Comparison of Different Data Augmentation Techniques for Improving

Epileptic Seizure Detection Based on 3D Acceleration, Heart Rate and

Temperature Data

Maleyka Seyidova

1,2

, Jasmin Henze

1 a

, Arne Pelzer

3 b

and Beate Rhein

Department of Healthcare, Fraunhofer Institute for Software and Systems Engineering (ISST), Speicherstraße 6,

44147 Dortmund, Germany

Institute of Computer and Communication Technology (ICCT), Faculty of Information, Media and Electrical Engineering,

TH K

oln - Cologne University of Applied Sciences, Betzdorfer Straße 2, 50679 Cologne, Germany

Branch for Hearing, Speech and Audio Technology HSA, Fraunhofer Institute for Digital Media Technology (IDMT),

Marie-Curie-Straße 2, 26129, Oldenburg, Germany

Keywords:

Data Augmentation, Epileptic Seizure Detection, Imbalanced Data, Convolutional Neural Networks.

Abstract:

Epilepsy, characterized by recurrent seizures, poses a signiﬁcant risk to an individual’s safety. To mitigate

these risks, one approach is to use automated seizure detection systems based on Convolutional Neural Net-

works (CNN), which rely on large amounts of data to train effectively. However, real-world seizure data

acquisition is challenging due to the short and infrequent nature of seizures, resulting in a data imbalance

which complicates accurate seizure detection. In this paper, various data augmentation techniques were uti-

lized to increase the amount of training data for CNN, aiming to investigate the potential of these techniques to

enhance the performance of the seizure detection algorithm by providing more seizure data. For this purpose,

two datasets, a unimodal (3D acceleration) and a multimodal dataset (3D acceleration, heart rate and tem-

perature), were used. To evaluate the effect of the different augmentation techniques, a CNN trained without

augmented data was used as a baseline. Experiments showed that data augmentation techniques improved the

seizure detection by lowering the baseline’s false alarm rate while maintaining its high sensitivity. The best

results were achieved with a combination of Rotation and Permutation in the multimodal dataset and Rotation,

as well as Magnitude Warping, in the unimodal dataset.

1 INTRODUCTION

Epilepsy is one of the most common neurologi-

cal disorders, affecting 50 million people worldwide

(McGeehan, 2018). One of the disease’s symptoms

are recurrent seizures. Epileptic seizures that involve

involuntary body movements can sometimes be ac-

companied by a loss of consciousness, have the poten-

tial to cause severe injuries and create life-threatening

situations. (Ahmad et al., 2022; Sazgar and Young,

2019; Beniczky et al., 2021; Schulze-Bonhage et al.,

2010)

Diagnosis and treatment of epilepsy highly de-

pend on accurate information about the seizures that

occur. Since patients regularly do not notice or for-

get that a seizure happened, automatic seizure detec-

https://orcid.org/0000-0001-7180-2578

https://orcid.org/0009-0006-4834-2618

tion could improve seizure documentation and by that

also diagnosis and treatment. (Ramgopal et al., 2014;

Bidwell et al., 2015)

To develop automated seizure detection systems,

an algorithm needs to be trained on the seizure data.

However, collecting real seizure data presents chal-

lenges. Acquiring real epileptic seizure data is costly

and time-consuming for medical experts and patients

(Siddiqui et al., 2020). Moreover, the unpredictable

and brief nature of seizures, often lasting mere sec-

onds to minutes, results in highly imbalanced datasets

with a skewed seizure and non-seizure distribution

(Siddiqui et al., 2020). This poses a signiﬁcant

challenge for accurate seizure detection, as the algo-

rithm may be biased towards the majority class (non-

seizure), leading to suboptimal performance in detect-

ing the minority class instances (seizure) (Siddiqui

et al., 2020). Additionally, neural networks need large

numbers of training data to perform well. A promis-

142

Seyidova, M., Henze, J., Pelzer, A. and Rhein, B.

Comparison of Different Data Augmentation Techniques for Improving Epileptic Seizure Detection Based on 3D Acceleration, Heart Rate and Temperature Data.

DOI: 10.5220/0012386800003657

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2024) - Volume 2, pages 142-153

ISBN: 978-989-758-688-0; ISSN: 2184-4305

ing strategy to address these issues is to increase the

number of seizure samples by generating seizure data

synthetically by utilizing time series data augmenta-

tion techniques to achieve a better classiﬁcation per-

formance. (Wen et al., 2021)

The MOND project, building on the results of

the EPItect project, works on developing an auto-

mated mobile method for detecting epileptic seizures

that is suitable for everyday use (Fraunhofer, 2023).

It employs an In-Ear sensor by cosinuss°, which is

equipped to measure 3D acceleration, photoplethys-

mography (PPG) and temperature. A previous uni-

modal approach achieved a sensitivity of 65.1% for

the detection of tonic-clonic seizures based on 3D ac-

celeration data (Houta et al., 2019). A multimodal

approach using 3D acceleration and heart rate data

achieved a sensitivity of 100% with a false alarm rate

of 138FA/24h using conventional machine learning

techniques (Henze et al., 2021). In another study,

seizures with motor components were detected us-

ing a multimodal approach, leveraging 3D acceler-

ation, heart rate and temperature data. Deep learn-

ing techniques were applied and the best CNN model

achieved 86.66% sensitivity with a false alarm rate of

1,804FA/24h (B

oring, 2021).

This work aimed to investigate the potential of dif-

ferent standard augmentation techniques to improve

the performance of seizure detection by increasing

the amount of data available for the MOND project.

It builds upon the best-performing CNN architecture

obtained in the previous work by B

oring (2021). For

this purpose, real seizure samples are augmented via

standard augmentation techniques, such as Permuta-

tion, Rotation, Jittering, Time Warping, Magnitude

Warping and Window Slicing using a unimodal and

a multimodal dataset.

To avoid testing methodology that might result in

overly optimistic results, it is necessary to include dif-

ferent seizure types in the dataset (Shoeb and Gut-

tag, 2010). Therefore, this work focused on motor

epileptic seizures that include various seizure types.

This approach enabled the utilization of the majority

of data and different seizure types from the MOND

dataset before employing augmentation techniques.

2 RELATED WORK

In recent years, data augmentation has emerged as a

promising approach to tackle the challenges of insuf-

ﬁcient seizure data and imbalanced seizure datasets.

(Lashgari et al., 2020)

In a survey conducted by Wen et al. (2021), data

augmentation was regarded as an effective method to

enhance both the quantity and quality of training data,

enabling the efﬁcient use of deep learning models.

The study also demonstrated the effectiveness of data

augmentation in many time series classiﬁcation prob-

lems where class imbalance is often observed.

Iwana and Uchida (2021) noted that many time se-

ries data augmentation techniques, like those for im-

ages, are based on random transformations. The sur-

vey presented Jittering as a frequently used method

and Permutation and Rotation as effective techniques

to be combined with other methods for sensor data

augmentation. The study emphasized that data aug-

mentation methods are task-dependent. The inher-

ent temporal dependency of time series data further

complicates the identiﬁcation of effective methods for

speciﬁc tasks, such that different time series datasets

may have unique properties and not every transfor-

mation technique is applicable to every type of time

series dataset. For instance, the Jittering technique,

which involves adding noise, assumes that the time

series data patterns are naturally noisy, which may

be true for sensor data. On the other hand, Rotation

was mentioned to have a potentially detrimental effect

on some time series classiﬁcation tasks where it can

cause a change in the label of the data.

In a paper on Parkinson’s disease monitoring via

the wearable sensor by Um et al. (2017), differ-

ent standard data augmentation methods were utilized

to augment the acceleration data. Combining vari-

ous data augmentation methods outperformed a sin-

gle data augmentation technique. The combination of

two techniques (Rotation and Permutation, Rotation

and Time Warping) achieved performance improve-

ment by 7.5-9.2% compared to the baseline. The best

performance among the combination of three tech-

niques was obtained using Rotation, Permutation and

Time Warping with an 86.88% accuracy. This was a

signiﬁcant improvement over the baseline accuracy of

77.54%. (Um et al., 2017)

Despite efforts to use data augmentation tech-

niques in seizure detection, to our best knowledge,

data augmentation has not been used speciﬁcally for

acceleration data, heart rate and temperature in the

context of seizure detection. This work investigated

the potential of various standard augmentation tech-

niques to improve seizure detection using the same

data prepossessing pipeline and the best-performing

CNN model from previous work (B

oring, 2021).

Comparison of Different Data Augmentation Techniques for Improving Epileptic Seizure Detection Based on 3D Acceleration, Heart Rate

and Temperature Data

143

3 MATERIALS AND METHODS

3.1 Dataset

The dataset used in this work consisted of ﬁve fea-

tures: 3D acceleration (acceleration along the x-axis,

y-axis and z-axis), temperature and heart rate. These

measurements were collected from epilepsy patients

at the Department of Epileptology at University Hos-

pital Bonn (Germany) through an In-Ear sensor from

cosinuss°. The sensor captured 3D acceleration at 50

Hz sampling frequency. Temperature is given in the

unit degree Celsius °C and was measured at a sam-

pling frequency of 1 Hz. Additionally, the heart rate,

computed from the prior 6 seconds of the PPG sig-

nal, was sampled at 1 Hz frequency. Heart rate data

is given in units of bpm (beats per minute) or min

−1

at a sampling frequency of 1 Hz with corresponding

quality indexes available in the dataset. Information

regarding the quality index is provided by the man-

ufacturer of the sensor. The measurements with a

quality index below 40 were considered inaccurate,

while those with a quality index above 65 were re-

garded as correct with a high degree of certainty. It

was not possible to draw any conclusions about the

accuracy of the heart rate measurements when they

fell within the range of 40 to 65. All data was avail-

able in the form of time series. Moreover, the records

of seizure events were separately available and anno-

tated via video-EEG by the clinicians. These records

provided information such as seizure starting time,

ending time, seizure types and additional event de-

tails. The dataset contained both the motor and non-

motor seizure types. However, this study aimed to

enhance the detection of motor seizures.

After excluding non-motor seizures, the dataset

retained four distinct types of motor seizure events:

• Tonic-clonic seizure (TCS): 11

• Focal impaired awareness seizure (FIAS): 70

• Focal aware seizure (FAS): 27

• Generalized motor seizure (GMS): 2

A total of 110 seizures from 45 patients remained.

The total duration of seizures is 2.675 hours with a

mean seizure length of 1.47minutes = 87seconds.

Figure 1 and Figure 2 show the stereotypical pat-

tern of a TCS seizure event for a patient from the

dataset, encompassing acceleration measurement in

x, y, z axis, heart rate and temperature. Acceleration

data for a TCS episode is depicted in Figure 1, begin-

ning with the tonic phase, indicating muscle stiffening

and progressing to the clonic phase, which is charac-

terized by rhythmic muscle jerking that increases in

intensity and magnitude. Figure 2 shows a segment

Figure 1: Section of raw 3D acceleration data of a pa-

tient.The acceleration along the x−, y−, and z− axes are

represented on the y-axis in gravity units, while the x-axis

shows the recording time. The region within two black ver-

tical lines is the TCS event experienced by the patient.

Figure 2: Section of heart rate, quality and temperature for

patient during the TCS event. The y-axis for the top (ma-

genta) plot represents heart rate in beats per minute (bpm).

The quality is given in the middle (cyan) plot. The red line

in the quality plot represents the minimum accepted qual-

ity=40. The y-axis of the third (green) plot represents tem-

perature in degrees Celsius °C.

of heart rate data plotted over time with corresponding

quality values during the same seizure. The quality of

heart rate readings was diminished during the seizure

event, falling below the quality index. Since the heart

rate data was collected based on the PPG signals,

movement during the seizure could cause the photodi-

ode to shift, likely contributing to the data’s low qual-

ity (Henze et al., 2021). Additionally, the body tem-

perature rose during the seizure and decreased after

the event.

3.2 Data Preparation

The data preparation strategy used in this work was

like that in previous work (B

oring, 2021). The raw

data was cleaned to remove duplicate sensor record-

ings, followed by a windowing process. In this work,

each patient’s recording was divided into 10-second

windows without overlap.

An important aspect to consider when dividing the

data into windows was the presence of missing time

steps. These missing values could be caused by sen-

HEALTHINF 2024 - 17th International Conference on Health Informatics

144

sor errors during the observation time of the patient.

Consequently, the windowing process led to some

windows containing incomplete or missing data. To

ensure the integrity of the data used in the analysis,

we addressed the issue of missing time steps. This

included considering the sampling frequency of the

sensor and calculating the expected number of values

within each window. For example, with a 50 Hz sam-

pling frequency for acceleration, a 10-second window

should contain 500 values. Likewise, for heart rate

and temperature with a 1 Hz sampling frequency, the

expected number of values was 10. To further ensure

that the windows used in the analysis represent the pa-

tient’s condition, a threshold was set at 15% (B

oring,

2021). Windows with more than 15% missing values

within a window were discarded. If a window had

less than 15% missing values, linear interpolation was

used. This ensured that the CNN received windows of

consistent size.

This 15% threshold determined whether a window

should be discarded or interpolated. A smaller thresh-

old resulted in more windows being discarded, while

a larger threshold resulted in more interpolated val-

ues, potentially skewing the analysis and misrepre-

senting the patient’s biodata. If a window was rejected

due to excessive missing data, the window continued

to slide along the time series until a section was found

that met the threshold criteria. This approach helped

to ensure that as much data as possible was utilized in

the analysis while also maintaining the integrity of the

data by discarding windows with many missing val-

ues. In the multimodal case, the whole window was

discarded if data from at least one of the modalities

(3D acceleration, heart rate or temperature) was miss-

ing. Due to that, the unimodal dataset consisted of

more samples than the multimodal dataset since there

were more missing values within the heart rate and

temperature data than in the 3D acceleration data. In

the window generation process described above, each

window was assigned a class label (0: non-Seizure,

1: Seizure). The window was labelled as seizure if

any part of the window fell within the seizure inter-

val (B

oring, 2021). The window was labelled as a

non-seizure if there was no overlap between it and a

seizure interval.

3.2.1 Multimodal Dataset

For the multimodal approach (3D acceleration, heart

rate and temperature), the data preparation process

resulted in a total of 785 seizure and 584,737 non-

seizure windows. In total, there were 1626.45 hours

of measurements, of which approximately 2.18 hours

belonged to motor seizures. A total of 88 seizure

events remained inside the data after data preparation

steps with the following distribution:

• FIAS: 54

• FAS: 21

• TCS: 11

• GMS: 2

3.2.2 Unimodal Dataset

For the unimodal dataset consisting only of 3D ac-

celeration, data preparation gave 941 seizure and

648,399 non-seizure windows. Out of 1803.72 hours

of total measurements, approximately 2.61 hours be-

longed to motor seizures. A total of 94 seizure events

were identiﬁed in the data, with the following distri-

bution:

• FIAS: 60

• FAS: 21

• TCS: 11

• GMS: 2

3.3 Data Augmentation Methods

Data augmentation is a technique used to artiﬁcially

increase the size of a dataset by creating modiﬁed

versions of existing data. In the context of imbal-

anced dataset distributions, data augmentation allows

additional data to be generated for underrepresented

classes, creating a more balanced dataset. Several

studies have shown that data augmentation techniques

can improve the generalizability of deep learning net-

works, thereby reducing overﬁtting and enabling the

networks to handle imbalanced datasets more effec-

tively. (Blagus and Lusa, 2013; Hasibi et al., 2019;

Shorten and Khoshgoftaar, 2019)

While the choice of data augmentation techniques

for time series data depends on the dataset’s proper-

ties and the task at hand, several basic techniques have

been identiﬁed in this area.(Iglesias et al., 2023; Um

et al., 2017).

Jittering is a type of magnitude domain trans-

formation that creates novel time series by introduc-

ing noise at every time step of the initial time series

(Iwana and Uchida, 2021). Rotation is a magnitude

domain transformation achieved by applying a Rota-

tion matrix with a deﬁned angle to multivariate time

series data (Um et al., 2017). In Permutation, the

time series segments are rearranged to produce a new

pattern (Um et al., 2017). The segments can be of

equal or variable size. To create a new time series

using the Permutation technique, the original time se-

ries is divided into N segments and then randomly re-

arranged. However, this random rearrangement does

Comparison of Different Data Augmentation Techniques for Improving Epileptic Seizure Detection Based on 3D Acceleration, Heart Rate

and Temperature Data

145

not maintain the temporal dependencies in the origi-

nal time series (Iglesias et al., 2023).

Time Warping is perturbing the pattern in the

temporal dimension. The timesteps are either

stretched or contracted to generate a new pattern.

Magnitude Warping is a type of magnitude domain

transformation that warps the magnitude of each time

series by convolving with smooth curves. Lastly, the

Window Slicing augmentation technique involves the

removal of a portion of a data sample to create an en-

tirely new sample.

Multiple techniques can be applied sequentially to

the original time series to create transformed time se-

ries data.

For augmentation, the entire seizure intervals were

used instead of the segmented windows. One rea-

son for this is that the windows were obtained af-

ter interpolation and only represent a portion of the

seizure interval. Applying the transformation on each

portion separately might not have resulted in a good

quality synthetic seizure sample. To ensure the max-

imum quality of the generated sample, the augmen-

tation techniques were applied directly to the whole

seizure interval before windowing.

3.4 Evaluation Procedure

This study used a Convolutional Neural Network

(CNN) as a binary classiﬁer to determine whether a

given window belonged to a seizure or a non-seizure

event. To assess the effectiveness of the augmen-

tation techniques, the CNN was trained with aug-

mented data and the performance was compared with

the baseline results (CNN trained only with the origi-

nal data).

Some metrics are commonly used in seizure de-

tection (Sun et al., 2009; Henze et al., 2021):

• Sensitivity

seizure

: Number of all detected seizures

/ Number of all seizures

• FA/24h: Number of false alarms per 24 hours

• Sensitivity: Number of all detected seizure win-

dows / Number of all seizure windows

• Speciﬁcity: Number of all correctly classiﬁed

non-seizure windows / Number of all non-seizure

windows

In seizure detection, the main objective is to deter-

mine if a seizure is detected as a whole event rather

than assessing the performance on individual sam-

ples (windows). In this research, a seizure was con-

sidered as detected if at least one sample (window)

within the entire seizure interval was correctly clas-

siﬁed. Based on this, a metric referred to as seizure

sensitivity (Sensitivity

seizure

) was calculated. This in-

forms about the proportion of seizures that are cor-

rectly classiﬁed. The metric Sensitivity refers to win-

dow sensitivity, which is the proportion of seizure

windows that are correctly classiﬁed. Speciﬁcity is

the proportion of non-seizure windows that are cor-

rectly classiﬁed.

Another metric of particular importance in the

case of seizure detection is the false alarm rate, which

is the number of false alarms (non-seizure predicted

as a seizure) within a 24-hour interval. In this re-

search, the false alarm rate for 24 hours was based on

the average length of the seizure interval (90s). In this

context, if any non-seizure window was misclassiﬁed

within each non-overlapping 90-second interval, the

entire interval was counted as a single false alarm.

Following this, the total count of false alarms was di-

vided by the total measurement duration in days to get

a False Alarm Rate per 24 hours (FA/24h). Sensitivity

and false alarm rate are the metrics that are often used

in reporting seizure detection systems (Beniczky and

Ryvlin, 2018).

4 EXPERIMENTS

4.1 Augmentation of Seizures

Each seizure event was augmented to create a syn-

thetic seizure. Different random parameter values

were used for each data augmentation technique.

For Jittering, the noise was generated from a nor-

mal distribution of chosen mean and standard devia-

tion. The mean was set at 0. The study by Um et

al. (2017) used a standard deviation value of 0.03,

whereas, in this work, standard deviation values of

0.02 and 0.05 were used to evaluate the impact of dif-

ferent noise levels in the data. By using different val-

ues of standard deviation, we were able to examine

the effect of low and high noise levels on the mod-

els’ performance. The generated noise was added to

each channel at every time step of the original seizure

sample.

Rotation was performed using a Rotation matrix

formed through angles and axes. The angle was

drawn uniformly from the interval [-π, π] while the

axis was chosen within a uniform distribution be-

tween [-1,1].

In the implementation of Permutation, the original

seizure sample was rearranged into a new time series

by dividing it into a speciﬁed number of segments and

randomly rearranging the order of the segments. The

minimum length of each segment was set to a speci-

ﬁed value. The minimum length of each segment was

HEALTHINF 2024 - 17th International Conference on Health Informatics

146

ﬁxed at 10 while the number of segments N was tested

at values of 2, 3, and 5

Random sinusoidal curves were generated using

arbitrary amplitude, frequency and phase values for

Magnitude- and Time-Warping.

Window Slicing croped the time series data to

90% of its original length. The starting point of the

Window Slicing was selected randomly and the re-

maining 10% of data was removed from either end.

To facilitate a direct comparison with other data aug-

mentation methods, the cropped time series was then

linearly interpolated back to its original length.

Figure 3 depicts the application of some single and

combined data augmentation techniques on the accel-

eration data in the y-axis, showing how those tech-

niques transform the original seizure data.

Some augmentation techniques, like Magnitude

and Time Warping, were applied only to the unimodal

dataset since the library from Um et al. (2017) that

was used for implementation did not support those

operations in our multimodal case. Additionally, Ro-

tation was not applied to heart rate and temperature

data to avoid unrealistic values like a decreasing heart

rate at the start of a seizure while it usually goes up.

1x, 2x, 4x and 8x proportion of augmented seizure

samples (1x, 2x, 4x and 8x as many augmented

seizure samples as real seizure samples) were gen-

erated from each technique for unimodal and multi-

modal datasets. For techniques only applicable to 3D

acceleration data, to create the augmented samples for

the multimodal dataset, the heart rate and temperature

were kept the same as the original and the transfor-

mation was only applied to the 3D acceleration data.

These samples then underwent the windowing pro-

cess and resulting windows were added to the original

data before training.

4.2 Train-Test Split

Both unimodal and multimodal dataset were divided

into two parts: a train set and a hard-coded test set.

Within the training set, 20% was reserved as a vali-

dation set to optimize model performance. The train-

ing data was used to train the model, while the test

data was used to evaluate the model’s performance

on previously unseen data. Training and evaluating

the model with data from the same patients existing

both in the train and test set can give over-optimistic

results. This is because values measured from an in-

dividual patient can be highly correlated due to their

unique characteristics. To address this, train-test split

was done based on distinct patients, not individual

data windows.

Figure 3: Example of a single and combined standard time

series data augmentation techniques applied to the y-axis

of 3D acceleration data. The techniques were implemented

using code from Um et al. (2017). Original (red), Jittering

(cyan), Permutation (yellow), Rotation (magenta), Rotation

+ Permutation (green), Rotation + Permutation and Time

Warping (blue).

Table 1: Distribution of seizure types in train and test sets

with the multimodal dataset.

Type Train Test

FIAS 42 12

FAS 18 3

TCS 7 4

GMS 1 1

Total 68 20

4.2.1 Train-Test Split with Multimodal Dataset

20 out of 88 seizures were included in the test set,

corresponding to 9 out of 45 patients. This proportion

of train-test split led to 125 seizure and 59,370 non-

seizure windows in the test set (16% of the dataset),

whereas 660 seizure and 525,367 non-seizure win-

dows (84% of the dataset) in the train set. Table 1

gives the distribution of seizure types in the train and

test set in the multimodal dataset.

Comparison of Different Data Augmentation Techniques for Improving Epileptic Seizure Detection Based on 3D Acceleration, Heart Rate

and Temperature Data

147

Table 2: Distribution of seizure types in train and test sets

with the unimodal dataset.

Type Train Test

FIAS 48 12

FAS 18 3

TCS 7 4

GMS 1 1

Total 74 20

4.2.2 Train-Test Split with Unimodal Dataset

20 out of 94 seizures were included in the test set cor-

responding to 9 out of 45 patients. This proportion

of train-test split led to 150 seizure and 67,104 non-

seizure windows in the test set (16% of the data set),

whereas 791 seizure and 581,295 non-seizure win-

dows (84% of the data set) in the train set. Table 2

gives the distribution of seizure types in the train and

test set in the unimodal dataset.

The same train-test split was used for all ex-

periments to ensure comparability. All generated

synthetic data were used in addition to the original

seizure samples in the train set while keeping the test

set unchanged.

4.3 CNN Architecture and Training

To develop a CNN network for seizure detection,

the model architecture from previous work (B

oring,

2021) was used. This model was an adaptation of

the architecture originally proposed by Meisel et al.

(2020). CNN were chosen because of their higher ro-

bustness and time invariance in comparison to other

network architectures when used on time-series data

(Ismail Fawaz et al., 2021). Additionally, they are

easy to implement and highly efﬁcient (Shoeibi et al.,

2021). B

oring (2021) employed stratiﬁed k-fold

cross-validation to assess the performance of vari-

ous CNN architectures in seizure detection and the

adapted architecture from Meisel et al. (2020) was

found to be the most effective in leveraging wear-

able sensor data for the detection of epileptic seizures.

The original architecture includes a 1D convolutional

layer with 64 ﬁlters of size 2, followed by a ReLU

activation function and a max pooling layer with a ﬁl-

ter size of 2. A dense layer with 50 neurons, with

a dropout layer (rate 0.7) present before and after it.

The ﬁnal layer is a dense layer with two output neu-

rons activated by the SoftMax function (Meisel et al.,

2020).

Since the different sampling frequencies of the

sensor data caused issues with the input dimensions,

the original architecture was modiﬁed by B

oring

(2021) to have multiple inputs, with each input layer

receiving a single feature window and processing it

separately through the same layer structure until the

last dropout layer. Then, a global average pooling

layer was applied to have the same size output from

each model. The outputs of these separate individual

models were then added together using a concatena-

tion layer. Finally, the dense layer with SoftMax was

applied to output the probabilities of the respective

class. The shape of the input layer was set to 500

for acceleration and 10 for heart rate and temperature

since the dataset had shapes (500,1) and (10,1), re-

spectively.

In this work, two models were trained: a three-

input model for three features (unimodal - only 3D

acceleration data) and a ﬁve-input model for ﬁve fea-

tures (multimodal - 3D acceleration data, heart rate

and temperature) given in Figure 4.

The ADAM optimizer was used during training,

and the batch size was set to 2048 (Kingma and Ba,

2014). The number of training iterations, also known

as epochs, was set to 2000. Binary cross entropy was

used as a loss function. An early stopping method was

used to monitor the validation loss so that the train-

ing stopped as soon as the model stopped improving.

The initial learning rate was set to 0.001, and a Keras

function called ReduceLROnPlateau was used to ad-

just the learning rate during training. The minimum

learning rate was set to 0.0001 (B

oring, 2021).

To reduce the training time, undersampling was

applied to balance the classes before the training.

Consequently, the non-seizure data was randomly un-

dersampled to match the size of the seizure data. Fur-

thermore, the dataset was standardized to ensure that

all features have a consistent scale. (B

oring, 2021)

4.4 Results

Experiments were conducted by training a CNN with

augmented data in various proportions. The evalua-

tion metrics were obtained for both the multimodal

and unimodal datasets. For comparison, baseline re-

sults were derived by training the CNN without any

augmented data. For choosing the best-performing

setting (data proportions and parameter settings) for

each data augmentation technique or combination of

techniques, the two metrics, seizure sensitivity and

false alarm rate per 24 hours were considered. If

there was one setting with both the highest seizure

sensitivity and the lowest false alarm rate, this setting

was chosen. If not, of those settings with the high-

est seizure sensitivity, the one with the lowest false

alarm rate was chosen, even if there was another set-

ting with a lower false alarm rate that lead to a reduced

seizure sensitivity. The description for the abbrevia-

HEALTHINF 2024 - 17th International Conference on Health Informatics

148

Figure 4: CNN model architecture with ﬁve inputs for 3D-acceleration, heart rate and temperature time series. Based on CNN

architecture in (B

oring, 2021). CNN model with three inputs has the same architecture with three inputs - 3D acceleration.

Table 3: Description of the abbreviations for data augmen-

tation techniques.

Abbreviation Description

Baseline Without augmentation

J Jittering

R Rotation

P Permutation

T Time Warping

M Magnitude Warping

W Window Slicing

PT Permutation + Time Warping

RT Rotation + Time Warping

RP Rotation + Permutation

RPT Rotation + Permutation + Time Wa.

tion corresponding to each augmentation technique is

provided in Table 3. Results are reported for all the

single techniques. Because of the limited space avail-

able, only results for combinations with good results

are reported.

Table 4 summarizes the best-case results of CNN

model performance trained with augmented data from

different data augmentation techniques on the multi-

modal scenario. It presents the seizure and window

sensitivity, false alarm rate per 24 hours and speci-

ﬁcity values for the best cases. The proportion of aug-

mented data used during the training is also shown.

Additionally, the ”Base” column provides the base-

line (without data augmentation) results for compar-

ison. 7 out of 10 techniques achieved the highest

value of 0.9 seizure sensitivity and detected 18 out

of 20 seizures, as in the baseline. The false alarm rate

ranges from 181 FA/24h (J, std=0.05) to 363 FA/24h

(W), with a mean value of 304 FA/24h, a signiﬁcant

reduction from the baseline (382 FA/24h). Window

sensitivity varies between 0.57 (J) and 0.73 (T), with

a mean value of 0.686, which is close to the base-

line result of 0.69. Conversely, speciﬁcity varies from

0.82 (T) to 0.92 (J), with a mean value of 0.87, repre-

senting a signiﬁcant improvement over the baseline of

0.80. Comparing the results of different proportions,

the 2x proportion was the most effective for six tech-

niques: J, T, W, PT, RT, and RPT. The 4x proportion

performed best for P (n=3), M and RP. The 1x propor-

tion was the most effective for only R. The 8x propor-

tion, which was not used for combined augmentation

techniques, did not produce a best-case outcome for

any of the single augmentation techniques in the mul-

timodal dataset.

Figure 5 visualizes the relationship between

seizure sensitivity and false alarm rate FA/24h for all

augmentation techniques in the multimodal scenario.

Seizure sensitivity is represented on the x-axis, while

FA/24h is shown on the y-axis. Each dot represents a

technique and the color corresponds to the technique’s

label. It can be seen that W, T, R, PT, RT, P and RP

achieved the same high seizure sensitivity as the base-

line with a lower false alarm rate than the baseline. J,

M, and RPT yielded a lower FA/24h at the expense

of a seizure sensitivity that was lower than the base-

line. The black arrow on the plot points to the RP

combination, which showed the best performance by

achieving the lowest false alarm rate (259 FA/24h) at

a high seizure sensitivity of 0.9.

The twin axes plot of speciﬁcity and window sen-

Comparison of Different Data Augmentation Techniques for Improving Epileptic Seizure Detection Based on 3D Acceleration, Heart Rate

and Temperature Data

149

Table 4: Best results of the experiments for each data augmentation technique in the multimodal dataset.

Metrics Base J R P T M W PT RT RP RPT

Sensitivity

seizure

0.9 0.8 0.9 0.9 0.9 0.85 0.9 0.9 0.9 0.9 0.85

FA/24h 382 181 351 317 358 242 363 343 337 259 298

Sensitivity 0.69 0.57 0.72 0.68 0.73 0.66 0.69 0.69 0.72 0.66 0.69

Speciﬁcity 0.80 0.92 0.85 0.87 0.82 0.91 0.82 0.86 0.83 0.90 0.89

Proportion No 2x 1x 4x 2x 4x 2x 2x 2x 4x 2x

Table 5: Best results of the experiments for each data augmentation technique in the unimodal dataset.

Metrics Base J R P T M W PT RT RP RPT

Sensitivity

seizure

0.8 0.7 0.85 0.85 0.95 0.8 0.85 0.85 0.85 0.85 0.8

FA/24h 453 228 293 422 551 213 540 553 297 337 345

Sensitivity 0.6 0.42 0.56 0.58 0.58 0.46 0.66 0.57 0.6 0.64 0.64

Speciﬁcity 0.80 0.90 0.90 0.83 0.62 0.94 0.64 0.66 0.85 0.88 0.82

Proportion No 1x 4x 8x 1x 8x 1x 4x 4x 4x 4x

Figure 5: Comparison based on seizure sensitivity and

FA/24h of data augmentation techniques in the multimodal

dataset.

Figure 6: Comparison based on sensitivity and speciﬁcity

of data augmentation techniques in the multimodal dataset

based on Table 4.

sitivity for the multimodal case is given in Figure 6.

The speciﬁcity is higher than the baseline with all the

augmentation techniques. Using R, T and RT, the

window sensitivity increased, while with J, P, M and

RP techniques, it decreased below the baseline. W,

PT and RPT achieved the same window sensitivity as

the baseline.

Table 5 summarizes the best results from the uni-

modal dataset. The seizure sensitivity values ranged

from 0.8 (Base, M, RPT) to 0.95 (T) with a mean

of 0.835. The highest value (0.95, 19 of 20 seizures

detected) was achieved with Time Warping (T). The

false alarm rate had values between 213 FA/24h (J)

and 553 FA/24h (T), with a mean value of 377

FA/24h. While the window sensitivity varied between

0.42 (J, std=0.05) and 0.66 (W), with a mean value

of 0.573, the speciﬁcity varied between 0.62 (T) and

0.94 (M) with a mean value of 0.80. Among all the

different proportions, using the 4x proportion per-

formed well for ﬁve techniques: R, RT, RP, PT and

RPT. While the 1x proportion performed best for J, T

and W, the 8x proportion was most successful in two

techniques: P(n=3) and M. Conversely, the 2x pro-

portion did not give any better overall performance in

any augmentation techniques.

For the unimodal scenario, Figure 7 displays

the relationship between seizure sensitivity and false

alarm rate per 24 hours. T, PT, W, P, RP, RT and R

achieved a higher seizure sensitivity than the baseline.

However, the false alarm rate was very high (greater

than the baseline) for W, PT, and T, while it is lower

than the baseline for P, RP, RT and R. Using RPT and

M augmentation techniques yields the same seizure

sensitivity as the baseline at a lower false alarm rate

than the baseline. Moreover, the technique J yielded

a low seizure sensitivity at a lower false alarm rate.

The technique T gave the highest seizure sensitivity

of 0.95 with an increased false alarm rate compared

to the baseline. The M performed the best in terms of

the lowest false alarm rate among all other techniques

at the same seizure sensitivity as the baseline. On the

other hand, the R achieved the best result with higher

seizure sensitivity at a lower false alarm rate than the

HEALTHINF 2024 - 17th International Conference on Health Informatics

150

Figure 7: Comparison based on seizure sensitivity and

FA/24h of data augmentation techniques in the unimodal

dataset.

Figure 8: Comparison based on sensitivity and speciﬁcity of

data augmentation techniques in the unimodal dataset based

on Table 5.

baseline.

Figure 8 provides the twin axes plot of speciﬁcity

and window sensitivity in the unimodal case. J, R, P,

M, RT, RP and RPT had a higher speciﬁcity than the

baseline, indicating a low false alarm rate for these

techniques. T, W, and PT yielded lower speciﬁcity.

The window sensitivity is lower in J, R, P, T, M, and

PT, while it was the same as the baseline for RT. RP

and RPT techniques improved both the window sen-

sitivity and the speciﬁcity. RT achieved the same win-

dow sensitivity as the baseline at higher speciﬁcity.

5 DISCUSSION

In the multimodal case, all data augmentation tech-

niques resulted in a reduction of the false alarm rate

per 24 hours, which was also accompanied by an in-

crease in speciﬁcity. With the exceptions of J, M,

RPT, which achieved a lower seizure sensitivity, all

other techniques achieved the baseline seizure sensi-

tivity. Among the techniques that achieved the same

seizure sensitivity as the baseline, T, R and RT also

increased the window sensitivity. In contrast, the

techniques P and RP reduced the window sensitiv-

ity while with the techniques W and PT, the win-

dow sensitivity remained the same as the baseline.

At the same seizure sensitivity as the baseline, RP

yielded fewer false alarms (259 FA/24h) than RT (337

FA/24h). However, RT detected a larger proportion

of seizure windows overall, with a window sensitivity

of 0.72 compared to the RP’s 0.66. Considering the

same high seizure sensitivity (0.9) and lowest false

alarm rate, the RP combination was the most promis-

ing technique for the multimodal case.

In the unimodal case, all data augmentation tech-

niques except J yielded a high seizure sensitivity

equal to or greater than the baseline. The use of

M and RPT methods achieved a seizure sensitivity

as the baseline with a lower false alarm rate. With

the same seizure sensitivity, the false alarm rate was

much lower with M (213 FA/24h) than with RPT

(345 FA/24h), while M detected fewer seizure win-

dows (window sensitivity of 0.46) than RPT (win-

dow sensitivity of 0.64). Using R, P, W, PT, RT, and

RP techniques improved the seizure sensitivity. How-

ever, among these techniques, the false alarm rate was

much higher for W and PT. This increase in false

alarms is associated with a reduction in speciﬁcity

(below 0.7), compared to the baseline (0.8). At a

seizure sensitivity of 0.85, R achieved the lowest false

alarm rate of 293 FA/24h but with a slightly lower

window sensitivity of 0.56 compared to the baseline

of 0.6. On the other hand, RT and RP result in a

high window sensitivity of 0.6 and 0.64, with slightly

higher false alarm rates of 297 FA/24h for RT and 337

FA/24h for RP. However, these false alarm rates were

still lower than the baseline (453 FA/24h). T is the

only technique that yielded the highest seizure sensi-

tivity of 0.95, with 19 of 20 seizures detected at the

expense of an increased false alarm rate. Thus, it can

be concluded that R, RT and RP were the most suit-

able techniques for the unimodal case. Moreover, it

was possible to achieve a much lower false alarm rate

(213 FA/24h) while maintaining the same seizure sen-

sitivity as the baseline (0.8) by utilizing the M data

augmentation technique. Comparing the R and M

techniques, the R technique detected one seizure more

than the M technique. In contrast, M yields 80 false

alarms per 24 hours less than R. Detecting seizures

accurately is critical in reducing the risk for epilepsy

patients, while having a low false alarm rate is desir-

able for practical use. Given this tradeoff, a clinician’s

view might be needed to determine the preferred ap-

proach.

Using data generated via all time series data

Comparison of Different Data Augmentation Techniques for Improving Epileptic Seizure Detection Based on 3D Acceleration, Heart Rate

and Temperature Data

151

augmentation techniques (except T) resulted in a

higher seizure sensitivity in the multimodal case as

compared to when they were used in the unimodal

case. However, considering the detection improve-

ment from their baseline version, no technique in the

multimodal case has produced more seizure sensitiv-

ity than its baseline. In contrast, all the techniques

(except J, M, and RPT) in the unimodal case have

given higher seizure sensitivity than its baseline. Us-

ing T in a unimodal case, a seizure sensitivity of 0.95

was achieved, which was overall the highest among

all the techniques in both unimodal and multimodal

cases.

The difference in the results achieved by unimodal

and multimodal datasets can be due to the difference

in the amount of data available for the multimodal and

unimodal datasets. This difference originated from

the lack of temperature and heart rate data for some

seizure events in the multimodal dataset (see 3.2.2).

As a result, different train and test set sizes in the

unimodal and multimodal case potentially impacted

the prediction outcomes. Using data augmentation

techniques, the average false alarm rate over all tech-

niques for the multimodal case was approximately

304 FA/24h. This is notably lower than the average

false alarm rate of 377 FA/24h observed in the uni-

modal case. Since the lower false alarm means that

more non-seizure windows were classiﬁed accurately,

it was observed that, as a result, the average speciﬁcity

with all techniques in the multimodal case was 0.87,

which is higher than the average speciﬁcity of 0.8 ob-

served in the unimodal scenarios. The window sen-

sitivity for all the techniques in the multimodal case,

averaging 0.686, was also higher than those of the uni-

modal case, which had an average window sensitivity

of 0.573.

Summarizing the above, using standard time se-

ries data augmentation techniques in the multimodal

case reduced false alarm rates while maintaining

the same high seizure sensitivity. In the unimodal

case, various data augmentation techniques improved

seizure sensitivity from the baseline while lowering

false alarm rates. Therefore, it was shown that the use

of data augmentation techniques can be a way to im-

prove the performance of CNN in seizure detection.

6 CONCLUSION

This work investigated the potential of data augmen-

tation techniques to improve a seizure detection sys-

tem. For this purpose, standard time series data aug-

mentation techniques were utilized to increase the

amount of available motor seizure data. The exper-

iments demonstrated that training the CNN with a

combination of synthetic data generated through the

aforementioned augmentation techniques and original

seizure data can result in an increased seizure sensi-

tivity and a reduced false alarm rate.

In the multimodal dataset, the most effective

augmentation approach was achieved by utilizing a

combination of Rotation and Permutation techniques

(RP), whereas in the unimodal dataset, the best results

were obtained by employing Rotation (R) or Mag-

nitude Warping (M) techniques. In the multimodal

case, the combination of Rotation and Permutation

(RP) achieved the same seizure sensitivity as the base-

line (0.9) while reducing the false alarm rate by 123

FA/24h. In the unimodal case, Rotation (R) increased

the seizure sensitivity by 0.05 (from 0.8 to 0.85) while

reducing the false alarm rate by 160 FA/24h (from

453FA/24h to 293FA/24h). Both improvements of the

false alarm rate are statistically highly signiﬁcant ap-

plying the binomial test. The aforementioned results

were observed when the synthetic data obtained from

the most effective techniques was combined with the

original dataset, which was used for CNN training.

Evaluating the impact of data augmentation across

diverse test sets, coupled with experimentation with

different Machine Learning algorithms, may enhance

the performance of seizure detection systems. Addi-

tionally, the inﬂuence of augmentation on detecting

different seizure types could also be explored with

larger test sets.

ACKNOWLEDGEMENTS

The authors acknowledge the ﬁnancial support by the

Federal Ministry of Health in the framework of the

MOND (project number G512F11007).

REFERENCES

Ahmad, I., Wang, X., Zhu, M., Wang, C., Pi, Y., Khan,

J. A., Khan, S., Samuel, O. W., Chen, S., and Li, G.

(2022). Eeg-based epileptic seizure detection via ma-

chine/deep learning approaches: A systematic review.

Computational Intelligence and Neuroscience, 2022.

Beniczky, S. and Ryvlin, P. (2018). Standards for testing

and clinical validation of seizure detection devices.

Epilepsia, 59:9–13.

Beniczky, S., Wiebe, S., Jeppesen, J., Tatum, W. O.,

Brazdil, M., Wang, Y., Herman, S. T., and Ryvlin,

P. (2021). Automated seizure detection using wear-

able devices: A clinical practice guideline of the inter-

national league against epilepsy and the international

HEALTHINF 2024 - 17th International Conference on Health Informatics

152

federation of clinical neurophysiology. Clinical Neu-

rophysiology, 132(5):1173–1184.

Bidwell, J., Khuwatsamrit, T., Askew, B., Ehrenberg, J. A.,

and Helmers, S. (2015). Seizure reporting technolo-

gies for epilepsy treatment: A review of clinical infor-

mation needs and supporting technologies. Seizure,

32:109–117.

Blagus, R. and Lusa, L. (2013). Smote for high-dimensional

class-imbalanced data. BMC bioinformatics, 14:1–16.

oring, Y. (2021). Implementierung eines convolutional

neural networks zur detektion motorischer epileptis-

cher anf

alle mittels biosignalen eines in-ohr-sensors.

Unpublished master’s project.

Fraunhofer (2023). Mobile, smart neurosensor system for

the detection and documentation of epileptic seizures

in everyday life.

Hasibi, R., Shokri, M., and Dehghan, M. (2019). Aug-

mentation scheme for dealing with imbalanced net-

work trafﬁc classiﬁcation using deep learning. arXiv

preprint arXiv:1901.00204.

Henze, J., Houta, S., Surges, R., Kreuzer, J., and Bis-

gin, P. (2021). Multimodal detection of tonic–clonic

seizures based on 3d acceleration and heart rate data

from an in-ear sensor. In Del Bimbo, A., Cucchiara,

R., Sclaroff, S., Farinella, G. M., Mei, T., Bertini,

M., Escalante, H. J., and Vezzani, R., editors, Pat-

tern Recognition. ICPR International Workshops and

Challenges, pages 490–502, Cham. Springer Interna-

tional Publishing.

Houta, S., Bisgin, P., and Dulich, P. (2019). Machine

learning methods for detection of epileptic seizures

with long-term wearable devices. In Eleventh Interna-

tional Conference on eHealth, Telemedicine, and So-

cial Medicine.

Iglesias, G., Talavera, E., Gonz

alez-Prieto,

A., Mozo,

A., and G

omez-Canaval, S. (2023). Data augmen-

tation techniques in time series domain: a survey

and taxonomy. Neural Computing and Applications,

35(14):10123–10145.

Ismail Fawaz, H., Forestier, G., Weber, J., Idoumghar, L.,

and Muller, P.-A. (2021). Deep learning for time series

classiﬁcation: a review. Data Mining and Knowledge

Discovery, 33(4):917–963.

Iwana, B. K. and Uchida, S. (2021). An empirical survey of

data augmentation for time series classiﬁcation with

neural networks. Plos one, 16(7):e0254841.

Kingma, D. P. and Ba, J. (2014). Adam: A

method for stochastic optimization. arXiv preprint

arXiv:1412.6980.

Lashgari, E., Liang, D., and Maoz, U. (2020). Data aug-

mentation for deep-learning-based electroencephalog-

raphy. Journal of Neuroscience Methods, 346:108885.

McGeehan, B. (2018). Supporting students with epilepsy in

the school setting. Communique, 47(3):8–10.

Meisel, C., El Atrache, R., Jackson, M., Schubach, S.,

Ufongene, C., and Loddenkemper, T. (2020). Ma-

chine learning from wristband sensor data for wear-

able, noninvasive seizure forecasting. Epilepsia,

61(12):2653–2666.

Ramgopal, S., Thome-Souza, S., Jackson, M., Kadish,

N. E., S

anchez Fern

andez, I., Klehm, J., Bosl, W.,

Reinsberger, C., Schachter, S., and Loddenkemper,

T. (2014). Seizure detection, seizure prediction, and

closed-loop warning sstems in epilepsy. Epilepsy and

Behavior, 37:291–307.

Sazgar, M. and Young, M. (2019). Absolute Epilepsy

and EEG Rotation Review: Essentials for Trainees.

Springer International Publishing.

Schulze-Bonhage, A., Sales, F., Wagner, K., Teotonio, R.,

Carius, A., Schelle, A., and Ihle, M. (2010). Views of

patients with epilepsy on seizure prediction devices.

Epilepsy & behavior, 18(4):388–396.

Shoeb, A. H. and Guttag, J. V. (2010). Application of ma-

chine learning to epileptic seizure detection. In Pro-

ceedings of the 27th international conference on ma-

chine learning (ICML-10), pages 975–982.

Shoeibi, A., Khodatars, M., Ghassemi, N., Jafari, M.,

Moridian, P., Alizadehsani, R., Panahiazar, M.,

Khozeimeh, F., Zare, A., Hosseini-Nejad, H., Khos-

ravi, A., Atiya, A. F., Aminshahidi, D., Hussain,

S., Rouhani, M., Nahavandi, S., and Acharya, U. R.

(2021). Epileptic seizures detection using deep

learning techniques: A review. International Jour-

nal of Environmental Research and Public Health,

18(11):5780.

Shorten, C. and Khoshgoftaar, T. M. (2019). A survey on

image data augmentation for deep learning. Journal

of big data, 6(1):1–48.

Siddiqui, M. K., Huang, X., Morales-Menendez, R., Hus-

sain, N., and Khatoon, K. (2020). Machine learning

based novel cost-sensitive seizure detection classiﬁer

for imbalanced eeg data sets. International Journal

on Interactive Design and Manufacturing (IJIDeM),

14:1491–1509.

Sun, Y., Wong, A. K., and Kamel, M. S. (2009). Classiﬁca-

tion of imbalanced data: A review. International jour-

nal of pattern recognition and artiﬁcial intelligence,

23(04):687–719.

Um, T. T., Pﬁster, F. M., Pichler, D., Endo, S., Lang,

M., Hirche, S., Fietzek, U., and Kuli

c, D. (2017).

Data augmentation of wearable sensor data for parkin-

son’s disease monitoring using convolutional neural

networks. In Proceedings of the 19th ACM inter-

national conference on multimodal interaction, pages

216–220.

Wen, Q., Sun, L., Yang, F., Song, X., Gao, J., Wang, X.,

and Xu, H. (2021). Time series data augmentation for

deep learning: A survey. In Proceedings of the Thirti-

eth International Joint Conference on Artiﬁcial Intel-

ligence. International Joint Conferences on Artiﬁcial

Intelligence Organization.

Comparison of Different Data Augmentation Techniques for Improving Epileptic Seizure Detection Based on 3D Acceleration, Heart Rate

and Temperature Data

153