Handling Unbalanced Data in Nocturnal Epileptic Seizure Detection

using Accelerometers

Kris Cuppens

1,2

, Peter Karsmakers

1,2

, Anouk Van de Vel

, Bert Bonroy

, Milica Milosevic

Lieven Lagae

, Berten Ceulemans

, Sabine Van Huffel

and Bart Vanrumste

1,2

MOBILAB, Thomas More Kempen, Geel, Belgium

iMinds Future Health Departement, Biomed, SCD SISTA, Electrical Engineering (ESAT), KU Leuven, Leuven, Belgium

Department of Paediatrics, University Hospital Leuven, Leuven, Belgium

University Hospital Leuven, Leuven, Belgium

Keywords:

Epilepsy Detection, Acceleration Data, Unbalanced Data, Support Vector Machines.

Abstract:

Data of nocturnal movements in epileptic patients is marked by an imbalance due to the relative small number

of seizures compared to normal nocturnal movements. This makes developing a robust classiﬁer more difﬁcult,

especially with respect to reducing the number of false positives while keeping a high sensitivity. In this paper

we evaluated different ways to overcome this problem in our application, by using a different weighting of

classes and by resampling the minority class. Furthermore, as we only have a limited number of training

samples available per patient, additionally it was investigated in which manner the training set size affects

the results. We observed that oversampling gives a higher performance than only adjusting the weights of

both classes. Compared to its alternatives oversampling based on the probability density function gives the

best results. On 2 of 3 patients, this technique gives a sensitivity of 95% or more and a PPV more than 70%.

Furthermore, an increased imbalance in the dataset leads to lower performance, whereas the size of the dataset

has little inﬂuence.

1 INTRODUCTION

Epileptic seizures mainly occur as paroxysmal events,

which means that they occur at a sudden, unexpected

timing. The frequency of the seizures varies from pa-

tient to patient.

25% to 30% of epileptic patients can not be treated

by medication or surgery, they suffer from so called

refractory epilepsy (Chapman et al., 2011) (Dalton

et al., 2010). To be able to track the progress of the

disease and to alarm caregivers during a seizure, these

patients should be monitored.

The type of epileptic seizures we focus on are hy-

permotor seizures, which manifest themselves as vio-

lent, uncontrolled movements of the arms and or legs,

e.g. by making a pedalling movement (Azar et al.,

2008)(R

emi et al., 2011). The movements can last

from a couple of seconds to multiple minutes. Due

to the possible heavy movements the patients can in-

jure themselves or even die (Husain and Sinha, 2011).

Patients may suffer from confusion after a seizure,

and when they don’t, they often recall the seizure

as a ’strange feeling’ and need comforting (Tinuper

et al., 2005)(Tinuper et al., 2007). Therefore our aim

is to develop an automated system which can detect

seizures during sleep to alarm the parents or care-

givers if a seizure occurs. A second aim is to keep

track of the number of seizures the patient has during

the night.

Much more normal movements occur compared

to seizure data which leads to very unbalanced data

when modeling the movements of these patients. In

this paper we compare techniques compensating for

the highly unbalanced seizure data, to generate a

classiﬁcation model that is able to detect epileptic

seizures. The patient group we focus on are chil-

dren as the prevalence of epilepsy in children is higher

and parents want to know when their child has a

seizure during the night. Furthermore during sleep

the convulsions occur more or less in a controlled re-

producible manner without too much interference of

other noise sources such as voluntary movement. For

recording the motion, the acceleration of the arms and

legs is recorded using 3D accelerometers. We want

to have a high sensitivity (preferably close to 100%)

with as little false positives as possible. But for our

447

Cuppens K., Karsmakers P., Van de Vel A., Bonroy B., Milosevic M., Lagae L., Ceulemans B., Van Huffel S. and Vanrumste B. (2013).

Handling Unbalanced Data in Nocturnal Epileptic Seizure Detection using Accelerometers.

In Proceedings of the 2nd International Conference on Pattern Recognition Applications and Methods, pages 447-452

DOI: 10.5220/0004264704470452

 SciTePress

Table 1: Overview of data available in the dataset.

Patient Nights of Hypermotor Normal

number monitoring seizures movements

A 2 9 287

B 5 26 784

C 2 7 381

Total 9 42 1452

application a high sensitivity is more important, as we

do not want to miss a possible violent movement.

To overcome the issue of imbalance, we evaluated

the following approaches

• use a weighting factor to train the SVM model,

• resample the training data by estimating the prob-

ability density function of the seizure data,

• resample the seizure data based on the SMOTE

technique (Chawla et al., 2002).

Note that undersampling of the majority class was

not considered due to the very limited number of mi-

nority class samples that were available. Furthermore

we evaluated the linear and the radial basis function

(RBF) kernel and the inﬂuence of the balance and the

number of training points on the performance of our

classiﬁer.

He et al. (He, 2009) give an overview of the state-

of-the-art techniques that are used to overcome the is-

sue of imbalance. Next to the sampling methods that

we discuss in our paper, cost-sensitive methods that

give a different cost value to misclassifying instances

of different classes, and kernel-based methods are re-

viewed. But the former two dominate the research

efforts.

2 METHODS

2.1 Dataset Overview

In this study, we use a dataset of patients suffer-

ing from hypermotor seizures, similar to what we al-

ready described in previous studies (Cuppens et al.,

2009)(Cuppens et al., 2012), but in this study we only

selected patients with at least 4 hypermotor seizures

in our dataset, because for training and testing of the

classiﬁcation model, we need at least 3 and 1 seizures

respectively. An overview of the dataset is shown in

Table 1. From the raw data, ﬁve features are extracted

to represent the data: the maximal resultant over both

arms, the mean standard deviation over all channels,

the length of the epoch, the mean root mean square

over all the channels and the minimal power in the

frequency band 1-3 Hz over all the channels.

The acquired data contains EEG and video which

is used for labeling of the data by the neurologists,

and 12 channel acceleration data sampled at 250 Hz

measured with 3D accelerometers at the four limbs.

Also ECG, EMG, EOG and audio are recorded, but

these modalities are not used in this study.

2.2 Test Outline

Different classiﬁcation models are estimated using

state-of-the-art Support Vector Machine (SVM) in-

ference. For this purpose the freely available libsvm

(Chang and Lin, 2011) implementation was adopted

Two groups of related application speciﬁc tests were

setup in order to investigate: a) which alternative to

adjust model inference for the high class imbalance

performed best, b) the inﬂuence of the degree of class

imbalance on the classiﬁcation results c) the inﬂuence

of the training set size on the results.

All the tests we perform, are conducted in a 10

fold randomization with different combinations of

seizures and normal movements. The tests are evalu-

ated using sensitivity (measure that indicates the per-

centage of seizures that is detected, also indicated as

recall), the positive predictive value (PPV, a measure

that indicates the percentage of true detections over all

detections, also indicated as precision) and speciﬁcity

(measure that indicates the number of true negatives

over all normal movements). The sensitivity and PPV

values are averaged over the 10 randomizations and

the standard deviation is calculated. Thanks to this

randomization the obtained results are less biased.

2.2.1 Methods for Coping with Class Imbalance

To overcome the issue of imbalance, we explored:

a) different weight factors the ﬁrst applied to the er-

ror term in the SVM learning objective correspond-

ing to the positive examples and the second to that

corresponding to the examples of the negative class.

For this purpose a weight factor T is introduced

which balances the weights of the error terms for both

classes in the SVM objective. b) Resampling of the

minority class, using 1) sampling from a probability

density function estimated using the available minor-

ity examples 2) generating new minority class exam-

ples using the SMOTE technique.

Resampling based on the estimation of the proba-

bility density function estimates the distribution of the

seizure data points along every feature using a non-

parametric kernel density estimation, with a Gaussian

http://www.csie.ntu.edu.tw/˜cjlin/libsvm/

ICPRAM2013-InternationalConferenceonPatternRecognitionApplicationsandMethods

448

function as kernel and a bandwidth equal to the stan-

dard deviation of the seizure data. Based on this esti-

mated distribution, new points are randomly sampled.

The SMOTE technique generates new data points by

introducing synthetic examples along the line seg-

ments that connect minority class examples with their

k minority class nearest neighbors.

In a ﬁrst series of tests, we used the same number

of seizures s and normal movements n for each pa-

tient in the training (s = 4, n = 190) and testing phase

(s = 2, n = 95), to investigate the performance of the

different approaches, and to be able to compare them

over all the patients. These movements were ran-

domly chosen from each patient but with a strict sepa-

ration between training and test set. We used this spe-

ciﬁc number of normal movement and seizures based

on the smallest numbers available for each patient, i.e.

7 seizures for patient C and 287 normal movements

for patient A. For each patient, we used two thirds of

the data for training and one third for testing.

In the training phase, we used a three-fold cross-

validation to ﬁnd the optimal hyper-parameters for the

SVM classiﬁer. These parameters include the regular-

ization parameter C, kernel parameter σ and weight

factor T . C affects the trade-off between complexity

and proportion of nonseparable samples (Cherkassky

and Mulier, 1998). If it is too large, we may overﬁt

and consequently store many support vectors, if it is

too small, too much smoothing may be applied giving

an underﬁt (Alpaydin, 2004). The kernel parameter σ

denotes the kernel width parameter. A larger σ leads

to a smoother ﬁt. The weight factor T gives a lower

or higher weight to the error cost of the negative class,

in such a way that C

−

= C ·T and C

= C ·(1 −T ).

The different values we tested for the parameters are:

• C : e

−3,−2,...,6

• σ :

√

−2.5,−2,...,2

• T ∈ {0, 0.01, 0.02, 0.05, 0.1, 0.2, 0.5, 0.7, 1}

The cost function gives a higher weight to the sen-

sitivity compared to the PPV more speciﬁc: cost =

−(2 ·sens +PPV ).

2.2.2 Inﬂuence of Class Imbalance

In the second series of tests, we use the optimal ap-

proach found in the ﬁrst tests, and investigate the

inﬂuence of the number of training examples and

the balance on the performance of the classiﬁcation

model. To investigate the inﬂuence of the balance,

we keep the number of seizures in the training set

ﬁxed (4), and increase the number of normal move-

ment from 4 (equally balanced) to 190. The same ra-

tio is used in the test set.

2.2.3 Inﬂuence of Training Set Size

For the investigation on the inﬂuence of the number

of training samples, we keep the balance the same (1

seizure for 33 normal movements), and we test this for

2, 3 and 4 seizures in the training set, and 2 seizures

and 66 normal movements in the test set.

3 RESULTS

Table 2 shows the results of the SVM classiﬁer us-

ing the different methods compensating for the im-

balance. To be able to compare the different meth-

ods, we calculated a cost value for each test, using

the same cost function for model selection cost =

−(2 ·sens + PPV ). This cost value is shown in the

last column of Table 2. Notice that the cost values

have a negative value, as a lower cost implies a bet-

ter performance. The higher the absolute value, the

better the classiﬁer performs. In general, we can ob-

serve that the oversampling techniques work better

than only using a different weighting for both classes,

although this is mainly due to the low performance on

patient B. Furthermore, the density estimation over-

sampling gives better results than the SMOTE tech-

nique, although the PPV values of using SMOTE are

higher compared to the density estimation oversam-

pling. But the sensitivity of the latter is higher and

this sensitivity has a higher weight in the cost func-

tion. For this density estimation oversampling, the

linear kernel performs a little better than the RBF ker-

nel, as the absolute cost value of the former is higher.

To investigate the inﬂuence of the size of the train-

ing set and the balance in the data, we used the over-

sampling based on the probability density estimation

with the linear kernel. In the ﬁrst test, we evaluated

three different sizes of training set with the same bal-

ance, and a test set with a ﬁxed amount of data. Ta-

ble 3 shows the results of this test. We do not observe

a clear trend when a larger set of data is used. Al-

though there is a difference in performance in each

test with respect to the sensitivity and PPV for each

patient, the overall cost stays more or less the same.

Table 4 shows the results of the second test, where

we tested the inﬂuence of the balance of the dataset

on the performance. These results show that the bal-

ance of the dataset has a large inﬂuence on the per-

formance. When the number of seizures and normal

movements is equal, the performance is best, with a

sensitivity and PPV of 100.00% for patient A and C.

For patient B and for all patients combined, the values

are around 50.00% for the sensitivity and 85.00% or

higher for the PPV. But when the imbalance increases,

HandlingUnbalancedDatainNocturnalEpilepticSeizureDetectionusingAccelerometers

449

Table 2: Results of performance of SVM classiﬁer using different methods to overcome the problem of imbalance.

Balance factor

linear n

◦

sens.(%) std. PPV.(%) std.(%) spec.(%) std.(%) -cost

kernel A 80.00 25.82 93.33 14.05 99.79 0.44 253.33

B 0.00 0.00 0.00 0.00 99.37 1.66 0.00

C 80.00 42.16 89.58 19.80 99.68 0.71 249.58

all 10.00 21.08 16.67 25.82 98.95 1.57 36.67 134.90

RBF n

◦

sens.(%) std.(%) PPV.(%) std.(%) spec.(%) std.(%) -cost

kernel A 90.00 21.08 90.00 16.10 99.68 0.51 270.00

B 5.00 15.81 25.00 50.00 99.68 0.51 35.00

C 50.00 52.70 100.00 0.00 100.00 0.00 200.00

all 15.00 24.15 14.29 20.25 98.21 1.79 44.29 137.32

Probability density estimation oversampling

linear n

◦

sens.(%) std.(%) PPV.(%) std.(%) spec.(%) std.(%) -cost

kernel A 90.00 21.08 68.67 29.15 98.63 1.41 248.67

B 70.00 34.96 12.44 6.50 88.95 3.92 152.44

C 100.00 0.00 83.33 22.22 99.37 0.89 283.33

all 65.00 24.15 12.90 5.84 89.47 4.30 142.90 206.84

RBF n

◦

sens.(%) std.(%) PPV.(%) std.(%) spec.(%) std.(%) -cost

kernel A 95.00 15.81 74.00 23.98 98.95 1.11 264.00

B 60.00 39.44 12.99 10.08 91.05 4.25 132.99

C 100.00 0.00 80.00 26.99 99.05 1.44 280.00

all 45.00 36.89 19.41 13.27 95.79 3.33 109.41 196.60

SMOTE oversampling

linear n

◦

sens.(%) std.(%) PPV.(%) std.(%) spec.(%) std.(%) -cost

kernel A 90.00 21.08 83.33 22.22 99.47 0.74 263.33

B 45.00 36.89 9.84 8.18 91.68 3.27 99.84

C 85.00 33.75 87.04 20.03 99.58 0.74 257.04

all 55.00 28.38 17.22 10.35 92.21 5.39 127.22 186.86

RBF n

◦

sens.(%) std.(%) PPV.(%) std.(%) spec.(%) std.(%) -cost

kernel A 90.00 21.08 75.67 26.44 99.05 1.16 255.67

B 30.00 25.82 15.59 19.47 95.68 3.42 75.59

C 80.00 34.96 87.04 20.03 99.58 0.74 247.04

all 45.00 28.38 16.58 14.53 94.00 3.75 106.58 171.22

Table 3: Results of performance using resampling based on a density estimation and a linear kernel in the SVM classiﬁer.

This table shows the inﬂuence of different sizes of the training set.

Training: 2 seizures, 66 normal Test: 2 seizures, 66 normal

◦

sens.(%) std. PPV.(%) std.(%) spec.(%) std. -cost

A 95.00 15.81 81.67 19.95 99.09 1.06 271.67

B 65.00 33.75 18.15 9.38 90.30 6.32 148.15

C 100.00 0.00 83.33 17.57 99.24 0.80 283.33

all 70.00 34.96 22.65 11.18 92.73 3.96 162.65 216.45

Training: 3 seizures, 99 normal Test: 2 seizures, 66 normal

◦

sens.(%) std.(%) PPV.(%) std.(%) spec.(%) std.(%) -cost

A 90.00 21.08 76.67 21.08 98.79 1.20 256.67

B 75.00 35.36 14.70 7.00 87.27 2.49 164.70

C 100.00 0.00 66.67 19.25 98.18 1.20 266.67

all 65.00 33.75 22.99 14.18 92.12 5.09 152.99 210.26

Training: 4 seizures, 132 normal Test: 2 seizures, 66 normal

◦

sens.(%) std.(%) PPV.(%) std.(%) spec.(%) std.(%) -cost

A 90.00 21.08 76.67 22.50 98.94 1.02 256.67

B 70.00 34.96 15.82 11.21 86.82 7.92 155.82

C 100.00 0.00 70.00 21.94 98.33 1.33 270.00

all 70.00 25.82 21.03 12.13 89.85 4.90 161.03 210.88

we see a gradual reduction of the performance, which

is also reﬂected in the cost value for each test, which

increases from -244.72 when the dataset is balanced,

to -207.04 when the imbalance is largest (4 seizures

compared to 190 normal movements).

ICPRAM2013-InternationalConferenceonPatternRecognitionApplicationsandMethods

450

Table 4: Results of performance using resampling based on a density estimation and a linear kernel in the SVM classiﬁer.

This table shows the inﬂuence of a different balance between seizures and normal movements in the training and test set.

◦

sens.(%) std.(%) PPV.(%) std.(%) spec.(%) std.(%) -cost

Training: A 100.00 0.00 100.00 0.00 100.00 0.00 300.00

4 seizures, 4 normal B 50.00 47.14 88.89 17.21 90.00 21.08 188.89

Test: C 100.00 0.00 100.00 0.00 100.00 0.00 300.00

2 seizures, 2 normal all 45.00 28.38 100.00 0.00 100.00 0.00 190.00 244.72

Training: A 95.00 15.81 100.00 0.00 100.00 0.00 290.00

4 seizures, 10 normal B 60.00 45.95 68.75 33.85 88.00 13.98 188.75

Test: C 100.00 0.00 100.00 0.00 100.00 0.00 300.00

2 seizures, 5 normal all 55.00 36.89 74.07 35.46 92.00 10.33 184.07 240.71

Training: A 95.00 15.81 90.00 16.10 98.00 3.22 280.00

4 seizures, 30 normal B 65.00 41.16 45.67 34.10 88.00 8.78 175.67

Test: C 100.00 0.00 100.00 0.00 100.00 0.00 300.00

2 seizures, 15 normal all 65.00 33.75 63.70 29.03 92.00 9.32 193.70 237.34

Training: A 95.00 15.81 90.00 16.10 99.14 1.38 280.00

4 seizures, 70 normal B 70.00 34.96 27.21 15.16 87.14 7.77 167.21

Test: C 100.00 0.00 78.33 23.64 97.71 2.63 278.33

2 seizures, 35 normal all 65.00 24.15 48.61 23.98 94.29 6.46 178.61 226.04

Training: A 90.00 21.08 76.67 22.50 98.94 1.02 256.67

4 seizures, 132 normal B 70.00 34.96 15.82 11.21 86.82 7.92 155.82

Test: C 100.00 0.00 70.00 21.94 98.33 1.33 270.00

2 seizures, 66 normal all 70.00 25.82 21.03 12.13 89.85 4.90 161.03 210.88

Training: A 90.00 21.08 68.67 29.15 98.63 1.41 248.67

4 seizures, 190 normal B 70.00 34.96 13.16 8.36 88.95 4.16 153.16

Test: C 100.00 0.00 83.33 22.22 99.37 0.89 283.33

2 seizures, 95 normal all 65.00 24.15 13.02 5.77 89.47 4.57 143.02 207.04

4 DISCUSSION

The oversampling methods give the best results.

However, we still observe that with an increasing

unbalance, the performance gets worse. It should

be noted that thanks to the oversampling, the SVM

model is still trained with an equally balanced train-

ing set, and that only the test set remains unbalanced.

When we test the statistical signiﬁcance of the dif-

ference in methods to overcome the issue of imbal-

ance, we observe that in patient A non of the methods

or type of kernel in the SVM approach give a signif-

icant difference for either sensitivity or PPV. In pa-

tient B there is a signiﬁcant difference (with p-values

lower than 0.05) between methods and kernels, ex-

cept for the difference in PPV for the type of kernel.

For patient C only the difference in sensitivity for the

method (not the kernel) is signiﬁcant.

In most of the trials, the performance on patient B

is lower. This is due to the different clinical manifes-

tation of the seizures of this patient, which are clearly

shorter in time and lower in intensity.

Note that for the calculation of the performance

on all patients, the varying number of seizures per pa-

tient is not taken into account. This means that the

inﬂuence of patient B is higher than those of patient

A en C as 62% of the total seizures belong to patient

B (26 on a total of 42). This explains why the per-

formance of all patients combined is most of the time

very similar to the performance on patient B.

In the real setting the balance between epileptic

and normal movements can differ from patient to pa-

tient, and therefore inﬂuence the patient-speciﬁc per-

formance.

The number of seizures we collected is small for

some patients, which makes the training of an SVM

classiﬁer more difﬁcult if we want to make patient-

speciﬁc models. Although the results of the tests on

the inﬂuence of the dataset size indicated that there

is no big difference when changing the amount of

data. However, this statement needs to be interpreted

carefully since the number of seizures was too small

to draw any conclusions about the relation between

seizure set size and classiﬁcation performance.

Due to the limited number of seizure examples,

the weighting method does not give very good results.

This can be explained by the fact that the number of

the support vector candidates is small, which reduces

the ﬂexibility of setting the decision hyperplane. We

also noticed that the cost of most of the different pa-

rameter combinations (C, σ, T ) was the same when

using only the weighting of the classes. This means

that due to the limited number of training examples,

the hyper-parameters only have a small inﬂuence on

HandlingUnbalancedDatainNocturnalEpilepticSeizureDetectionusingAccelerometers

451

the decision hyperplane.

In our tests we used 2 seizures in the test set using

10 randomizations. Therefore, the resolution of the

sensitivity is only 5%. This is also visible in the tables

showing the results.

For using the SMOTE technique, we have only a

limited number of seizures in our setup when using

the 3-fold cross-validation in determining the optimal

parameters for the SVM. The new data points are gen-

erated on the line segments connecting the minority

class examples, but sometimes there is only one near-

est neighbor (for determining the line segments) for

generating new data points. This can explain why

the SMOTE technique gives lower results, although

it also works well for patient A and C.

We also evaluated a cost function taking into ac-

count the decision values of the SVM classiﬁcation

(indicating the distance from the data points to the

decision plane). However, this did not give any better

results compared to our original cost function, in most

cases the performance was even lower.

5 CONCLUSIONS

We have tested different approaches to overcome

the imbalance problem in our application of detect-

ing nocturnal epileptic seizures in children using ac-

celerometers. Oversampling of the minority class

seems to give the best results, especially the density

estimation oversampling. On 2 of 3 patients, this tech-

nique gives a sensitivity of 95% or more and a PPV

more than 70%.

ACKNOWLEDGEMENTS

Research supported by Research Council KUL:

GOA-MANET, IWT: TBM070713-Accelero, Bel-

gian Federal Science Policy Ofﬁce IUAP P6/04

(DYSCO, ’Dynamical systems, control and optimiza-

tion, 2007-2011); EU: Neuromath (COSTBM0601).

Kris Cuppens is funded by a Ph.D. grant of the

Agency for Innovation by Science and Technology

(IWT).

REFERENCES

Alpaydin, E. (2004). Introduction to Machine Learning.

MIT Press.

Azar, N. J., Tayah, T. F., Wang, L., Song, Y., and Abou-

Khalil, B. W. (2008). Postictal breathing pattern

distinguishes epileptic from nonepileptic convulsive

seizures. Epilepsia, 49(1):132–7.

Chang, C.-C. and Lin, C.-J. (2011). LIBSVM: A library

for support vector machines. ACM Transactions on

Intelligent Systems and Technology, 2:27:1–27:27.

Chapman, D., Panelli, R., Hanna, J., and Jeffs, T. (2011).

Sudden unexpected death in epilepsy: continuing the

global conversation. Camberwell, Australia.

Chawla, N., Bowyer, K., Hall, L., and Kegelmeyer, W.

(2002). SMOTE: Synthetic minority over-sampling

technique. Journal of Artiﬁcial Intelligence Research,

16:321–357.

Cherkassky, V. and Mulier, F. (1998). Learning from Data:

Concepts, Theory, and Methods. John Wiley & Sons.

Cuppens, K., Karsmakers, P., Van de Vel, A., Luca, S.,

Bonroy, B., Milosevic, M., Croonenborghs, T., Ceule-

mans, B., Lagae, L., Van Huffel, S., and Vanrumste,

B. (2012). Accelerometer based home monitoring for

detection of nocturnal hypermotor seizures based on

novelty detection. internal report.

Cuppens, K., Lagae, L., Ceulemans, B., Van Huffel, S.,

and Vanrumste, B. (2009). Detection of nocturnal

frontal lobe seizures in pediatric patients by means of

accelerometers: a ﬁrst study. In Proc. of the Annual

International Conference of the IEEE Engineering in

Medicine and Biology Society (EMBC 2009), volume

1-20, pages 6608–6611.

Dalton, A., Patel, S., Chowdhury, A., Welsh, M., Pang, T.,

Schachter, S.,

Olaighin, G., and Bonato, P. (2010).

Detecting epileptic seizures using wearable sensor

technologies. In Proc. of the First AMA-IEEE Medical

Technology Conference on Individualized Healtcare.

He, H.and Garcia, E. (2009). Learning from imbalanced

data. IEEE Transactions on Knowledge and Data En-

gineering, 21:1263–1284.

Husain, A. M. and Sinha, S. R. (2011). Nocturnal epilepsy

in adults. J Clin Neurophysiol., 28(2):141–5.

emi, J., Silva Cunha, J. P., Vollmar, C., Topc¸u

oglu, O. B.,

Meier, A., Ulowetz, S., Beleza, P., and Noachtar, S.

(2011). Quantitative movement analysis differentiates

focal seizures characterized by automatisms. Epilepsy

Behav., 20(4):642–7.

Tinuper, P., Provini, F., Bisulli, F., and Lugaresi, E. (2005).

Hyperkinetic manifestations in nocturnal frontal lobe

epilepsy. semeiological features and physiopathologi-

cal hypothesis. Neurol Sci., 26(suppl 3):210–4.

Tinuper, P., Provini, F., Bisulli, F., Vignatelli, L., Plazzi, G.,

Vetrugno, R., Montagna, P., and Lugaresi, E. (2007).

Movement disorders in sleep: guidelines for differen-

tiating epileptic from non-epileptic motor phenomena

arising from sleep. Sleep Med Rev., 11(4):255–67.

ICPRAM2013-InternationalConferenceonPatternRecognitionApplicationsandMethods

452