Intraoperative Electrocorticography Signal Synthesis to Improve the

Classiﬁcation of Epileptiform Tissue

Leonor Almeida

1,∗ a

, Sem Hoogteijling

2,3,∗ b

, In

es Silveira

1 c

, Dania Furk

1 d

, Irene Heijink

2,3 e

Maryse van’t Klooster

2 f

, Hugo Gamboa

1 g

, Lu

ıs Silva

1 h

and Maeike Zijlmans

2,3 i

Laborat

orio de Instrumentac¸

ao, Engenharia Biom

edica e F

ısica da Radiac¸

ao (LIBPhys-UNL), Departamento de F

ısica,

Faculdade de Ci

encias e Tecnologia, FCT, Universidade Nova de Lisboa, 2829-516 Caparica, Portugal

Department of Neurology and Neurosurgery, Brain Center, University Medical Center Utrecht, Part of ERN EpiCARE,

P.O. box 85500, 3508 GA Utrecht, The Netherlands

Stichting Epilepsie Instellingen Nederland (SEIN), The Netherlands

Keywords:

Synthetic Data, Epilepsy, Epileptiform Activity, Epileptogenic Tissue, ioECoG, GAN.

Abstract:

Epilepsy surgery is a viable option for treating drug-resistant cases where anti-seizure medications fail, but

accurately localizing epileptic tissue remains challenging. This process can be guided by the visual assess-

ment of intraoperative electrocorticography (ioECoG). Data scarcity limits developing machine learning (ML)

models for automatic epileptic tissue classiﬁcation. To address this, we propose a generative model based

on Generative Adversarial Networks (GANs) to synthesize realistic ioECoG signals. Our approach identiﬁed

three distinct ioECoG patterns using Agglomerative Clustering, which guided training individual Deep Convo-

lutional Wasserstein GANs with Gradient Penalty (DCwGAN-GP). Synthetic data (SD) was evaluated across

multiple dimensions: ﬁdelity using temporal (e.g., Wasserstein distance (WD)), frequency and time-frequency

metrics; diversity through dimensionality reduction; and utility by comparing ML performance with and with-

out SD. It replicated temporal and frequency characteristics of real signals (ﬁdelity), though lacked variability

(diversity) due to potential data misclassiﬁcations. Speciﬁcally, the WD between real and synthetic signals

outperformed literature benchmarks (i.e., 0.043 ± 0.025 vs. 0.078). Classiﬁers trained on a combination of

real and SD achieved 88% accuracy, compared to 85% with real data alone. These results demonstrate the po-

tential of SD to replicate real signals, address data scarcity, augment ioECoG datasets, and advance ML-based

epilepsy surgery research.

1 INTRODUCTION

Epilepsy is a central nervous system disorder (Shoka

et al., 2023) characterized by abnormal brain activ-

ity, affecting approximately 50 million people glob-

ally (WHO, 2019). While anti-seizure medication

is the ﬁrst line of treatment (Consales et al., 2021),

about a third of patients have drug-resistant epilepsy

https://orcid.org/0009-0003-1185-9750

https://orcid.org/0000-0002-4455-6700

https://orcid.org/0000-0003-0675-3444

https://orcid.org/0009-0008-7139-8665

https://orcid.org/0009-0005-9266-1168

https://orcid.org/0000-0002-6594-8965

https://orcid.org/0000-0002-4022-7424

https://orcid.org/0000-0001-9811-0571

https://orcid.org/0000-0003-1258-5678

∗

These authors contributed equally to this work.

and do not achieve seizure control through this pro-

cedure (Duncan and Taylor, 2023). For these peo-

ple, epilepsy surgery offers a viable solution, poten-

tially curing seizures and improving their quality of

life (Consales et al., 2021). The surgery procedure

involves the removal of the epileptogenic tissue, that

is, the brain tissue responsible for triggering seizures

(Zijlmans et al., 2019). To localize epileptic tis-

sue, pathological electrographic activity in the intra-

operative electrocorticogram (ioECoG) can be used,

including spikes, sharp waves and ictiform spike pat-

terns or the more recently discovered high frequency

oscillations (HFOs) (Fern

andez and Loddenkemper,

2013) (Greiner et al., 2016) (Wang et al., 2024)

(Zweiphenning et al., 2022). This method involves

placing electrodes directly on surgically exposed cor-

tex to precisely map the tissue to be removed (Tatum,

2021). However, differentiating epileptiform activ-

Almeida, L., Hoogteijling, S., Silveira, I., Furk, D., Heijink, I., Klooster, M. V., Gamboa, H., Silva, L. and Zijlmans, M.

Intraoperative Electrocorticography Signal Synthesis to Improve the Classiﬁcation of Epileptiform Tissue.

DOI: 10.5220/0013398500003911

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

In Proceedings of the 18th International Joint Conference on Biomedical Engineer ing Systems and Technologies (BIOSTEC 2025) - Volume 1, pages 1141-1153

ISBN: 978-989-758-731-3; ISSN: 2184-4305

1141

ity from normal ioECoG patterns remains a challenge

(Zweiphenning et al., 2022). Clinicians often rely

on visual interpretation, which is prone to human er-

rors and can miss critical information (Li et al., 2020)

(Saminu et al., 2022). Also, the typically visible

pathological interictal EEG activity may not be opti-

mally predicting the diseased tissue: epileptic spikes

are seen over a broader area than just the seizure on-

set zone and epileptic HFOs can be difﬁcult to discern

from artefacts and physiological HFOs. This high-

lights the need for more reliable ways to distinguish

epileptic and non-epileptic tissue.

Recent research shows promise in using ML for

automatic detection of interictal epileptiform dis-

charges in intracranial EEG (iEEG) signals (Saminu

et al., 2022) by efﬁciently identifying epilepsy

biomarkers such as spikes, improving diagnostic ac-

curacy and assist in epilepsy surgery (Chaibi et al.,

2024). A step further is to use ML not to detect

short epileptiform events but also background activ-

ity that differs between epileptic and non-epileptic tis-

sue (Hoogteijling et al., 2024). Data scarcity limits

the effectiveness of ML models, which require large

datasets to accurately distinguish epileptiform pat-

terns (Du et al., 2024a). To overcome this, the gener-

ation of SD (SD) for medical applications has gained

importance not only for augmenting training datasets

and therefore improving the performance of ML mod-

els in EEG based tasks, but also for addressing pa-

tient privacy concerns (Pascual et al., 2021) (Aznan

et al., 2019) (Park et al., 2024) (Carrle et al., 2023)

(Wu et al., 2024) (Fang et al., 2024) (Nia et al., 2024).

Common data augmentation methods include

GANs, Variational Auto-Enconders and Diffusion

Models (Liang et al., 2023). While GANs have made

signiﬁcant strides in SD generation, their application

in the medical domain, particularly for EEG data, re-

mains challenging due to its high variability, which

leads to issues such as mode collapse, complex train-

ing dynamics, and instability (Habashi et al., 2023),

(Liang et al., 2023). Nevertheless, GANs remains the

leading method for SD generation (Liang et al., 2023)

(Carrle et al., 2023) (Fan et al., 2024) (Habashi et al.,

2023). Over the past decade, several GAN archi-

tectures have emerged to address earlier limitations.

Conditional GAN (cGAN), introduced in 2014 (Mirza

and Osindero, 2014), has been effective in enhancing

training stability and reducing mode collapse (Panwar

et al., 2020). (Pascual et al., 2021) utilized cGAN

for synthesizing 4-second EEG signals, while (Car-

rle et al., 2023) applied it for depression diagnosis,

demonstrating the utility of cGANs in generating tar-

geted SD for speciﬁc clinical applications. Similarly,

(Aznan et al., 2019) used Deep Convolutional GANs

(DCGAN) to generate 3-second dry-EEG signals for

a Steady State Visually Evoked Potential task. Other

variations, such as Wasserstein GANs with Gradient

Penalty (wGAN-GP), offer even more stable gradi-

ents and robust training, resulting in higher-quality

SD (Fang et al., 2024) (Nia et al., 2024). Recent ad-

vancements have integrated Long Short Term Mem-

ory (LSTM) layers (Du et al., 2024b) or U-Net gener-

ators (Pascual et al., 2019) (Pan et al., 2023) to better

capture time-series characteristics. Auxiliary decod-

ing techniques, such as those used by (Liang et al.,

2023) and (Aznan et al., 2020) further improve sig-

nal quality by classifying input data alongside distin-

guishing real from SD.

(Hartmann et al., 2018) pioneered EEG sig-

nal synthesis with EEG-GAN, using an improved

wGAN-GP for generating single-channel signals.

This framework utilized convolutional layers for up-

sampling and downsampling, and was evaluated us-

ing Fr

echet Inception Distance, Inception Score, and

sliced Wasserstein Distance (WD) (Xu et al., 2022)

compared different GAN variants, ﬁnding that DCw-

GAN performed best for generating multichannel

EEG data. In recent work, (Wu et al., 2024) intro-

duced a conditional transformer-based wGAN-GP for

synthesizing stereoelectroencephalography (SEEG)

epileptic signals. This approach leveraged trans-

formers to capture time dependencies, with eval-

uation through t-SNE, Cosine Similarity, Jensen-

Shannon Distance, and classiﬁer performance, with

results outperforming traditional augmentation meth-

ods. Similarly, (Du et al., 2024a) proposed a

Deep Convolutional Wasserstein GAN with Gradient-

Penalty (DCwGAN-GP) to generate time-frequency

representations, improving classiﬁcation performance

when combined with real data (RD). Finally, (Cook

et al., 2024) presented an architecture for generating

EEG signals to predict brain age, comparing cGAN,

wGAN, and wGAN-GP models. These were evalu-

ated through the Kolmogorov-Smirnov test and clas-

siﬁer performance, with all models showing improve-

ment when combined SD with RD.

This study aimed to develop a generative model

using GANs to produce synthetic ioECoG data that

captures both epileptic and non-epileptic characteris-

tics. To achieve this, we performed a clustering anal-

ysis on the original dataset to identify distinct data

patterns, which were subsequently processed using

the DCwGAN-GP generative model. The resulting

synthetic signals were evaluated for ﬁdelity, diver-

sity, and utility, ensuring comprehensive assessment

across these domains.

The document is structured as follows: Section

2 details the methods, including data handling, gen-

SyntBioGen 2025 - Special Session on Synthetic biosignals generation for clinical applications

1142

erative process, and evaluation metrics. Section 3

presents the results, followed by their discussion in

Section 4. Finally, Section 5 summarizes the conclu-

sions and proposes future work.

2 METHOD

2.1 Sample

As part of the Synthetic and Scalable Data Platform

for Medical Empowered AI (AISym4Med) project,

conducted in collaboration with the Utrecht Medical

Center (UMCU), this study analyzed data from 18 pa-

tients who underwent intraoperative electrocorticog-

raphy (ioECoG)-guided epilepsy surgery at UMCU,

Netherlands, from 2008 onward. Speciﬁcally, only

pre-resection data were utilized in this analysis.

2.1.1 Participants

The ioECoG data, sourced from the UMCU-SEIN

RESPect database, is stored in a Brain Imaging Data

Structure (BIDS) format (Demuru et al., 2022), with

all patients providing informed consent. Included pa-

tients had ioECoG sampled at 2048 Hz, electrode

placement photos, and were at least 1-year post-

surgical seizure free (Engel 1A). Exclusion criteria in-

cluded prior brain surgery, multiple epileptic foci, or

surgeries involving an amygdala-hippocampectomy.

2.1.2 Procedure

During surgery, the ioECoG was recorded using elec-

trode grids or strips with 1 cm interelectrode dis-

tance. Recordings were made while propofol anes-

thesia was paused to prevent suppression of epilep-

tiform activity (Sun et al., 2024). The ioECoG en-

abled the identiﬁcation of epileptiform patterns in real

time, assisting neurosurgeons by tailoring the extent

of the resection. More details about the data col-

lection can be found in (Hoogteijling et al., 2024).

The pipeline of this study starts with processing and

analysing the raw ioECoG. The signals were clustered

into different groups based on speciﬁc data patterns.

Each cluster is then fed into its own generative model,

producing SD that aligns with the structure deﬁned

during the clustering stage.

2.2 Signal Processing and Analysis

To preserve the statistical integrity of the ioECoG

signals for authentic data generation, they were only

trimmed and ﬁltered, following the recommendations

of (Delorme, 2023). The signals were shortened at

their midpoint, reducing the average length from 360

to 60 seconds, matching the duration used by (van

Klink et al., 2014). A 4th-order Butterworth ﬁlter

was applied (Rasheed and Miften, 2023) (Wu et al.,

2024) with a 0.16 Hz low-cut to remove drift arti-

facts (Miller, 2019) and a 512 Hz high-cut to re-

tain high-frequency oscillations (HFOs) (Zweiphen-

ning et al., 2022). Additionally, an Inﬁnite Impulse

Response Notch ﬁlter at 50 Hz intervals was used

to eliminate powerline noise. Subsequently, an in-

depth analysis of both the time and frequency do-

mains was performed. Given that the patients were

under anesthesia, it was anticipated that speciﬁc fre-

quency bands, particularly the delta and alpha bands,

would exhibit increased power compared to other fre-

quency band(Shin et al., 2020). Accordingly, the rel-

ative power of each frequency band, as well as the

dominant frequency, were systematically examined.

2.3 Clustering

An unsupervised clustering analysis was performed

to group signals into distinct clusters based on their

intrinsic characteristics. This data-driven approach

facilitated a deeper understanding of the variability

within the dataset and enabled the design of a more

tailored generative process. By associating speciﬁc

labels with individual generative models, this method

improved the precision and diversity of SD genera-

tion.

2.3.1 Feature Extraction

For the clustering analysis, 33 features were ex-

tracted using the TSFEL (v0.1.7) (Barandas et al.,

2020), Numpy (v1.26.4) (Harris et al., 2020), and

Scipy (v1.13.0) (Virtanen et al., 2020) libraries, cov-

ering temporal, frequency, time-frequency, and non-

linear domains. Temporal features followed meth-

ods from (Rasheed and Miften, 2023) and (Du et al.,

2024c), known for clustering epileptic EEG signals.

Time-frequency features were based on (Chaibi et al.,

2024), while frequency and non-linear features were

computed speciﬁcally for this study, some including:

Relative Band Power, Dominant Frequency, Hurst

Exponent and Autocorrelation.

Following feature extraction, the features were

standardized using the StandardScaler from Scikit-

Learn (v1.4.1.post1) (Pedregosa et al., 2018) and fea-

tures with a correlation above 95% were removed to

reduce redundancy (Zhou et al., 2022). To identify

key features, the signals were categorized as resected

and non-resected according to the clinicians’ labels,

Intraoperative Electrocorticography Signal Synthesis to Improve the Classiﬁcation of Epileptiform Tissue

1143

and the mean values of each feature were compared

between the groups.

2.3.2 Cluster Analysis

To determine the appropriate number of clusters for

the clustering algorithm, the Elbow method and Sil-

houette analysis were applied. After selecting the

most discriminative features and determining the opti-

mal number of clusters, the Agglomerative Clustering

algorithm from Scikit-Learn (Pedregosa et al., 2018)

with the Ward’s method (Kononenko and Kukar,

2007) was used to classify the signals into distinct

patterns. The clusters were then labelled based on

the signal characteristics and their alignment with the

clinician labels (i.e., resected and non-resected).

2.4 Generative Model

This section describes the proposed generative model,

DCwGAN-GP, covering the input data, architecture,

parameters, and training process. The model was de-

veloped using Pytorch (v2.3.1) (Paszke et al., 2019)

and an NVIDIA RTX 6000 Ada GPU (NVIDIA Cor-

poration, 2024). To generate SD for each cluster,

a separate generative model was created per cluster,

following the same architecture and training process,

with only the batch size differing for the ﬁnal cluster.

2.4.1 Data Preparation

After initial processing, the clustered ioECoG signals

were further prepared for the generative model. This

involved segmenting the 60-second signals into 20-

second windows and downsampling them to 512 Hz

for computational efﬁciency. The signals were then

normalized and transformed into tensors as input for

the model.

2.4.2 Model Architecture

The model architecture, shown in Figure 1, uses 1D

CNNs for both the Generator and Critic, trained with

a Wasserstein Distance loss and Gradient Penalty.

While CNNs are typically utilized for image gen-

eration, 1D CNNs are stated to effectively capture

patterns in one-dimensional EEG data (Aznan et al.,

2019). The Generator processes a latent vector Z of

size 100 through a series of 1D Transposed Convolu-

tional Layers, conﬁgured with a kernel size of 4, stride

of 2, and padding of 1 to upsample the input (Aznan

et al., 2019). Batch normalization and ReLU activa-

tion are applied between layers. A Tanh activation

is used in the ﬁnal layer to generate synthetic sam-

ples, G(z). The Critic processes both real and syn-

thetic samples using 1D Convolutional Layers with

the same conﬁguration of kernel size 4, stride 2, and

padding 1. It uses Spectral Normalization for stability

(Zhong et al., 2023) and Leaky ReLU activation with

alpha 0.2 (Park et al., 2024), except in the ﬁnal block.

The Critic outputs a score distinguishing real from

synthetic samples, with the loss back-propagating to

improve both networks.

2.4.3 Model Training and Tuning

The training process for the DCwGAN-GP ran for a

set number of epochs, processing batches of RD in

each iteration. The Critic was trained ﬁve times for

each Generator update, where each Critic update be-

gan with zeroing its gradients to prevent accumula-

tion. The Critic then computed scores for both real

C(x) and SD C(G(z)), applying the gradient penalty,

calculated as equation 1 (Shu et al., 2023):

GP = λE

x∼p

data

(x)

(∥∇C( ˆx)∥

− 1)

(1)

The Critic’s loss was then calculated using equa-

tion 2 (Shu et al., 2023):

L = E

x∼p

[C(G(z))] − E

x∼p

data

(x)

[C(x)] + GP (2)

This loss aimed to maximize the difference be-

tween scores for real and SD while penalizing large

gradients. The loss was backpropagated, and the

Critic’s parameters were updated using the Adam op-

timizer. For the Generator, its gradients were zeroed,

and a new batch of SD G(z) was generated using the

latent tensor Z. The Critic evaluated this SD, produc-

ing a score C(G(z)). The generator’s loss is then cal-

culated as 3 (Shu et al., 2023):

loss

= −E[C(G(z))] (3)

The loss was backpropagated, and the Generator’s

parameters were updated using the Adam optimizer.

At the end of each epoch, learning rate schedulers ad-

justed the learning rates for both models to optimize

training progression. Following training, the model

underwent hyperparameter tuning over 51 iterations

to achieve optimal performance. The ﬁnal version

included six layers in the Generator and ﬁve in the

Critic. The training parameters are detailed in Table

1. The batch size was set to 40, with exception on the

ﬁnal model, which was reduced to 30.

2.5 Evaluation Metrics

This section outlines the evaluation metrics used to

assess the quality of the generated ioECoG signals,

categorized into ﬁdelity, diversity and utility.

SyntBioGen 2025 - Special Session on Synthetic biosignals generation for clinical applications

1144

Figure 1: Generative model’s architecture. Each Neural Network is comprised of three types of blocks: input, middle and

ﬁnal block. The input and ﬁnal block are constituted by one layer, while the middle block was constituted by several layers.

The number of layers contained in the middle block was deﬁned through the optimization of the model.

Table 1: Training Parameters for DCwGAN-GP.

Training Parameters

Number of Epochs 2000

Batch Size 40

Learning Rate 0.00005

Critic Interactions 5

Gradient Penalty Lambda 10

Latent Dimension 100

Number of samples Data points in each cluster

2.5.1 Fidelity Analysis

Fidelity analysis evaluates how closely synthetic sig-

nals resemble real ones across time, frequency, and

time-frequency domains without replicating them ex-

actly (Naeem et al., 2020) (Figueira and Vaz, 2022).

In this dimension, we calculated metrics for three

pairings: real-real, real-synthetic, and synthetic-

synthetic. The real-real pairing serves as a reference

to evaluate the ﬁdelity of the synthetic data. To per-

form this analysis, each sample was compared with

every other sample within its respective pairing group.

In the time domain, the mean and standard devi-

ation (STD) of both real and synthetic signals were

evaluated, and the Wasserstein Distance (WD) was

calculated to quantify differences in their probabil-

ity distributions. In the frequency domain, the Power

Spectral Density (PSD) was computed for both sets

of signals to ensure similar spectral characteristics,

and a plot was created to visually compare the power

distribution across frequencies. Finally, for the time-

frequency domain, scalograms were generated using

the Morlet Wavelet Transform to capture dynamic

changes in frequency content over time. Similarity

between real and synthetic scalograms was quanti-

ﬁed using the Pearson Correlation Coefﬁcient, Co-

sine Similarity, Structural Similarity Index (SSIM)

and Mean Squared Error (MSE). This metrics were

calculated with the Scipy (Virtanen et al., 2020) and

Scikit-Learn Image Processing libraries (van der Walt

et al., 2014).

2.5.2 Diversity Analysis

To evaluate the diversity of SD in comparison to

RD, Principal Component Analysis (PCA) and t-

distributed Stochastic Neighbor Embedding (t-SNE)

were employed as dimensionality reduction tech-

niques. These methods enabled a visual comparison

of the distribution of synthetic samples against the

original data (Jansen, 2020).

2.5.3 Utility Analysis

To evaluate the utility of the SD, several classiﬁca-

tion models were developed and trained on three

datasets: RD, SD, and a combination of both. The

aim was to determine if the SD could improve

the model’s ability to distinguish between clusters,

with better performance on the combined dataset

indicating the utility of the SD. The statistical

signiﬁcance of the observed differences was then

evaluated using a Z-test, which compares the per-

formance metrics between datasets to determine

if the observed variations are likely due to chance.

Model Training and Evaluation

The same features used in the clustering process (Sec-

tion 2.3.1) were applied in the classiﬁers. The data

was split into 70% training and 30% testing sets, us-

ing stratiﬁed sampling to maintain class proportions.

Various ML classiﬁers were developed using the

Intraoperative Electrocorticography Signal Synthesis to Improve the Classiﬁcation of Epileptiform Tissue

1145

Scikit-Learn library (Pedregosa et al., 2018), includ-

ing Logistic Regression, Support Vector Machines

(SVM), Random Forest, Neural Network Multi-Layer

Perceptron, Gradient Boosting, and Balanced Ran-

dom Forest. These models were chosen for their abil-

ity to handle imbalanced datasets (Chen et al., 2004),

capture complex non-linear patterns (Jaiswal, 2024),

and perform well with tabular data (Tuychiev, 2023),

with the ﬁrst two serving as baselines for compar-

ison. A 10-fold Stratiﬁed Cross-Validation method

was used with the shufﬂe parameter set to ”True” to

minimize bias from any inherent order, ensuring the

models generalization to unseen data and avoiding

overﬁtting. Performance metrics such as Accuracy,

F1-Score, Precision, Recall and AUC were used to

compare the models’ effectiveness.

3 RESULTS

3.1 Clustering Analysis

Three clusters were chosen for the Agglomerative

Clustering algorithm. Based on the alignment with

the clinician binary classiﬁcations (non-resected and

resected tissue) and the data characteristics, the clus-

ters were labelled as: Cluster 0 (Noisy), Cluster 1

(Epileptic), and Cluster 2 (Non-Epileptic). Speciﬁ-

cally, Cluster 0 consisted of 28 signals, with 57.14%

of the data points classiﬁed as non-resected and

42.86% as resected. Cluster 1 included 71 sig-

nals, comprising 39.89% non-resected and 61.11%

resected data points. Cluster 2 contained 31 signals,

of which 83.87% were non-resected and 16.13% were

resected. Some overlap occurred in the 3D scatter

plots between Clusters 0 and 1, and Clusters 1 and 2.

Figure 2 illustrates three representative signals from

each cluster.

Figure 2: Illustration of three representative signals of each

cluster. The signals were normalized, therefore the ampli-

tude is in arbitrary units.

Cluster 0 exhibited irregular and noisy ﬂuctua-

tions with no clear structure. Cluster 1, on the other

hand, had more structured patterns, while Cluster 2

showed a level of organization higher than Cluster 0,

but its patterns were still not as structured as those

in Cluster 1. Furthermore, the frequency analysis re-

vealed that Cluster 0 had a low PSD with no dominant

frequency peaks (as shown in Figure 5), reﬂecting its

noisy nature. In contrast, Cluster 1 demonstrated a

strong delta band peak with higher variability in the

PSD. Cluster 2 showed consistent delta-band peaks,

although with less variability compared to Cluster 1.

3.2 Evaluation of Synthetic Data

3.2.1 Fidelity Analysis

Time Analysis

The mean signals and their respective STD for real

and SD across Clusters 0, 1, and 2, depicted in Fig-

ure 3, reveal consistent waveform patterns. However,

synthetic signals in Cluster 0 display a slightly higher

amplitude and increased noise compared to their real

counterparts. In turn, the WD analysis illustrated in

Figure 4 shows relatively small differences between

the data pairings (real-real (RR), real-synthetic (RS)

and synthetic-synthetic (SS)). WD is highest for the

RR pairing, followed by RS, and lowest for SS. Clus-

ter 1 consistently has the highest WD across all pair-

ings, Cluster 2 is moderate, and Cluster 0 has the low-

est values. For the RS pairing, the WD values were

0.043 ± 0.025 for Cluster 0, 0.065 ± 0.045 for Cluster

1, and 0.045 ± 0.036 for Cluster 2.

Frequency Analysis

In the frequency domain depicted in Figure 5,

the synthetic signals match closely with the real

ones in low-frequency behavior, especially in the

delta band. However, the SD exhibit lower variabil-

ity, as indicated by their lower STD across all clusters.

Time-Frequency Analysis

Figure 6 evaluates the similarity between real and

synthetic scalograms across real-real (RR), real-

synthetic (RS), and synthetic-synthetic (SS) pairings.

Cosine Similarity increased from 0.729 ± 0.037 (RR)

to 0.795 ± 0.005 (SS), indicating greater similarity

within SD. SSIM and Pearson’s Correlation followed

a similar pattern, with Cluster 1 and Cluster 2 scor-

ing higher than Cluster 0. The Pearson correlation of

cluster 0 improved signiﬁcantly in SS (+0.275), but

remained lower overall compared to other clusters.

Cluster 0 consistently exhibited the lowest similar-

ity across all metrics, while Clusters 1 and 2 demon-

strated moderate-to-high similarity, particularly in RS

pairings, where Pearson’s Correlation reached 0.694

± 0.087 and 0.672 ± 0.063, and Cosine Similarity

scored 0.757 ± 0.066 and 0.747 ± 0.051, respectively.

MSE showed a clear reduction, decreasing by 0.182

SyntBioGen 2025 - Special Session on Synthetic biosignals generation for clinical applications

1146

Figure 3: Comparison of the mean and respective Standard Deviation of the signals between real and synthetic clusters. The

signals are normalized, therefore the amplitude is in arbitrary units.

Figure 4: Comparison of the Wasserstein Distance between

different data pairing. In this case, this metric is being

compared between the real and synthetic (RS) signals, and

within the real (RR) and synthetic (SS) signals separately.

± 0.052 from RR to SS, with Cluster 1 achieving the

lowest error in RS (0.499 ± 0.078). Despite these

trends, the differences between RR and SS pairings

remained relatively modest overall, with Cluster 0

again showing the largest variations in SSIM (0.180

± 0.051) and Pearson’s Correlation (0.275 ± 0.119).

These results highlight greater internal consistency

within SD and the varying performance of clusters,

with Clusters 1 and 2 outperforming Cluster 0 across

all similarity metrics.

3.2.2 Diversity Analysis

In the PCA results illustrated in Figure 7 (upper row),

the SD overlap with the RD, replicating the overall

structure but is more concentrated in the center, sug-

gesting less diversity. In the t-SNE plots (second

line), synthetic signals are more scattered and less

grouped, indicating they fail to capture the local re-

lationships present in the RD.

3.2.3 Utility Analysis

The Logistic Regression model yielded the best re-

sults. Table 2 presents improvements in four per-

formance metrics and a slight decline in AUC when

trained on combined real and SD (RS) versus training

on RD alone (R). However, these differences were not

statistically signiﬁcant. Additional classiﬁcation re-

sults are available in the Appendix.

Table 2: Logistic Regression Scores. Comparison between

the scores of the model trained on the combined data (RS)

and tested on real (R), with the model trained and tested

solely on the RD.

Train Test Train Test

R R RS R

Accuracy 0.89 0.85 0.93 0.88

Precision 0.89 0.73 0.92 0.76

Recall 0.92 0.80 0.94 0.81

F1-Score 0.90 0.75 0.93 0.78

AUC 0.97 0.94 0.99 0.92

4 DISCUSSION

This study aimed to develop a generative model in

order to produce synthetic ioECoG data from ioECoG

signals of patients undergoing epilepsy surgery.

The analysis of the real ioECoG dataset revealed

three distinct clusters, each associated with unique

signal patterns. Cluster 1 could be linked to patho-

logical signals, displaying epileptiform activity char-

acterized by spikes and slow-wave complexes, simi-

lar to those described by (Li et al., 2020). Moreover,

Cluster 1 and Cluster 2, labelled as epileptic and non-

epileptic, showed a strong alignment with the clinical

Intraoperative Electrocorticography Signal Synthesis to Improve the Classiﬁcation of Epileptiform Tissue

1147

Figure 5: Comparison of the mean PSD and respective STD between real and synthetic clusters.

Figure 6: Similarity metrics in the time-frequency domain between different data pairing: real-real (RR), real-synthetic (RS)

and synthetic-synthetic (SS).

Figure 7: Diversity evaluation. The upper row is correspondent to the PCA evaluation, while the second line is correspondent

to the t-SNE evaluation. The columns correspond to each cluster. The SD is represented by black dots, while the RD is

represented by gray crosses.

categories, matching 61.1 and 83.9% of the clinician’s

classiﬁcations, respectively. This alignment was also

supported by the study of (Gajic et al., 2015), who

showcased epileptic signals with more spaced events

and with fewer ﬂuctuations, while non-epileptic sig-

nals showed more noise and greater variability. In

turn, Cluster 0 contained a mix of non-epileptic and

epileptic signals, likely due to contamination from

neighboring channels within the same electrode grid,

as observed in other intracranial EEG studies (Sindhu

et al., 2023). Additionally, this overlap may simply

reﬂect the random nature of noise, which can occur

with similar frequency in both types of tissue. Its ori-

gin may be attributed to machine artifacts, which can

mimic pathological signals and interfere with accu-

rate classiﬁcation (Nejedly et al., 2020). Having a

SyntBioGen 2025 - Special Session on Synthetic biosignals generation for clinical applications

1148

noisy cluster could also aid clinical implementation,

as it removes the need to exclude channels affected

by artifacts, simplifying the process and preserving

data integrity. In the frequency domain, all clusters

showed a peak in the delta band, as expected dur-

ing propofol-induced deep sleep (Moini and Piran,

2020), with Cluster 0 also exhibiting an alpha peak

that might be related to propofol use (Purdon et al.,

2015).

Overall, in Figure 3, the synthetic signals closely

match the real ones in amplitude and waveform, in-

dicating that the model effectively captures the global

temporal patterns. Minor STD deviations, similar to

those reported by (Hartmann et al., 2018), suggest

slight differences in variability between real and syn-

thetic signals. In turn, WD scores showed that syn-

thetic signals preserved the similarity trends across

clusters, with Cluster 1 having the highest WD prob-

ably due to its distinct epileptiform patterns, followed

by Clusters 2 and 0, where Cluster 0 displayed the

greatest similarity with lower WD values. Slightly

lower SS values suggested reduced variability in syn-

thetic signals. RS WD scores in this study seem to

outperform benchmarks from the literature, such as

0.078 reported by (Hartmann et al., 2018) and 0.450

by (Xu et al., 2022). Furthermore, the 20-second syn-

thetic time-series offers a more realistic representa-

tion compared to the shorter windows commonly used

in other studies using scalp EEG data —such as 3 sec-

onds (Aznan et al., 2019), 4 seconds (Pascual et al.,

2019), and even as short as 0.05 seconds (Pan et al.,

2023)—allowing for improved ﬁdelity and signal di-

versity.

Figure 5 shows that the synthetic signals closely

match the real ones in power distribution, mirror-

ing their trend particularly in the delta band, despite

the real signals displaying higher variability (STD).

These results align with those reported by (Park et al.,

2024) and (Hartmann et al., 2018), where synthetic

PSDs showed similar prominent peaks to their real

counterparts, with slight power variations. In contrast,

(Carrle et al., 2023) found that synthetic PSDs did not

fully capture the prominent peaks of RD, indicating a

stronger reliability from the generative model in this

study. However, the lower STD in synthetic PSDs

suggests reduced variability compared to the real sig-

nals.

In the evaluation of the time-frequency domain

(Figure 6), the lower similarity values in the RR pair-

ing across all metrics reﬂect greater variability in the

RD compared to the higher SS scores, indicating that

synthetic signals display higher inter-similarity and

fail to capture the full variability of the RD. This is

most evident in Cluster 0, which performed worst in

all metrics but showed the largest improvement from

RR to SS, particularly in SSIM and Pearson’s Cor-

relation. However, the differences between RR and

SS pairings are relatively small. Although Cluster 0

had the lowest RS scores, they were consistent with

the RR pairing, suggesting that while the SD does not

fully capture RD variability, it still maintains strong

correlations with the real signals. To the best of our

knowledge, this type of analysis was not found in

comparable studies (Wu et al., 2024), which focused

on signal similarity rather than time-frequency repre-

sentations.

The limited variability of the synthetic signals is

conﬁrmed by the diversity evaluation results. While

synthetic signals replicated most real patterns, as

shown by the overlap in the PCA plot, their concentra-

tion toward the center indicates only partial diversity

capture, with the tails of the real distributions being

overlooked. Similarly, the t-SNE plots reﬂect ﬁndings

from (Wu et al., 2024), though our study showed less

overlap, highlighting a lack of diversity in the gener-

ated samples. This issue likely stems from misclas-

siﬁcations during clustering, particularly in Cluster 1,

introducing unwanted variability in the RD. During

training, the Critic consistently labelled these mis-

classiﬁed signals as synthetic, causing the generator

to avoid synthesizing them. As a result, the synthetic

signals focused on common patterns and lacked the

broader variability seen in the RD, especially at the

distribution tails.

All classiﬁers showed improved performance

across Accuracy, Precision, Recall, and F1-score

when using the augmented ioECoG dataset, high-

lighting the utility of the generated signals. The Lo-

gistic Regression Model showcased the best results

when trained on the augmented dataset, but tested

only with the real signals. The 3% increase in accu-

racy aligns with improvements reported in other stud-

ies, such as 2.51% (Du et al., 2024a), 5% (Xu et al.,

2022), and 6% (Wu et al., 2024). However, the de-

crease in AUC may indicate that the SD exhibit dis-

tributional differences compared to RD, conﬁrming

their inability to fully capture the variability and com-

plexity of the RD.

Finally, a comprehensive visual review and vali-

dation of all generated signals remains necessary to

conﬁrm their suitability for clinical applications.

5 CONCLUSIONS AND FUTURE

WORK

This study developed a generative model, DCwGAN-

GP, capable of synthesizing realistic ioECoG signals

Intraoperative Electrocorticography Signal Synthesis to Improve the Classiﬁcation of Epileptiform Tissue

1149

from epileptic and non-epileptic tissue with the po-

tential to enhance ML models and support epilepsy

surgery treatment. The generated signals replicated

the key characteristics of the RD across time, fre-

quency, and time-frequency domains, contributing to

improved performance in ML classiﬁers trained to

distinguish epileptic, non-epileptic, and noisy signals.

The 3% performance improvement demonstrated the

utility of the SD in enhancing classiﬁer performance.

While the DCwGAN-GP model demonstrated re-

liable ﬁdelity in capturing temporal patterns within in-

dividual channels, it did not account for inter-channel

temporal dynamics or spatial relationships between

electrodes. Incorporating techniques such as Long

Short-Term Memory (LSTM) networks, Recurrent

Neural Networks (RNNs) or Transformers for tempo-

ral dependencies, along with graph-based models for

spatial relationships, could further enhance the real-

ism and clinical relevance of the generated signals. In

addition, clinical validation of each signal should be

considered to ensure reliability and clinical applica-

bility.

Future work could focus on improving the ground-

truth labelling of signals to enhance accurate dis-

tinction of epileptiform activity, thereby strengthen-

ing the generative process. Machine Learning and

Deep Clustering techniques hold promise for reﬁn-

ing pattern recognition within ioECoG data, poten-

tially reducing misclassiﬁcations and achieving more

nuanced synthetic signals that reﬂect real-world di-

versity. Another critical direction involves explicitly

modeling spatial and temporal dependencies across

signals.

ACKNOWLEDGEMENTS

This work was supported by the European Union’s

Horizon Europe research and innovation programme

under grant agreement No. 101095387: AISym4Med

– Synthetic and Scalable Data Platform for Med-

ical Empowered AI (HORIZON-HLTH-2022-IND-

13), by the European Research Council (ERC) under

grant No. 803880, by EpilepsieNL, the Christelijke

Verening voor de Verpleging van Lijders aan Epilep-

sie, and a VIDI grant number 09150172210057.

REFERENCES

Aznan, N. K. N., Atapour-Abarghouei, A., Bonner, S.,

Connolly, J., Moubayed, N. A., and Breckon, T.

(2019). Simulating brain signals: Creating synthetic

eeg data via neural-based generative models for im-

proved ssvep classiﬁcation. In International Joint

Conference on Neural Networks (IJCNN). IEEE.

Aznan, N. K. N., Atapour-Abarghouei, A., Bonner, S., Con-

nolly, J. D., and Breckon, T. P. (2020). Leveraging

synthetic subject invariant eeg signals for zero cali-

bration bci.

Barandas, M., Folgado, D., Fernandes, L., Santos, S.,

Abreu, M., Bota, P., Liu, H., Schultz, T., and Gam-

boa, H. (2020). Tsfel: Time series feature extraction

library. SoftwareX, 11:100456.

Carrle, F. P., Hollenbenders, Y., and Reichenbach, A.

(2023). Generation of synthetic eeg data for training

algorithms supporting the diagnosis of major depres-

sive disorder. Frontiers in Neuroscience, 17.

Chaibi, S., Mahjoub, C., Ayadi, W., and Kachouri,

A. (2024). Epileptic eeg patterns recognition

through machine learning techniques and relevant

time–frequency features. Biomedical Engineering /

Biomedizinische Technik, 69:111–123.

Chen, C., Liaw, A., and Breiman, L. (2004). Using ran-

dom forest to learn imbalanced data. Technical report,

Department of Statistics, UC Berkley.

Consales, A., Casciato, S., Asioli, S., Barba, C., Caulo, M.,

Colicchio, G., Cossu, M., de Palma, L., Morano, A.,

Vatti, G., Villani, F., Zamponi, N., Tassi, L., Gennaro,

G. D., and Marras, C. E. (2021). The surgical treat-

ment of epilepsy. Neurological Sciences, 42:2249–

2260.

Cook, Z., Sinha, G., Wang, J., Zhao, C., Belacel, N., Does-

burg, S., Medvedev, G., Ribary, U., Vakorin, V., and

Xi, P. (2024). Enhancing brain age prediction: A gen-

erative ai approach for eeg machine learning models.

In 2024 IEEE International Instrumentation and Mea-

surement Technology Conference (I2MTC), pages 1–

6. IEEE.

Delorme, A. (2023). Eeg is better left alone. Scientiﬁc Re-

ports, 13:2372.

Demuru, M., van Blooijs, D., Zweiphenning, W., Hermes,

D., Leijten, F., and Zijlmans, M. (2022). A prac-

tical workﬂow for organizing clinical intraoperative

and long-term ieeg data in bids. Neuroinformatics,

20:727–736.

Du, X., Ding, X., Xi, M., Lv, Y., Qiu, S., and Liu, Q.

(2024a). A data augmentation method for motor im-

agery eeg signals based on dcgan-gp network. Brain

Sciences, 14:375.

Du, X., Wang, X., Zhu, L., Ding, X., Lv, Y., Qiu, S., and

Liu, Q. (2024b). Electroencephalographic signal data

augmentation based on improved generative adversar-

ial network. Brain Sciences, 14.

Du, Y., Li, G., Wu, M., and Chen, F. (2024c). Unsupervised

multivariate feature-based adaptive clustering analysis

of epileptic eeg signals. Brain Sciences, 14:342.

Duncan, J. S. and Taylor, P. N. (2023). Optimising epilepsy

surgery. The Lancet Neurology, 22:373–374.

Fan, Y., Wang, B., and Zhang, T. (2024). A dual-

discriminator gan for sleep eeg signal synthesis. Jour-

nal Of Bioinformatics and Neuroscience (JBINS).

Fang, L., Li, Y., Shao, M., Yu, A., Felemban, B. F., Aly,

A. A., Rani, S., and Lyu, X. (2024). Enhancing med-

SyntBioGen 2025 - Special Session on Synthetic biosignals generation for clinical applications

1150

ical signal processing and diagnosis with ai-generated

content techniques. IEEE Journal of Biomedical and

Health Informatics.

Fern

andez, I. S. and Loddenkemper, T. (2013). Electro-

corticography for seizure foci mapping in epilepsy

surgery. Journal of Clinical Neurophysiology,

30(6):554–570.

Figueira, A. and Vaz, B. (2022). Survey on synthetic data

generation, evaluation methods and gans. Mathemat-

ics, 10:1–41.

Gajic, D., Djurovic, Z., Gligorijevic, J., Gennaro, S. D.,

and Savic-Gajic, I. (2015). Detection of epileptiform

activity in eeg signals based on time-frequency and

non-linear analysis. Frontiers in Computational Neu-

roscience, 9.

Greiner, H. M., Horn, P. S., Tenney, J. R., Arya, R., Jain,

S. V., Holland, K. D., Leach, J. L., Miles, L., Rose,

D. F., Fujiwara, H., and Mangano, F. T. (2016). Prere-

section intraoperative electrocorticography (ecog) ab-

normalities predict seizure-onset zone and outcome in

pediatric epilepsy surgery. Epilepsia, 57(4):582–589.

Habashi, A. G., Azab, A. M., Eldawlatly, S., and Aly, G. M.

(2023). Generative adversarial networks in eeg anal-

ysis: an overview. Journal of NeuroEngineering and

Rehabilitation, 20:40.

Harris, C. R., Millman, K. J., van der Walt, S. J., Gommers,

R., Virtanen, P., Cournapeau, D., Wieser, E., Taylor,

J., Berg, S., Smith, N. J., Kern, R., Picus, M., Hoyer,

S., van Kerkwijk, M. H., Brett, M., Haldane, A., del

ıo, J. F., Wiebe, M., Peterson, P., G

erard-Marchant,

P., Sheppard, K., Reddy, T., Weckesser, W., Abbasi,

H., Gohlke, C., and Oliphant, T. E. (2020). Array pro-

gramming with NumPy. Nature, 585(7825):357–362.

Hartmann, K. G., Schirrmeister, R. T., and Ball, T. (2018).

Eeg-gan: Generative adversarial networks for elec-

troencephalograhic (eeg) brain signals. Electrical En-

gineering and Systems Science.

Hoogteijling, S., Schaft, E. V., Dirks, E. H., Straumann,

S., Demuru, M., van Eijsden, P., Gebbink, T., Otte,

W. M., Huiskamp, G. M., van’t Klooster, M. A., et al.

(2024). Machine learning for (non-) epileptic tissue

detection from the intraoperative electrocorticogram.

Clinical Neurophysiology, 167:14–25.

Jaiswal, S. (2024). Multilayer perceptrons in machine learn-

ing: A comprehensive guide.

Jansen, S. (2020). Machine Learning for Algorithmic Trad-

ing: Predictive models to extract signals from market

and alternative data for systematic trading strategies

with Python. Packt Publishing, 2nd edition.

Kononenko, I. and Kukar, M. (2007). Chapter 12 - cluster

analysis. In Kononenko, I. and Kukar, M., editors,

Machine Learning and Data Mining, pages 321–358.

Woodhead Publishing.

Li, Q., Gao, J., Zhang, Z., Huang, Q., Wu, Y., and Xu, B.

(2020). Distinguishing epileptiform discharges from

normal electroencephalograms using adaptive fractal

and network analysis: A clinical perspective. Fron-

tiers in Physiology, 11:828.

Liang, S., Kuang, S., Wang, D., Yuan, Z., Zhang, H., and

Sun, L. (2023). An auxiliary synthesis framework for

enhancing eeg-based classiﬁcation with limited data.

IEEE Transactions on Neural Systems and Rehabili-

tation Engineering, 31:2120–2131.

Miller, K. J. (2019). A library of human electrocortico-

graphic data and analyses. Nature Human Behaviour,

3:1225–1235.

Mirza, M. and Osindero, S. (2014). Conditional generative

adversarial nets. ArXiv.

Moini, J. and Piran, P. (2020). Chapter 6 - cerebral cor-

tex. In Moini, J. and Piran, P., editors, Functional and

Clinical Neuroanatomy, pages 177–240. Academic

Press.

Naeem, M. F., Oh, S. J., Uh, Y., Choi, Y., and Yoo, J.

(2020). Reliable ﬁdelity and diversity metrics for gen-

erative models. In III, H. D. and Singh, A., editors,

Proceedings of the 37th International Conference on

Machine Learning, volume 119 of Proceedings of Ma-

chine Learning Research, pages 7176–7185. PMLR.

Nejedly, P., Kremen, V., Sladky, V., Cimbalnik, J., Klimes,

P., Plesinger, F., Mivalt, F., Travnicek, V., Viscor, I.,

Pail, M., Halamek, J., Brinkmann, B. H., Brazdil, M.,

Jurak, P., and Worrell, G. (2020). Multicenter intracra-

nial eeg dataset for classiﬁcation of graphoelements

and artifactual signals. Scientiﬁc Data, 7(1):179.

Nia, A. F., Tang, V., Talou, G. M., and Billinghurst, M.

(2024). Synthesizing affective neurophysiological sig-

nals using generative models: A review paper. Journal

of Neuroscience Methods, 406.

NVIDIA Corporation (2024). Nvidia rtx 6000 ada gener-

ation graphics card. https://www.nvidia.com/en-us/

design-visualization/rtx-6000/.

Pan, Y., Li, N., Zhang, Y., Xu, P., and Yao, D. (2023). Short-

length ssvep data extension by a novel generative ad-

versarial networks based framework. Cognitive Neu-

rodynamics.

Panwar, S., Rad, P., Jung, T.-P., and Huang, Y. (2020). Mod-

eling eeg data distribution with a wasserstein genera-

tive adversarial network to predict rsvp events. IEEE

Transactions on Neural Systems and Rehabilitation

Engineering, 28:1720–1730.

Park, J., Mahey, P., and Adeniyi, O. (2024). Improving eeg

signal classiﬁcation accuracy using wasserstein gener-

ative adversarial networks.

Pascual, D., Aminifar, A., Atienza, D., Ryvlin, P., and Wat-

tenhofer, R. (2019). Synthetic epileptic brain activities

using generative adversarial networks. arXiv preprint

arXiv:1907.10518.

Pascual, D., Amirshahi, A., Aminifar, A., Atienza, D.,

Ryvlin, P., and Wattenhofer, R. (2021). Epilepsygan:

Synthetic epileptic brain activities with privacy preser-

vation. IEEE Transactions on Biomedical Engineer-

ing, 68:2435–2446.

Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J.,

Chanan, G., Killeen, T., Lin, Z., Gimelshein, N.,

Antiga, L., Desmaison, A., Kopf, A., Yang, E., De-

Vito, Z., Raison, M., Tejani, A., Chilamkurthy, S.,

Steiner, B., Fang, L., Bai, J., and Chintala, S. (2019).

Pytorch: An imperative style, high-performance deep

learning library. In Advances in Neural Information

Intraoperative Electrocorticography Signal Synthesis to Improve the Classiﬁcation of Epileptiform Tissue

1151

Processing Systems 32, pages 8024–8035. Curran As-

sociates, Inc.

Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V.,

Thirion, B., Grisel, O., Blondel, M., M

uller, A.,

Nothman, J., Louppe, G., Prettenhofer, P., Weiss, R.,

Dubourg, V., Vanderplas, J., Passos, A., Cournapeau,

D., Brucher, M., Perrot, M., and

Edouard Duchesnay

(2018). Scikit-learn: Machine learning in python.

Purdon, P. L., Sampson, A., Pavone, K. J., and Brown, E. N.

(2015). Clinical Electroencephalography for Anesthe-

siologists: Part I: Background and Basic Signatures.

Anesthesiology, 123(4):937–960.

Rasheed, S. S. and Miften, F. S. (2023). Improve of neona-

tal seizure detection based on eeg signal using k-mean

clustering. In 2023 Al-Sadiq International Conference

on Communication and Information Technology (AIC-

CIT), pages 181–184. IEEE.

Saminu, S., Xu, G., Shuai, Z., Kader, I. A. E., Jabire, A. H.,

Ahmed, Y. K., Karaye, I. A., and Ahmad, I. S. (2022).

Application of deep learning and wt-sst in localization

of epileptogenic zone using epileptic eeg signals. Ap-

plied Sciences, 12:4879.

Shin, H. W., Kim, H. J., Jang, Y. K., You, H. S., Huh, H.,

Choi, Y. J., Choi, S. U., and Hong, J. S. (2020). Mon-

itoring of anesthetic depth and eeg band power using

phase lag entropy during propofol anesthesia. BMC

Anesthesiology, 20:49.

Shoka, A. A. E., Dessouky, M. M., El-Sayed, A., and Hem-

dan, E. E.-D. (2023). Eeg seizure detection: concepts,

techniques, challenges, and future trends. Multimedia

Tools and Applications, 82:42021–42051.

Shu, K., Zhao, Y., Wu, L., Liu, A., Qian, R., and Chen, X.

(2023). Data augmentation for seizure prediction with

generative diffusion model.

Sindhu, K. R., Ngo, D., Ombao, H., Olaya, J. E., Shrey,

D. W., and Lopour, B. A. (2023). A novel method

for dynamically altering the surface area of intracra-

nial eeg electrodes. Journal of Neural Engineering,

20(2):026002.

Sun, D., van ’t Klooster, M. A., Ringeling, E. M., Schaft,

E. V., van Rijen, P. C., Leijten, F. S., Demuru, M.,

Robe, P. A., Hoff, R. G., and Zijlmans, M. (2024).

Pausing propofol during neurosurgery to record in-

traoperative electrocorticography is feasible;10 years

of clinical experience. Clinical Neurophysiology,

167:84–91.

Tatum, W. O., editor (2021). Handbook of EEG Interpreta-

tion. Springer Publishing Company.

Tuychiev, B. (2023). A guide to the gradient boost-

ing algorithm. https://www.datacamp.com/tutorial/

guide-to-the-gradient-boosting-algorithm. Accessed:

2024-09-07.

van der Walt, S., Sch

onberger, J. L., Nunez-Iglesias, J.,

Boulogne, F., Warner, J. D., Yager, N., Gouillart, E.,

and Yu, T. (2014). scikit-image: image processing in

python. PeerJ, 2:e453.

van Klink, N., van’t Klooster, M., Zelmann, R., Leijten, F.,

Ferrier, C., Braun, K., van Rijen, P., van Putten, M.,

Huiskamp, G., and Zijlmans, M. (2014). High fre-

quency oscillations in intra-operative electrocorticog-

raphy before and after epilepsy surgery. Clinical Neu-

rophysiology, 125(11):2212–2219.

Virtanen, P., Gommers, R., Oliphant, T. E., Haberland, M.,

Reddy, T., Cournapeau, D., Burovski, E., Peterson, P.,

Weckesser, W., Bright, J., van der Walt, S. J., Brett,

M., Wilson, J., Millman, K. J., Mayorov, N., Nel-

son, A. R. J., Jones, E., Kern, R., Larson, E., Carey,

C. J., Polat,

I., Feng, Y., Moore, E. W., VanderPlas,

J., Laxalde, D., Perktold, J., Cimrman, R., Henriksen,

I., Quintero, E. A., Harris, C. R., Archibald, A. M.,

Ribeiro, A. H., Pedregosa, F., van Mulbregt, P., and

SciPy 1.0 Contributors (2020). SciPy 1.0: Fundamen-

tal Algorithms for Scientiﬁc Computing in Python.

Nature Methods, 17:261–272.

Wang, Z., Guo, J., van’t Klooster, M., Hoogteijling, S., Ja-

cobs, J., and Zijlmans, M. (2024). Prognostic value

of complete resection of the high-frequency oscilla-

tion area in intracranial eeg: A systematic review and

meta-analysis. Neurology, 102(9):e209216.

WHO (2019). Epilepsy: a public health imperative. World

Health Organization.

Wu, X., Zhang, D., Li, G., Gao, X., Metcalfe, B.,

and Chen, L. (2024). Data augmentation for in-

vasive brain-computer interfaces based on stereo-

electroencephalography (seeg). Journal of Neural En-

gineering.

Xu, Y., Yang, J., and Sawan, M. (2022). Multichannel syn-

thetic preictal eeg signals to enhance the prediction of

epileptic seizures. IEEE Transactions on Biomedical

Engineering, 69:3516–3525.

Zhong, H., Yu, S., Trinh, H., Lv, Y., Yuan, R., and Wang,

Y. (2023). Fine-tuning transfer learning based on dc-

gan integrated with self-attention and spectral nor-

malization for bearing fault diagnosis. Measurement,

210:112421.

Zhou, H., Wang, X., and Zhu, R. (2022). Feature selection

based on mutual information with correlation coefﬁ-

cient. Applied Intelligence, 52:5457–5474.

Zijlmans, M., Zweiphenning, W., and van Klink, N.

(2019). Changing concepts in presurgical assess-

ment for epilepsy surgery. Nature Reviews Neurology,

15(10):594–606.

Zweiphenning, W., van ’t Klooster, M. A., van Klink, N.

E. C., Leijten, F. S. S., Ferrier, C. H., Gebbink, T.,

Huiskamp, G., van Zandvoort, M. J. E., van Schoon-

eveld, M. M. J., Bourez, M., Goemans, S., Straumann,

S., van Rijen, P. C., Gosselaar, P. H., van Eijsden,

P., Otte, W. M., van Diessen, E., Braun, K. P. J., Zi-

jlmans, M., Bloemen-Carlier, E. M., Cibulkov

a, V.,

de Munnink, R., van der Salm, S., Eijkemans, M. J.,

van Eck, J. M. O., Velders, A., van Asch, C. J., Zwem-

mer, J., van Regteren-van Griethuysen, R., Smeding,

H., van der Berg, L., de Bresser, J., de Kort, G. A.,

and Dankbaar, J.-W. (2022). Intraoperative elec-

trocorticography using high-frequency oscillations or

spikes to tailor epilepsy surgery in the netherlands (the

hfo trial): a randomised, single-blind, adaptive non-

inferiority trial. The Lancet Neurology, 21:982–993.

SyntBioGen 2025 - Special Session on Synthetic biosignals generation for clinical applications

1152

APPENDIX

Table 3: Comparison of classiﬁcation models trained on combined data (RS) and tested on real data (R) versus models trained

and tested solely on RD.

Models

Metric Random Forest SVM Neural Network Gradient Boosting Balanced RF

Accuracy

tR: 1.0 TR: 0.87 tR: 0.89 TR: 0.85 tR: 1.0 TR: 0.77 tR: 1.0 TR: 0.85 tR: 0.94 TR: 0.84

tRS: 1.0 TR: 0.87 tRS: 0.95 TR: 0.85 tRS: 1.0 TR: 0.83 tRS: 1.0 TR: 0.80 tRS: 0.97 TR: 0.78

Precision

tR: 1.0 TR: 0.63 tR: 0.89 TR: 0.69 tR: 1.0 TR: 0.63 tR: 1.0 TR: 0.63 tR: 0.94 TR: 0.65

tRS: 1.0 TR: 0.63 tRS: 0.95 TR: 0.63 tRS: 1.0 TR: 0.61 tRS: 1.0 TR: 0.72 tRS: 0.96 TR: 0.60

Recall

tR: 1.0 TR: 0.64 tR: 0.92 TR: 0.70 tR: 1.0 TR: 0.66 tR: 1.0 TR: 0.63 tR: 0.96 TR: 0.66

tRS: 1.0 TR: 0.64 tRS: 0.97 TR: 0.63 tRS: 1.0 TR: 0.62 tRS: 1.0 TR: 0.75 tRS: 0.98 TR: 0.60

F1-Score

tR: 1.0 TR: 0.64 tR: 0.90 TR: 0.70 tR: 1.0 TR: 0.64 tR: 1.0 TR: 0.63 tR: 0.95 TR: 0.66

tRS: 1.0 TR: 0.64 tRS: 0.95 TR: 0.63 tRS: 1.0 TR: 0.62 tRS: 1.0 TR: 0.72 tRS: 0.98 TR: 0.60

Abbreviations: tR: train on real data; TR: test on real data; tRS: train on real + synthetic data; RF: Random Forest.

Intraoperative Electrocorticography Signal Synthesis to Improve the Classiﬁcation of Epileptiform Tissue

1153