transform filters (Arnold
et al., 1996), recurrence quantification analysis (Liang et al.,
2014) and NN classification methods (Leistritz et al., 1999;
Griessbach et al., 1997; Lfhede et al., 2007). More recently, the
burst-suppression probability (BSP) was proposed as an alternative
approach to model the level of BS (Chemali et al., 2013). As a next
step, closed-looped systems for the automated delivery of anesthe-

523

Please cite this article in press as: Papadelis C et al. Real-time multi-channel monitoring of burst-suppression using neural network technology during
pediatric status epilepticus treatment. Clin Neurophysiol (2016), http://dx.doi.org/10.1016/j.clinph.2016.05.358

500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521

524
525
526
527
528
529
530
531
532
533
534
535
536
537
538

CLINPH 2007856

No. of Pages 12, Model 5G

14 June 2016
C. Papadelis et al. / Clinical Neurophysiology xxx (2016) xxxxxx

Fig. 7. Multichannel EEG recordings (14 bipolar channels) from two representative patients and the corresponding topoplot BSI maps. Green vertical boxes indicate
prominent differences in the BS patterns across channels. Inset figure indicates the spatial distribution of BSI over the scalp (the degree of suppression is color-coded: blue
indicates low, and red indicates higher suppression activity). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this
article.)
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565

sia in animals have been proposed to deliver propofol (Vijn and

Sneyd, 1998; Ching et al., 2013) and etomidate (Cotten et al.,
2011) by using the BSI as the control signal.
Extensive and pioneer work in the field has been done by a
research group at Massachusetts General Hospital that focuses
mostly on adults (Ching et al., 2013; Chemali et al., 2013;
Shanechi et al., 2013; Westover et al., 2013). However, methods,
which were validated under different conditions or in different
patient population, do not, a priori, be assumed to generalize to
the pediatric SE ICU setting. Here, we expand the work of this
group with a focus on the pediatric population. In humans, only
a single study (Westover et al., 2013) has described the development of a real-time algorithm to automatically discriminate periods of S from non-suppression (B) in EEG from critically ill adult
patients. Our on-line system enables automatic detection of BS
patterns and the estimation of BSI in children with SE having an
excellent performance (even without patient-by-patient tuning)
in the classification of segments (into B, S, and A) with an overall
accuracy of 94% in the confusion matrix (see Fig. 6). The NN with
the 10 nodes in the hidden layer showed an excellent overall sensitivity and specificity for the detection of B and S segments (92.6%
and 92.81%) when trained with the LM algorithm. Simulation of the
on-line setup showed also excellent classification results. Compared to the only other real-time human BS detection algorithm
(Westover et al., 2013), our system offers comparable sensitivity
and specificity for the detection of B (96% and 90% compared to
97% and 93%) and S (90% and 96% compared to 93% and 97%) seg-

ments. However, our system was validated by using a large

amount of information (10 times larger data than in Westover
et al., 2013, and 100 times larger than in the off-line algorithm
of Srkel et al., 2002). It also dealt with the issue of automated
detection and rejection of artifacts, though its validation was based
on preliminary data. The training of our NN required minimal
memory and computation time (< 2 h for 55 datasets from 12 different patients).
A potential limitation of our system is the fact that it uses fixed
duration segments (i.e. 500 ms) to localize segment boundaries.
Compared to adaptive segmentation methods where the EEG signal is divided into quasistationary epochs of variable lengths, fixed
duration segments approach may lead to the misclassification of
some portion of the BS pattern. However, so minor discrepancies
will not affect significantly the reliable estimation of BSI, if we consider the fact that the patient monitoring lasts for hours (or even
days) and the fixed duration of the EEG segment is only 500 ms.
Indeed, our statistical analysis between the manual annotations
considering variable lengths of burst and suppression and the NN
classification did not reveal significant differences (see
Section 3.8.).

566

4.2. NN technology gives flexibility to the system to be tailored for

different conditions

587

The BS signal is very unstationary, consisting of, for example,

changes in amplitudes, and oscillations with different frequencies

589

Please cite this article in press as: Papadelis C et al. Real-time multi-channel monitoring of burst-suppression using neural network technology during
pediatric status epilepticus treatment. Clin Neurophysiol (2016), http://dx.doi.org/10.1016/j.clinph.2016.05.358

567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586

588

590

CLINPH 2007856

No. of Pages 12, Model 5G

14 June 2016
10

C. Papadelis et al. / Clinical Neurophysiology xxx (2016) xxxxxx

616

along the pattern. It also varies in terms of frequency and amplitude across different drugs and different pathological conditions.
For example, animal studies have found that different drugs induce
very different patterns of BS (Kenny et al., 2014). In rodents,
remarkable differences were observed between B and S durations
and peak-to-peak amplitudes between propofol, etomidate,
thiopental, and isoflurane. In rabbits, differences were observed
between propofol and isoflurane. BS patterns also vary widely in
amplitude and spectral characteristics in different pathological
conditions (Srkel et al., 2002; Niedermeyer et al., 1999;
Niedermeyer, 2009). Different patterns of BS were observed in
our study (periodic discharges, patterns with distinct BS, and patterns with indistinct BS) (see Fig. 3). Our system showed an excellent performance (in terms of sensitivity and specificity) for the
detection of BS patterns independently of the observed pattern
(no differences were observed between datasets with different patterns). This could be due to the fact that the NN considers information from the amplitude, the frequency content, and the
complexity of the EEG signal. Thus, the full spectrum of information that can be extracted from a signal was used for the classification of segments to B, S, and A. This highlights the flexibility of our
system in the classification of EEG segments with different signal
features. Our system can be potentially trained using even larger
datasets of annotated recordings from specific patient populations.
Thus, it can be tailored to offer excellent performance for these
specific patient groups and medical conditions.

617

4.3. Spatiotemporal analysis

618

637

Previously proposed methods for the automatic detection of BS

patterns have dealt with the EEG as a single channel usually
recorded from central areas assuming that it represents the activity
of the entire brain. Our spatiotemporal analysis results indicate
that such an assumption ignores EEG focalities in the BS patterns
(see Fig. 7). Statistical analysis of the BSI values when only two
bipolar channels were used in the BSI estimation showed significant different results compared to when the entire sensor array
was used in the estimation. This indicates that few channel recordings are not representative of a global BS state. It is in line with
recent findings from intracranial electrocortigrams showing that
the BS patterns could be substantially asynchronous across the cortex (Lewis et al., 2013). Previous studies have also suggested the
use of a combination of multichannel EEGs and two channel monitoring (Tasker et al., 1989), since they found marked inter-patient
variability in the cortical activity of infants and young children
with SE treated with thiopentone. Whether performing multichannel EEG recordings as a regular practice for monitoring SE pediatric
patients is beneficial compared to the single (or few) channel
approach, it remains to be demonstrated in the clinical setup.

638

4.4. Detection of artifacts

639

Every automatic EEG analysis is easily corrupted by artifactual

electrical potentials. In apparatus without continuous monitoring
and indication of such artifacts there is always a possibility of serious misinterpretation of results, particularly when records are
obtained automatically without continuous supervision. In the
continuous monitoring of SE patients, technical artifacts are of
greater concern rather than the myogenic and blinking artifacts.
This is because, myogenic artifacts tends to be minimal during
ICU EEG recordings of BS patterns, since the patient is in a state
of profound coma where lies down motionless (Ching et al.,
2012a,b; Niedermeyer, 2009; Brown et al., 2011). However, technical artifacts are unavoidable in a busy clinical environment as ICU.
For example, artifacts caused by electrode movements can be misclassified as B and affect significantly the BSI estimation. Artifact

591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615

619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636

640
641
642
643
644
645
646
647
648
649
650
651
652

detection or rejection in BS has been rarely explored on previous

publication on automated segmentation of burst-suppression
(Srkel et al., 2002). Our system is able to automatically identify
artifactual segments from multiple channel recordings with relatively high sensitivity and specificity and ignore them in the continuous estimation of the BSI. However, it is important to
highlight that the ability of our algorithm to detect the artifacts
is based on preliminary data, since the training of our system
was based on data containing remarkably less segments with artifacts (0.4%) compared to segments with B (77%) or S (22.6%).
Thus, we expect that the NN training for the detection of A segments can be further dramatically improved if it will be trained
by using more datasets contaminated with different kind of
artifacts.

653

5. Conclusions

667

We propose the development of a real-time validated system

that has the potential to assist ICU attendants with continuous
monitoring of BS in pediatric patients with SE. The system shows
excellent agreement between the automated and the manual classification of BS segments and requires minimal memory and computation time. The system can potentially be used also in basic
research applications as well as in monitoring patients with a wide
range of pathologies who either require medical-induced coma or
present patterns of BS in their EEG signal, such as in patients suffering from traumatic brain injury with elevated intracranial pressures, as well as patients suffering from severe depression.

668

Acknowledgments

679

This work has been supported by the Boston Childrens Hospital

Innovestment Grant (Funding Cycle 2014) Papadelis (PI). Preliminary data of this work were presented at the American Clinical
Neurophysiology Society (ACNS) 2014 Annual Meeting, and at
the pediatric highlights of the American Epilepsy Society (AES)
2014 Annual Meeting. We would like to thank Mrs. Eleonora
Tamilia, Madelyn Rubenstein, and Saba Jafarpour for their help
with retrieving patients data.
Conflict of interest: None of the authors have potential conflicts
of interest to be disclosed.

680

Appendix A

690

A.1. SCG algorithm

691

The SCG algorithm (Moller, 1993) is a supervised learning algorithm that is based upon a class of optimization techniques known
in numerical analysis as the Conjugate Gradient Methods. The SCG
trains the NN by adjusting its weights. Since SCG does not perform
a line search at each iteration (which is usually a time consuming
process), it is faster than other algorithms such as the LM, BR and
BFGS.

692

A.2. LM algorithm

699

The LM algorithm (Marquardt, 1963) also trains the NN by

updating the weight and bias values based on the LevenbergMarquardt optimization. The LM is an estimation to Newtons method
and although it requires more memory compared to other algorithms, it is particularly recommended as the first-choice supervised learning algorithm for moderate-sized networks.

700

Please cite this article in press as: Papadelis C et al. Real-time multi-channel monitoring of burst-suppression using neural network technology during
pediatric status epilepticus treatment. Clin Neurophysiol (2016), http://dx.doi.org/10.1016/j.clinph.2016.05.358

654
655
656
657
658
659
660
661
662
663
664
665
666

669
670
671
672
673
674
675
676
677
678

681
682
683
684
685
686
687
688
689

693
694
695
696
697
698

701
702
703
704
705

CLINPH 2007856

No. of Pages 12, Model 5G

14 June 2016
C. Papadelis et al. / Clinical Neurophysiology xxx (2016) xxxxxx
706

A.3. BR algorithm

707

714

The BR algorithm (MacKay, 1992) is a training function which

updates weight and bias values similar to LM algorithm based on
LevenbergMarquardt optimization. Within the LevenbergMarquardt algorithm, there is a process called regularization to
improve the network generalization. BR uses some modification
in the performance function to produce a network that generalizes
properly. It basically reduces a linear combination of weights and
squared errors and estimates the accurate combination.

715

A.4. BFGS algorithm

716

723

The BFGS quasi-Newton backpropagation algorithm (BFG)

(Dennis and Schnabel, 1983) trains the NN by updating weight
and bias values according to the BFGS quasi-Newton method. Usually Newtons method is used as an alternative to the conjugate
gradient methods for fast optimization. Unlike SCG, BFGS performs
a line search in each iteration, so it operates more computation in
each iteration and it requires more memory compared to the conjugate gradient algorithms.

724

A.5. RP algorithm

725

729

The Resilient backpropagation (RP) algorithm (Riedmiller and

Braun, 1993) is a training function that updates weight and bias
values according to the resilient backpropagation algorithm.
Besides the LevenbergMarquardt algorithm, RP is one of the fastest weight update processes.

730

Appendix B. Supplementary data

731
733

Supplementary data associated with this article can be found, in

the online version, at http://dx.doi.org/10.1016/j.clinph.2016.05.
358.

734

References

735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769

Akrawi WP, Drummond JC, Kalkman CJ, Patel PM. A comparison of the
electrophysiologic characteristics of EEG burst-suppression as produced by
isoflurane, thiopental, etomidate, and propofol. J Neurosurg Anesthesiol
1996;8:406.
http://dx.doi.org/10.1097/00008506-199601000-00010
[PubMed].
Arnold M, Doering A, Witte H, Drschel J, Eisel M. Use of adaptive Hilbert
transformation for EEG segmentation and calculation of instantaneous
respiration rate in neonates. J Clin Monit 1996;12:4360 [PubMed].
Brenner RP. The interpretation of the EEG in stupor and coma. Neurologist
2005;11:27184.
Brown EN, Purdon PL, Van Dort CJ. General anesthesia and altered states of arousal:
a systems neuroscience analysis. Annu Rev Neurosci 2011;34:60128
[PubMed].
Bruhn J, Bouillon TW, Shafer SL. Bispectral index (BIS) and burst suppression:
revealing a part of the BIS algorithm. J Clin Monit Comput 2000a;16:5936
[PubMed].
Bruhn J, Rpcke H, Rehberg B, Bouillon T, Hoeft A. Electroencephalogram
approximate entropy correctly classifies the occurrence of burst suppression
pattern as increasing anesthetic drug effect. Anesthesiology 2000b;93:9815.
http://dx.doi.org/10.1097/00000542-20001000000018 [PubMed].
Chemali J, Ching S, Purdon PL, Solt K, Brown EN. Burst suppression probability
algorithms: state-space methods for tracking EEG burst suppression. J Neural
Eng 2013;10:056017 [PubMed].
Chin RF, Neville BG, Scott RC. A systematic review of the epidemiology of status
epilepticus. Eur J Neurol 2004;11:80010 [PubMed].
Chin RF, Neville BG, Peckham C, Bedford H, Wade A, Scott RC. Incidence, cause, and
short-term outcome of convulsive status epilepticus in childhood: prospective
population-based study. Lancet 2006;368:2229. http://dx.doi.org/10.1016/
S0140-6736(06)69043-0 [PubMed].
Ching S, Purdon PL, Vijayan S, Kopell NJ, Brown EN. A neurophysiological-metabolic
model for burst suppression. Proc Natl Acad Sci USA 2012a;109:3095100.
Ching S, Purdon PL, Vijayan S, Kopell NJ, Brown EN. A neurophysiological-metabolic
model for burst suppression. Proc Natl Acad Sci USA 2012b;109:3095100.
Ching S, Liberman MY, Chemali JJ, Westover MB, Kenny JD, Solt K, et al. Real-time
closed-loop control in a rodent model of medically induced coma using burst

708
709
710
711
712
713

717
718
719
720
721
722

726
727
728

732

11

suppression. Anesthesiology 2013;119:84860. http://dx.doi.org/10.1097/

ALN.0b013e31829d4ab4 [PubMed].
Claassen J, Lokin JK, Fitzsimmons B-FM, Mendelsohn FA, Mayer SA. Predictors of
functional disability and mortality after status epilepticus. Neurology
2002;58:13942. http://dx.doi.org/10.1212/WNL.58.1.139 [PubMed].
Clark DL, Rosner BS. Neurophysiologic effects of general anesthetics. I. The
electroencephalogram and sensory evoked responses in man. Anesthesiology
1973;38:56482 [PubMed].
Cotten JF, Le Ge R, Banacos N, Pejo E, Husain SS, Williams JH, et al. Closed-loop
continuous infusions of etomidate and etomidate analogs in rats: a comparative
study of dosing and the impact on adrenocortical function. Anesthesiology
http://dx.doi.org/10.1097/ALN.0b013e31821950de
2011;115:76473.
[PubMed].
Cottenceau V, Petit L, Masson F, Guehl D, Asselineau J, Cochard J-F. The use of
bispectral index to monitor barbiturate coma in severely brain-injured patients
with refractory intracranial hypertension. Anesth Analg 2008;107:167682
[PubMed].
Patterson
Dan
W.
Artificial
Neural
Networks,
Theory
and
Applications. Singapore: Prentice Hal; 1996.
DeGiorgio DM. Burst-suppression patterns in status epilepticus. The burstsuppression duration ratio: a marker of the end of clinical status epilepticus. J
Epileps 1993;6:325.
DeLorenzo RJ, Hauser WA, Towne AR, Boggs JG, Pellock JM, Penberthy L. A
prospective, population-based epidemiologic study of status epilepticus in
Richmond, Virginia. Neurology 1996;46:102935 [PubMed].
Dennis JE, Schnabel RB. Numerical methods for unconstrained optimization and
nonlinear equations. Englewood Cliffs 1983.
Dinh C, Luessi M, Sun L, Haueisen J, Hamalainen MS. Mne-X: MEG/EEG real-time
acquisition, real-time processing, and real-time source localization framework.
Biomed Tech (Berl) 2013;58:4184. http://dx.doi.org/10.1515/bmt-2013-4184.
Geman S, Bienenstock E, Doursat R. Neural networks and the bias/variance
dilemma. Neural Comput 1992;4:158.
Griessbach G, Eiselt M, Drschel J, Witte H, Galicki M. Common optimization of
adaptive preprocessing units and a neural network during the learning period.
Application in EEG pattern recognition. Neural Netw 1997;10:115363.
Hirsch LJ, LaRoche SM, Gaspard N, Gerard E, Svoronos A, Herman ST, Mani R, Arif H,
Jette N, Minazad Y, Kerrigan JF, Vespa P, Hantus S, Claassen J, Young GB, So E,
Kaplan PW, Nuwer MR, Fountain NB, Drislane FW. American Clinical
Neurophysiology Societys Standardized Critical Care EEG Terminology: 2012
version. J Clin Neurophysiol 2013;30::127 [PubMed].
Kaiser J.F. On a simple algorithm to calculate the energy of a signal. Proc
International Conference on Acoustics, Speech, and Signal Processing (ICASSP
90), IEEE, 1990; 1: 381384. doi:10.1109/ICASSP.1990.115702.
Kenny JD, Westover MB, Ching S, Brown EN, Solt K. Propofol and sevoflurane induce
distinct burst suppression patterns in rats. Front Syst Neurosci 2014;8:237.
http://dx.doi.org/10.3389/fnsys.2014.00237.
Leistritz L, Jger H, Schelenz C, Witte H, Putsche P, Specht M, et al. New approaches
for the detection and analysis of electroencephalographic burstsuppression
patterns in patients under sedation. J Clin Monit Comput 1999;15:35767.
Lewis LD, Ching S, Weiner VS, Peterfreund RA, Eskandar EN, Cash SS, Brown EN,
Purdon PL. Local cortical dynamics of burst suppression in the anaesthetized
brain. Brain 2013;136:272737.
Liang Z, Wang Y, Ren Y, Li D, Voss L, Sleigh J, Li X. Detection of burst suppression
patterns in EEG using recurrence rate. Sci World J 2014;2014:295070.
Lfhede J, Lfgren N, Thordstein M, Flisberg A, Kjellmer I, Lindecrantz K. Comparison
of three methods for classifying burst and suppression in the EEG of post
asphyctic newborns. Conf Proc IEEE Eng Med Biol Soc 2007;2007:51369.
Logroscino G, Hesdorffer DC, Cascino G, Annegers JF, Hauser WA. Time trends in
incidence, mortality, and case-fatality after first episode of status epilepticus.
Epilepsia 2001;42:10315.
MacKay David J.C. Bayesian Interpolation. Neural Comput. 1992; 4: 41547.
Marquardt D. An Algorithm for Least-Squares Estimation of Nonlinear Parameters.
SIAM J Appl Math 1963;11:43141.
Moller MF. A scaled conjugate gradient algorithm for fast supervised learning.
Neural Netw 1993;6:52533.
Musialowicz T, Mervaala E, Klviinen R, Uusaro A, Ruokonen E, Parviainen I. Can
BIS monitoring be used to assess the depth of propofol anesthesia in the
treatment of refractory status epilepticus? Epilepsia 2010;51:15806.
Muthuswamy J, Sherman DL, Thakor NV. Higher-order spectral analysis of burst
patterns in EEG. IEEE Trans Biomed Eng 1999;46:929.
Niedermeyer E, Sherman DL, Geocadin RJ, Hansen HC, Hanley DF. The burst
suppression electroencephalogram. Clinical EEG 1999;30:99105.
Niedermeyer
E.
The
burstsuppression
electroencephalogram.
Am
J
Electroneurodiagnostic Technol 2009;49:33341.
Riedmiller M, Braun H. A direct adaptive method for faster backpropagation
learning: The RPROP algorithm. Proc IEEE ICNN 1993:58691.
Riker RR, Fraser GL, Wilkins ML. Comparing the bispectral index and suppression
ratio with burst suppression of the electroencephalogram during pentobarbital
infusions in adult intensive care patients. Pharmacotherapy 2003;23:108793.
Rossetti AO, Milligan TA, Vullimoz S, Michaelides C, Bertschi M, Lee JW. A
randomized trial for the treatment of refractory status epilepticus. Neurocrit
Care 2011;14:410.
Srkel M, Mustola S, Seppnen T, Koskinen M, Lepola P, Suominen K, Juvonen T,
Tolvanen-Laakso H, Jntti V. Automatic analysis and monitoring of burst
suppression in anesthesia. J Clin Monit Comput 2002;17:12534.

Please cite this article in press as: Papadelis C et al. Real-time multi-channel monitoring of burst-suppression using neural network technology during
pediatric status epilepticus treatment. Clin Neurophysiol (2016), http://dx.doi.org/10.1016/j.clinph.2016.05.358

770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854

CLINPH 2007856

No. of Pages 12, Model 5G

14 June 2016
12
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872

C. Papadelis et al. / Clinical Neurophysiology xxx (2016) xxxxxx

Shanechi MM, Chemali JJ, Liberman M, Solt K, Brown EN. A brain-machine interface
for control of medically-induced coma. PLoS Comput Biol 2013;9:e1003284.
http://dx.doi.org/10.1371/journal.pcbi.1003284.[PubMed].
Shannon CE. A Mathematical Theory of Communication. Bell Syst Tech J
1948;27:379423.
Sherman D.L., Brambrink A.M., Walterspacher D., Dasika V.K., Ichord R., Thakor N.V.
Detecting EEG bursts after hypoxic-ischemic injury using energy operators. Proc
19th IEEE EMBS Conference. 1997; 11881190.
Tasker RC, Boyd SG, Harden A, Matthew DJ. EEG monitoring of prolonged
thiopentone administration for intractable seizures and status epilepticus in
infants and young children. Neuropediatrics 1989;20:14753. http://dx.doi.org/
10.1055/s-2008-1071281 [PubMed].
Tasker RC, Vitali SH. Continuous infusion, general anesthesia and other intensive
care treatment for uncontrolled status epilepticus. Curr Opin Pediatr
2014;26:6829.
Thomsen CE, Rosenfalck A, Nrregaard Christensen K. Assessment of anaesthetic
depth
by
clustering
analysis
and
autoregressive
modelling
of
electroencephalograms. Comput Methods Programs Biomed 1991;34:12538.

Trinka E, Cock H, Hesdorffer D, Rossetti AO, Scheffer IE, Shinnar S, et al. A definition
and classification of status epilepticus Report of the ILAE Task Force on
Classification of Status Epilepticus. Epilepsia 2015;56:151523. http://dx.doi.
org/10.1111/epi.13121 [PubMed].
Vijn PC, Sneyd JR. Anaesthesia and EEG burst suppression in rats: Bolus injections
and closed-loop infusions. Br J Anaesth 1998;81:41521.
Westover MB, Shafi MM, Ching S, Chemali JJ, Purdon PL, Cash SS, et al. Real-time
segmentation of burst suppression patterns in critical care EEG monitoring. J
Neurosci
Methods
2013;219:13141.
http://dx.doi.org/10.1016/j.
jneumeth.2013.07.003 [PubMed].
Wilkes R, Tasker RC. Pediatric intensive care treatment of uncontrolled status
epilepticus. Crit Care Clin 2013;29:23957.
Wilkes R, Tasker RC. Intensive care treatment of uncontrolled status epilepticus in
children: systematic literature search of midazolam and anesthetic therapies.
Pediatr Crit Care Med 2014;15:6329.
Yoon JR, Kim YS, Kim TK. Thiopental-induced burst suppression measured by the
bispectral index is extended during propofol administration compared with
sevoflurane. J Neurosurg Anesthesiol 2012;24:14651 [PubMed].
Young GB. The EEG in coma. J Clin Neurophysiol 2000;17:47385 [PubMed].

873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892

Please cite this article in press as: Papadelis C et al. Real-time multi-channel monitoring of burst-suppression using neural network technology during
pediatric status epilepticus treatment. Clin Neurophysiol (2016), http://dx.doi.org/10.1016/j.clinph.2016.05.358