Advancements in Personalized Federated Learning for Epileptic

Seizure Detection

Rachitha E. and M S Bhargavi

Department of Computer Science and Engineering, Bangalore Institute of Technology, Bengaluru, India

Keywords: Personalized Federated Learning, Privacy Preservation, Epileptic Seizure Detection, EEG Signals, Deep

Learning, Neural Networks.

Abstract: Personalized federated learning for Epileptic seizure detection represents a promising avenue for improving

the accuracy and efficiency of seizure detection systems while safeguarding individual privacy. Epilepsy is a

neurological disorder characterized by recurrent seizures, and timely detection of these events is critical for

effective management and intervention. Traditional centralized approaches to seizure detection face

challenges related to data privacy, scalability, and diversity of data sources. Federated learning (FL) offers a

decentralized paradigm where models are trained cooperatively across various clients or data silos without

centralizing sensitive data. This study discusses the state-of-the-art in personalized federated learning for

epileptic seizure detection. The study focuses on the fundamentals of federated learning and its applicability

to healthcare settings, especially with regard to epilepsy management. Recent advancements in personalized

seizure detection algorithms tailored to federated learning settings, machine / deep learning models, client

/data distribution and performance are reviewed. Furthermore, challenges and opportunities in deploying

federated learning systems for epileptic seizure detection are examined. Finally, insights into the current

landscape of personalized federated learning for epileptic seizure detection are discussed with experimental

analysis inspiring further research.

1 INTRODUCTION

Epilepsy remains an eminent medical problem with

widespread implications, affecting over 65 million

individuals worldwide (World Health Organization,

2019). Despite advancements in medical and surgical

interventions, a substantial proportion of epilepsy

patients continue to experience uncontrolled seizures,

leading to significant morbidity and impaired quality

of life (Kwan & Brodie, 2000). The capacity to

anticipate epileptic seizures earlier in their beginning

holds extraordinary potential for upgrading patient

outcomes by enabling timely therapeutic

interventions and seizure prevention strategies

(Schulze-Bonhage, 2008).

Epileptic seizure detection has been a topic of

intense research, driven by the pressing need to

improve the management and well-being of

individuals with epilepsy. Over recent times,

researchers have investigated various approaches to

https://orcid.org/0000-0002-6576-6555

predict epileptic seizures, with increasing emphasis

on leveraging advanced computational techniques,

particularly deep learning, for analyzing

electroencephalography (EEG) signals (Zhang et al.,

2020). Plenty of methodologies, ranging from

conventional AI strategies, advanced deep learning

approaches to Federated Learning (FL) concepts,

have been explored in pursuit of accurate and reliable

seizure prediction models.

Federated learning performs collaborative model

training across distributed data sources, such as

multiple healthcare institutions or individual patient

devices (Kairouz et al., 2019) as portrayed in Figure

1. Privacy-preserving approaches are pivotal in

developing a personalized Federated Learning

framework for epileptic seizure detection (Qin et al.,

2021). Traditional centralized approaches to seizure

detection encounter obstacles related to information

protection, versatility, and the heterogeneity of

information sources. In response, federated learning

222

E., R. and Bhargavi, M. S.

Advancements in Personalized Federated Learning for Epileptic Seizure Detection.

DOI: 10.5220/0013311500004646

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

In Proceedings of the 1st International Conference on Cognitive & Cloud Computing (IC3Com 2024), pages 222-231

ISBN: 978-989-758-739-9

emerges as a promising solution, offering a

decentralized paradigm for cooperative model

preparation across distributed data sources while

preserving individual privacy.

Figure 1: Federated Learning framework in healthcare

setting (Tian, 2019).

Federated learning leverages the diversity of data

sources in seizure detection by utilizing local data

processing, aggregating model updates, integrating

heterogeneous data, enhancing generalization,

scaling efficiently, and allowing for personalization.

This approach results in more accurate, generalizable,

and robust seizure detection models that respect

patient privacy and encourage collaboration among

several medical institutions. EEG signals provide

valuable perception into the dynamic electrical

activity of the brain and serve as key biomarkers for

detecting epileptic seizure events. Recent

advancements in Federated learning methodologies

have spurred a fundamental change in seizure

prediction research, offering new avenues for

extracting complex patterns and temporal dynamics

from EEG data.

In this study, we synthesize recent research

findings in Federated learning methodologies applied

to epileptic seizure forecast utilizing EEG signals. By

consolidating recent advancements, this research

provides insights into the current developments,

identifies emerging trends, and proposes future

directions for advancing seizure detection in epilepsy

management. By synthesizing recent research

findings and identifying emerging trends, this study

seeks to add to the ongoing dialogue in the field of

epilepsy research and neuroinformatics. Our

objective is to elucidate the promise of Federated

learning techniques in revolutionizing seizure

prediction and their relevance for clinical application,

patient management, and healthcare resource

allocation. Ultimately, our objective is to stimulate

further research endeavors and foster collaborative

efforts aimed at translating innovative seizure

prediction algorithms into impactful clinical tools for

the benefit of epilepsy patients worldwide.

2 DATASETS FOR EPILEPTIC

SEIZURE DETECTION USING

The prosperity of personalized federated learning in

epileptic seizure detection relies heavily on the

openness of diverse and well-annotated EEG datasets.

In this section, we introduce key datasets that serve as

foundational resources for training and evaluating

personalized seizure detection models. These datasets

encompass an extent of seizure types, patient

demographics, and recording modalities, giving

priceless bits of knowledge to the headway of strong

and versatile seizure identification.

TUH EEG Seizure Corpus (TUH EEG Seizure

Corpus, 2024): The 26,846 clinical EEG recordings

that were gathered at Temple University Hospital

make up this extensive archive. The TUSZ subset

specifically focuses on seizure events, providing

valuable data for studying epilepsy, seizure detection

algorithms, and related research in neuroscience and

clinical applications. Over 150 hours of EEG data,

615 EDF files, and 247 sessions are included.

CHB-MIT Scalp EEG Database (CHB-MIT

Scalp EEG Database, 2024): This collection consists

of EEG recordings of individuals with epilepsy,

comprising both seizure and non-seizure recordings.

This database, gathered by Boston Children's

Hospital, includes EEG recordings from kids with

unmanageable epilepsy. Several days were spent

monitoring the subjects after they

discontinued

anticonvulsant medications. This was done to observe

their seizures and determine their suitability for

surgical intervention. The recordings were obtained

from 22 people and organized into 23 cases, 17

females, aged 1.5–19 years and 5 males, aged 3–22

years.

Epilepsy EEG Database (EEG-ID, 2024): EEG

recordings from epileptic patients are included along

with seizure event annotations. It provides a wide

variety of seizure kinds and patient profiles to create

customized models. The EEG database includes

recordings of 21 patients with medically intractable

focal epilepsy. At the University Hospital Freiburg in

Germany's Epilepsy Center, data were collected

during invasive preoperative epilepsy monitoring.

The epileptic focus was found in the neocortical brain

structures in eleven patients, the hippocampal region

in eight patients, and both in two patients. A Neurofile

Advancements in Personalized Federated Learning for Epileptic Seizure Detection

223

NT digital video EEG system with 128 channels, a

256 Hz sampling rate, and a 16-bit log-to-digital

converter was used to record the EEG data.

EPILEPSIAE Dataset

(EPILEPSIAE Dataset,

2024): The dataset includes EEG reports from

epileptic patients that are accompanied by seizure

start and finish timings. Annotated EEG information

from around 200 epileptic patients is available in the

EU database. 50 of these individuals have internal

brain recordings with up to 122 channels. Every

dataset offers EEG data at a maximum sampling rate

of 2500 Hz for at least 96 hours of nonstop recording.

The EEG data is supplemented with patient data and

magnetic resonance imaging data.

3 EXPLORATION OF RESEARCH

IN EPILEPTIC SEIZURE

DETECTION

Epilepsy is a neurological disorder characterized by

recurrent seizures that presents a significant challenge

in healthcare due to the critical need for timely

detection and intervention. In this section, we explore

different machine learning approaches used by

researchers for the effective detection of epileptic

seizures.

3.1 Traditional Methods

Early research on seizure prediction primarily used

traditional machine learning algorithms and

handcrafted features extracted from EEG signals

(Mirowski et al., 2009; Mormann et al., 2007; Tzallas

et al., 2009). These features included non-linear,

frequency and time domain characteristics, such as

statistical moments, entropy measures, wavelet

coefficients and spectral densities. Though these

methods were effective, they required domain-

specific knowledge. Also, these methods pose

difficulties in capturing complex, nonlinear patterns

in EEG data (Wu et al., 2024).

3.2 Evolution of Deep Learning

The progress in deep learning, particularly the

application of long short-term memory (LSTM),

recurrent neural networks (RNN), and convolutional

neural networks (CNN) has transformed seizure

prediction. These models utilize automated feature

extraction and can effectively model both spatial and

temporal relationships, leading to improved

predictions (Hosseini et al., 2020; Acharya et al.,

2018; Shoeibi et al., 2021). Further, these models can

learn both local and global patterns in EEG data

effectively, improving predictive accuracy (Kunekar

et al., 2024; Xu et al., 2020). Another important

milestone in the evolution of deep learning models is

Transfer learning, where models pre-trained on larger

datasets are optimized for seizure detection datasets

(Liu et al., 2019; Chen et al., 2022; Yu et al., 2023).

The incorporation of multi-modal data is another

approach used to enhance seizure prediction. When

EEG signals were combined with contextual

information related to EEG data, the accuracy of the

predictions improved reducing false alarms

(Moridani & Farhadi, 2017).

3.3 Personalized Federated Learning

Federated Learning allows multiple machines to

collaboratively train models by combining updates

from decentralized data sources, all while preserving

data privacy and security (Konecny et al., 2016). This

method enhances prediction models by utilizing

diverse datasets without incorporating sensitive

patient information. Model personalization involves

adapting the global models to reflect individual

patient characteristics and seizure patterns. Fine-

tuning of the models locally, personalized updates,

ensembling of model outputs and client-side

adaptation enhance the global model’s detection

accuracy while safeguarding the privacy of datasets

on the client side. Personalized federated learning

might offer the desired benefits because the patterns

of seizure vary from one person to another. There is a

difference in how individuals experience seizures due

to their onset, duration, and character. Triggers for

these seizures are not the same in patients.

Personalized federated learning customizes

models according to individual patterns as well as

responses. It facilitates the adjustment of models over

time, especially in situations such as epilepsy where

seizure patterns can change or respond differently to

treatment. By adapting models to individual patients

and updating them with data from wearables or other

sources, it is expected that the models would be able

to better tune to new situations for each patient.

Traditional learning requires that data from different

users or devices be independently and identically

distributed which is not always true for real-world

cases. Such heterogeneity of data can be addressed

through Personalized Federated Learning by adapting

models to individual client’s data distributions

considering variations thereby resulting in better

performance on diverse data.

In FL, the training task is divided among various

devices or institutions for labeling the data. This

IC3Com 2024 - International Conference on Cognitive & Cloud Computing

224

approach overcomes the energy inefficiency during

model training where traditional approaches may

consume substantial energy limiting their suitability

for long-term, continuous monitoring

. The

Personalized federated learning method will turn out

to be highly beneficial in epileptic cases where

accurate identification of seizure events is required

while preserving patient privacy. The information

about each patient is kept at the client end itself

without being stored in the central server and only the

gradients or model weights are shared to the central

server. By incorporating this measure, the

confidentiality of the patient data is maintained.

FL models can be used in wearable or

implantable devices to monitor patients in real time

in order to detect any anomalous patterns. This

allows for immediate intervention, like alerting

caregivers or administering treatments when a

seizure is detected. Federated learning enables

scalable and efficient personalized seizure detection

models across numerous patients and healthcare

institutions. This approach enables access to

advanced seizure detection technologies regardless of

the patient’s geographical location or resources

available in the healthcare system. By distributing

computational tasks, using local devices for training,

reducing data transmission needs, and aggregating

updates only centrally, federated learning enhances

the scalability and computational efficiency of

seizure detection systems. This decentralized method

aids in handling large volumes of data and scales

dynamically. Model will be improved continuously in

a centralized manner but resource consumption will

be low.

In federated learning, challenges such as

communication overhead and model convergence are

tackled in many ways. Compression of model updates

at the central server and aggregation is done

periodically to handle communication overhead.

Model convergence can be tackled using learning

rates in model tuning adaptively and using advanced

optimization algorithms (Kairouz et al., 2021;

McMahan et al., 2017). These methods reduce

communication costs and handle diverse data across

devices, ensuring efficient seizure detection model

training. By addressing these issues, federated

learning can enhance computational efficiency, and

maintain reliable and accurate seizure detection

capabilities.

In summary, personalized federated learning

holds an incredible commitment to improving

epileptic seizure detection by tailoring models to

individual patients, preserving privacy, adapting to

changes over time, reducing annotation burdens,

enabling real-time monitoring, and enhancing

scalability and accessibility across healthcare

systems. Recent research has explored the utilization

of federated learning techniques for epileptic seizure

detection. In the area of expertise in Personalized

Federated Learning for Epileptic Seizure Detection,

the research landscape may be characterized by a

scarcity of academic papers available for review.

Personalized federated learning for epileptic seizure

detection is a relatively new and emerging field

within both the medical and machine learning

communities. Consequently, there are very few

published works addressing this intersection.

Saleh Baghersalimi et al. (Baghersalimi et al.,

2021) introduced a FL approach for epileptic seizure

detection on mobile platforms, using NVIDIA Jetson

Nano units. They train neural networks on

preprocessed ECG segments, showcasing FL’s

performance over centralized training, with

personalized FL offering further improvements. The

trade-off between model detection accuracy and

training efficiency is discussed, highlighting the

benefits of FL in preserving data privacy while

achieving comparable performance to centralized

training. The authors also address challenges related

to energy consumption on mobile platforms by

carefully designing the synchronization process.

They optimize synchronization frequency to balance

accuracy and energy consumption, achieving

promising results for practical deployment. Overall,

their work demonstrates the viability of deploying

FL-based seizure detection systems on resource-

constrained devices to attain better capabilities.

Raghdah Saemaldahr et al. (Saemaldahr and Ilyas,

2023) suggested a multi-tier architecture for epileptic

seizure detection. By leveraging a large number of

seizure patterns from patients spread throughout the

globe, this design protects patient privacy while

sharing data. The two-level edge layer model affects

the preictal state determination process on both a

local and global level through model-assisted

decision-making. As a learned local model, the

Spiking Encoder is connected to the Graph CNN

using a time-series analysis with dual granularity

method. Every local model calculates the preictal

likelihood in the coarse-grained personalization by

utilizing the combined seizure understandings

gathered from the various medical institutions via FL.

In fine-grained personalization, the Adaptive Neuro

Fuzzy Inference System is used to identify

individuals with epileptic seizures by analyzing the

Advancements in Personalized Federated Learning for Epileptic Seizure Detection

225

aftereffects of the FL model, heart rate variability

data, and patient-specific clinical features.

S. Baghersalimi et al. (Baghersalimi et al., 2023)

suggest an FL approach that ensures patient privacy

and complies with healthcare standards by having

different hospitals work together on model training

without directly exchanging patient data. In the

experimental setting, patient data is partitioned by age

range and kind of seizure among hospitals. Geometric

Mean, Specificity, and Sensitivity are examples of

evaluation metrics. Four hospitals make up the

scenario for the decentralized training setup. Each

hospital trains locally, shares model updates, and

assesses ensemble models for seizure detection. The

learning parameters entail training DNNs by using the

Adam optimizer with TensorFlow on pre-processed

3-second ECG and EEG segments. Utilizing the

Raspberry Pi and Kendryte K210 platforms, energy

consumption analysis is carried out. The suggested

ensemble learning approach significantly improves

seizure detection correctness, especially in large-

scale FL.

W. Ding et al. (Ding et al., 2023) present a

Federated Edge Server-based Epileptic Seizure

Detection (Fed-ESD) system tailored for the Internet

of Medical Things (IoMT) environment. Through

comprehensive experimentation and analysis, the

authors showcase the system's robust performance in

detecting epileptic seizures. They compare Fed-ESD

with advanced approaches, demonstrating its

superiority in correctness and efficiency.

Additionally, scalability experiments highlight Fed-

ESD's ability to maintain detection performance with

a rising number of edge nodes. Privacy,

communication latency, and energy consumption

analyses underscore the system's practical viability,

especially in battery-operated IoMT devices. They

conclude by proposing future directions, such as

improving interpretability and addressing data

heterogeneity, to further enhance Fed-ESD's

applicability in real-world IoMT scenarios. Overall,

the study presents Fed-ESD as a promising tool for

distributed epileptic seizure prediction in medical

IoMT applications.

Marcos Lupion, et al. (Lupion et al., 2023) offer a

novel method for identifying epileptic seizures that

utilizes federated machine learning algorithms and

inexpensive IoT devices. It starts by outlining the

shortcomings of the available detection techniques

and then suggests a system that gets around those

drawbacks, like high cost and short battery life.

Wearable technology is used in the system

architecture to send data to a central device, where a

federated machine learning algorithm is used to detect

seizures while protecting user privacy. According to

preliminary results, compared to conventional

methods, detection rates and efficiency gains are

promising. The research provides a thorough

approach to continuous seizure monitoring that may

prove advantageous in terms of cost, effectiveness,

and privacy protection. This makes it ready for future

study and practical implementation.

S. Vasanthadev S et al. (Suryakala et al., 2024)

presents an innovative strategy to overcome the

difficulties in distinguishing epileptic seizures. It

begins by outlining the privacy and generalization

concerns that limit centralized machine learning's

ability to handle EEG data. Privacy concerns are

addressed by the decentralized strategy, which

permits model training utilizing local datasets without

sharing raw EEG recordings. Background

information on epilepsy, EEG technology, and the

significance of precise seizure detection in healthcare

are discussed. Time complexity analysis has been

performed to assess how FL can be applicable in real-

world scenarios. Comparisons with previous research

have been carried out which demonstrates improved

sensitivity, specificity, and accuracy. Overall, the

paper presents a well-structured study with valuable

insights into epileptic seizure detection using FL,

offering significant implications for healthcare data

analysis and patient care.

Amin Aminifar et al. (Aminifar et al., 2024)

present a framework for privacy-preserving federated

learning designed for wearable technology over IoT

infrastructure and mobile health under resource

constraints. It addresses the challenges of

decentralized healthcare data, emphasizing real-time

epileptic seizure detection. The framework integrates

high-quality hospital data with distributed deep

learning to reduce computing and communication

overheads while maintaining prediction accuracy. It

employs federated learning and Secure Multiparty

Computation (SMC) to ensure data privacy, evaluated

through implementation on Amazon's AWS cloud

platform. It does, however, recognize certain

drawbacks, such as the assumption of homogeneous

device resources, and suggests further research into

how to support heterogeneous devices and

computational power in FL scenarios.

These researches collectively contribute to

advancing the field of personalized federated learning

for epileptic seizure detection, addressing challenges

related to privacy preservation, model personalization

and model robustness. An overview of recent

research in FL for epileptic seizure detection is

provided in Table 1.

IC3Com 2024 - International Conference on Cognitive & Cloud Computing

226

4 CHALLENGES AND FUTURE

DIRECTIONS

Challenges in personalized federated learning for

epileptic seizure detection include the scarcity of

annotated EEG datasets, limiting the availability of

labeled data necessary for training accurate seizure

detection models. Moreover, existing models struggle

to generalize over diverse patient populations due to

variations in seizure patterns. Data heterogeneity is a

significant challenge as seizure patterns and nature

diverge significantly among patients, as well as

within the same individual

over time. This

heterogeneous nature of seizure patterns complicates

model development, requiring adaptable and robust

architectures to incorporate these variations. Ensuring

data privacy and security is vital in FL but yet

challenging. FL must safeguard sensitive medical

data during transmission and aggregation. It may be

vulnerable to attacks such as model and data

poisoning. The communication overhead is another

challenge as there is a requirement for frequent model

updates which can strain network resources. Another

issue is the interpretability of deep learning models.

The complexity of models and their uninterpretable

decisions hinder their application in clinical practice

as understanding, interpretation and trust of the

decisions are very important in clinical practices.

Table 1: Overview of recent research in Federated Learning for epileptic seizure detection.

Authors (Year) Dataset Machine / Deep

Learning Models

Data / Client

Distribution

Performance Metrics

S. Baghersalimi

et al., 2021

EPILEPSIAE

dataset - one-lead

ECG and 19-

channel EEG data

of 30 patients

MLP,

1 Dimensional

Convolutional Neural

Network,

Residual 1-DCNN

29 patient data across

4 clients

Sensitivity: 90.24%,

Specificity: 91.58%

Raghdah

Saemaldahr et

al., 2023

CHB-MIT EEG

dataset, Bonn

EEG dataset,

NSC New Delhi

EEG dataset

Spiking Encoder,

Graph Convolutional

Neural Network,

Adaptive Neuro-Fuzzy

Inference System

(ANFIS)

Not specified Sensitivity: 96.33%,

Specificity: 96.14% for

CHB-MIT dataset

S. Baghersalimi

et al., 2023

EPILEPSIAE

dataset,

TUSZ dataset

Ensemble learning

(ECG+EEG) using

Deep neural networks

4 hospital clients and

non-IID client data

distributions

EPILEPSIAE Dataset

Avg Sensitivity: 89%,

Avg Specificity: 88.4%

TUSZ Dataset

Avg Sensitivity: 87.9%,

Avg Specificity: 83.8%

W. Ding et al.,

2023

EPILEPSIAE

dataset

Lightweight

spatiotemporal

transformer network

29 edge nodes Sensitivity: 76.8%,

Specificity: 81.74%,

Accuracy: 79.2%

Marcos

Lupi´on, et al.,

2023

Simulated

dataset

CNN with LSTM

layers

Wearable IoT

devices data of 4

users

Train time: 278s,

Recall: 0.88,

Precision: 0.88,

F-Score: 0.88

S. Vasanthadev et

al., 2024

UCI Machine

Learning

Repository

Neural Network,

Decision Tree,

Logistic Regression

5 client nodes Accuracy:

Neural Network: 99%,

Decision Tree: 94%,

Logistic Regression: 89%

Amin Aminifar et

al., 2024

CHB-MIT

dataset

Deep Neural

Network (DNN)

Data Distributed

among multiple

hospitals or patients'

mobile devices

/sensors

Accuracy

DNN: 88.4%

Advancements in Personalized Federated Learning for Epileptic Seizure Detection

227

Looking ahead, addressing these challenges requires

collaborative initiatives to overcome data scarcity

issues. Partnerships between research institutions,

healthcare providers, and patient advocacy groups

can facilitate data sharing and resource pooling for

model advancement and approval. Future research

directions must aim at addressing current limitations

and advancing the field of epileptic seizure

prediction. Novel methodologies of federated

learning approaches must be proposed to enhance

model scalability, privacy, and collaboration across

multiple healthcare institutions. Opportunities in FL

must be identified for integrating multimodal data

sources, such as EEG, clinical metadata, and genetic

information, to enhance the correctness and reliability

of seizure prediction models. Future research

endeavors should prioritize the development of

interpretable and explainable deep machine learning

models tailored for seizure detection. Improving

clinicians' understanding and trust in model

predictions will be crucial for their adoption in

clinical practice. Additionally, advancements in

wearable EEG technology offer promise for

translating seizure prediction algorithms into

practical clinical tools. Real-time monitoring systems

can enable prompt action and customized care,

eventually leading to better patient outcomes in

epilepsy management.

5 EXPERIMENTAL ANALYSIS

AND RESULTS

This section discusses the efficacy of Federated

learning in epileptic seizure detection through

experimental analysis. In this research, the dataset

was gathered from the UCI Machine Learning

Repository (Data.world

, last accessed 2023/05/21

). Five

distinct folders, each holding 100 files make up the

original dataset. Each file in the folder represents a

single individual. Every file contains a 23.6-second

recording of brain activity. It includes 500 people's

EEG recordings, with 4097 data points collected

across 23.6 seconds from each of them. These data

points are divided into 23 chunks per individual, each

chunk containing 178 data points for 1 second. This

results in 11,500 rows of data, with 178 columns of

EEG data points and the 179th column representing

the label. The explanatory variables X1 to X178 form

the 178-dimensional input vector. Table 2 depicts the

experimental setup used for analysis.

Table 2: Experimental Setup.

Aspect Details

Model CNN-LSTM

Input 1D sequential EEG data

Architecture

Conv1D: 32 filters, kernel size 3,

ReLU activation

MaxPooling: kernel size 2

LSTM: 64 hidden units

FC: 32 units, ReLU activation

FC Output: sigmoid activation

Personalization

Personalized Federated Learning

with 10 users

Local Training

Random sampling of 10% data for

each user Adam optimizer, learning

rate = 0.001, 30 epochs

Loss Function

Binary Cross-Entropy Loss

(BCELoss)

Validation

Global model evaluated on

validation set after aggregation

Figure 2: Centralized Model for Personalized Epileptic Seizure Detection.

IC3Com 2024 - International Conference on Cognitive & Cloud Computing

228

Figure 3: Federated Learning Model for Personalized Epileptic Seizure Detection.

The Convolutional neural networks (CNNs) and long

short-term memory (LSTM) networks are combined

in the CNN-LSTM architecture. It employs a 1D

convolutional layer for feature extraction followed by

max-pooling to reduce spatial dimensions. The

LSTM layer captures temporal dependencies, while

fully connected layers process the extracted features

for prediction. This architecture is well-suited for

tasks involving 1D sequential data, such as EEG

recordings, as it effectively captures both spatial and

temporal patterns.

Figure 2 shows the training loss and training

accuracy of centralized model learning whereas

Figure 3 shows the training loss and training accuracy

of the FL model. The x-axis represents the number of

epochs, which are iterations over the training data.

The y-axis in the training loss graph depicts the loss,

and the y-axis in the training accuracy graph depicts

the accuracy.

Table 3: Comparison of Federated Learning and

Centralized Learning Models.

Model Accuracy Achieved

Federated Learning Model

for Personalized Epileptic

Seizure Detection

Global Model: 93.09%

Centralized Model for

Personalized Epileptic

Seizure Detection

98.09%

Table 3 portrays the accuracy scores of the

Federated learning model and the Centralized

machine learning model for seizure detection. As

evident from Table 3, the decentralized approaches in

federated learning may entail a reduction in accuracy

compared to centralized methods due to the

heterogeneous nature of data samples. However, they

offer valuable privacy and scalability benefits that are

paramount in privacy-sensitive applications. Striking

a balance between accuracy and privacy remains a

key challenge in the design and implementation of

federated learning systems.

6 CONCLUSION

In conclusion, this study provides a complete outline

of recent advancements in epileptic seizure forecast

utilizing federated learning techniques applied to

EEG signals. Through a thorough examination of

relevant literature and methodologies, key insights

have been gleaned regarding the efficacy, challenges,

and future prospects of seizure prediction in a

federated learning setup. Developments in

personalized federated learning for the detection of

epileptic seizures present a viable way to address

patient variability and the requirement to protect

patient privacy in medical records. Models can be

trained on decentralized patient data by utilizing

federated learning techniques, which leads to

improved performance customized to individual

seizure patterns. This strategy encourages scalability

and generalizability across a range of patient

populations while protecting patient privacy by

permitting training without disclosing raw data.

Additionally, customized seizure detection models

facilitate the creation of individualized treatment.

Research collaborations in this field promote

information exchange and group advancement

toward better seizure detection technologies and

patient treatment.

Advancements in Personalized Federated Learning for Epileptic Seizure Detection

229

REFERENCES

World Health Organization. Epilepsy Fact Sheet (2019.

Retrieved from https://www.who.int/news-room/fact-

sheets/detail/epilepsy,

Kwan, P., & Brodie, M. J. (2000). Early Identification of

Refractory Epilepsy. New England Journal of

Medicine, 342(5), 314–319,

https://doi.org/10.1056/NEJM200002033420503

Schulze-Bonhage, A. (2008). Challenges in epilepsy:

Identifying and characterizing treatment-resistant

patients. The Neurologist, 14(2), 35–44,

https://doi.org/10.1097/NRL.0b013e3181668884

Zhang, Y., Qin, Y., Zhou, M., & Wang, H. (2020). Deep

learning for epileptic seizure prediction: A

comprehensive review. Frontiers in Neuroscience, 14,

576124, https://doi.org/10.3389/fnins.2020.576124

Kairouz, P., McMahan, H. B., Avent, B., Bellet, A., Bennis,

M., Bhagoji, A. N., Bonawitz, K., Charles, Z.,

Cormode, G., Cummings, R., D'Oliveira, R.,

Rouayheb, S. E., Evans, D., Gardner, J. R., Garrett, Z.,

Gascoigne, J., Ghazi, B., Gibbons, P. B., Gruteser, M.,

Zhang, Z. 2019. Advances and Open Problems in

Federated Learning. arXiv preprint arXiv:1912.04977.

https://arxiv.org/abs/1912.04977

Qin, Y., Wang, H., Zhao, Y., Zhou, M., & Li, J. (2021).

Privacy-Preserving Federated Learning for Epileptic

Seizure Detection: Challenges and Opportunities.

Frontiers in Neuroscience, 15, 641739,

https://doi.org/10.3389/fnins.2021.641739

Tian, L. (2019). Federated Learning: Challenges, Methods

and Future Directions. Available online:

https://blog.ml.cmu.edu/2019/11/12/federated-

learning-challenges-methods-and-future-directions/

TUH EEG Seizure

Corpus,https://isip.piconepress.com/projects/tuh_eeg/

html/downloads.shtml, last accessed 2024/4/18.

CHB-MIT Scalp EEG

Database,https://physionet.org/pn6/chbmit/, last

accessed 2024/4/20.

Epilepsy EEG Database (EEG-ID),

https://www.epilepsy.uni-freiburg.de/freiburg-

seizure-prediction-project/eeg-database, last accessed

2024/4/24

EPILEPSIAE Dataset

,https://pubmed.ncbi.nlm.nih.gov/20863589/, last

accessed 2024/4/26

P. Mirowski, D. Madhavan, Y. LeCun, and R. Kuzniecky.

(2009). Classification of Patterns of EEG

Synchronization for Seizure Prediction. Clinical

Neurophysiology, vol. 120, no. 11, pp. 1927-1940.

F. Mormann, R. G. Andrzejak, C. E. Elger, and K. Lehnertz.

(2007). Seizure Prediction: The Long and Winding

Road. Brain, vol. 130, no. 2, pp. 314-333.

A. T. Tzallas, M. G. Tsipouras, and D. I. Fotiadis. (2019).

Epileptic Seizure Detection in EEGs Using Time-

Frequency Analysis. IEEE Transactions on

Information Technology in Biomedicine, vol. 13, no.

5, pp. 703-710.

G. Wu, K. Yu, H. Zhou, X. Wu, and S. Su. (2024). Time-

Series Anomaly Detection Based on Dynamic

Temporal Graph Convolutional Network for Epilepsy

Diagnosis, Bioengineering, vol. 11, no. 1, p. 53, doi:

10.3390/bioengineering11010053.

M. P. Hosseini, D. Pompili, and K. V. Elisevich. (2020).

Optimized Deep Learning for EEG Big Data and

Seizure Prediction BCI via Internet of Things. IEEE

Trans. Big Data, vol. 6, no. 2, pp. 250-260.

U. R. Acharya, S. L. Oh, Y. Hagiwara, J. H. Tan, and H.

Adeli. (2018). Deep Convolutional Neural Network

for the Automated Detection and Diagnosis of Seizure

Using EEG Signals. Comput. Biol. Med., vol. 100, pp.

270-278.

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

Moridian, R. Alizadehsani, M. Panahiazar, F.

Khozeimeh, A. Zare, H. Hosseini-Nejad, A. Khosravi,

A. F. Atiya, D. Aminshahidi, S. Hussain, M. Rouhani,

S. Nahavandi, and U. R. Acharya. (2021). Epileptic

Seizures Detection Using Deep Learning Techniques:

A Review. Int. J. Environ. Res. Public Health, vol. 18,

no. 11, p. 5780.

P. Kunekar, M. K. Gupta, and P. Gaur. (2024). Detection of

epileptic seizure in EEG signals using machine

learning and deep learning techniques. Journal of

Engineering and Applied Science, vol. 71, no. 21.

Y. Xu, J. Yang, S. Zhao, H. Wu, and M. Sawan. (2020). An

End-to-End Deep Learning Approach for Epileptic

Seizure Prediction. in 2020 2nd IEEE International

Conference on Artificial Intelligence Circuits and

Systems (AICAS), pp. 31-35. doi:

10.1109/AICAS48895.2020.9079870.

H. Liu, M. Li, H. Zhang, W. Li, and C. Shen. (2019).

Seizure prediction using multimodal data fusion and

novel deep learning model," Biomed. Signal Process.

Control, vol. 52, pp. 204-211.

Chen, Fangyi, Ina Chen, Muhammad Zafar, Saurabh R.

Sinha, and Xiao Hu. (2022). Seizures detection using

multimodal signals: a scoping review: Physiological

Measurement, vol. 43, no. 7, Art. no. 07TR01, doi:

10.1088/1361-6579/ac7a8d.

Yu, Shuang, Rima El Atrache, Jianbin Tang, Michele

Jackson, Adam Makarucha, Sarah Cantley, Theodore

Sheehan. (2023). Artificial intelligence-enhanced

epileptic seizure detection by wearables," Epilepsia,

vol. 64, no. 5, pp. 1009-1018. doi: 10.1111/epi.17774.

Moridani, M., and H. Farhadi. (2017). Heart rate variability

as a biomarker for epilepsy seizure prediction,"

Bratislava Medical Journal, vol. 118, no. 01, pp. 3-8,

doi: 10.4149/BLL_2017_001.

J. Konečnỳ, H. B. McMahan, F. X. Yu, P. Richtárik, A. T.

Suresh, and D. Bacon. (2016). Federated Learning:

Strategies for Improving Communication Efficiency.

arXiv preprint arXiv:1610.05492.

IC3Com 2024 - International Conference on Cognitive & Cloud Computing

230

Kairouz, P., McMahan, H. B., Avent, B., Bellet, A., Bennis,

M., Bhagoji, A. N., Bonawitz, K., Charles, Z.,

Cormode, G., Cummings, R., d'Oliveira, R. G. L., El

Rouayheb, S., Evans, D., Gardner, J., Garrett, Z.,

Gascón, A., Ghazi, B., Hardy, S., Hsu, J., Huo, Z.,

Jaggi, M., Javidi, T., Joshi, G., Khodak, M., Konecny,

J., Korolova, A., Koushanfar, F., Koyejo, S., Lepoint,

T., Liu, Y., Mittal, P., Mohri, M., Nock, R., Özgür, A.,

Pagh, R., Qi, H., Ramage, D., Raskar, R., Song, S.,

Song, W., Stich, S. U., Sun, Z., Suresh, A. T., Tramèr,

F., Vepakomma, P., Wang, J., Xie, L., Xu, M., Yang,

Q., Yu, F. X., Yu, H., & Zhao, S. (2021). Advances

and Open Problems in Federated Learning.

Foundations and Trends in Machine Learning, vol. 14,

no. 1–2, pp. 1-210.

McMahan, B., Moore, E., Ramage, D., Hampson, S., and y

Arcas, B. A. (2017). Communication-Efficient

Learning of Deep Networks from Decentralized Data,"

in Proc. of the 20th International Conference on

Artificial Intelligence and Statistics (AISTATS), Fort

Lauderdale, FL, USA, pp. 1273-1282.

Baghersalimi, S., Teijeiro, T., Atienza, D., & Aminifar, A.

(2021). Personalized Real-Time Federated Learning

for Epileptic Seizure Detection. IEEE Journal of

Biomedical and Health Informatics, DOI:

10.1109/JBHI.2021.3096127

Saemaldahr, R., & Ilyas, M. Patient-Specific Preictal

Pattern-Aware Epileptic Seizure Prediction with

Federated Learning. (2023). Sensors, 23(14), 6578,

https://doi.org/10.3390/s23146578

Baghersalimi, S., Teijeiro, T., Aminifar, A., & Atienza, D.

(2023). Decentralized Federated Learning for

Epileptic Seizures Detection in Low-Power Wearable

Systems. IEEE Transactions on Mobile Computing,

PP(99), 1-16, DOI: 10.1109/TMC.2023.3320862

W. Ding, C. Liu, M. Abdel-Basset, H. Hawash, and S.

Abdel-Razek. (2023). Fed-ESD: Federated learning

for efficient epileptic seizure detection in the fog-

assisted internet of medical things. Information

Sciences. DOI: 10.1016/j.ins.2023.02.052

M. Lupión, J. F. Sanjuan, J. Medina-Quero, and P. Martínez

Ortigosa. (2023). Epilepsy Seizure Detection Using

Low-Cost IoT Devices and a Federated Machine

Learning Algorithm. in ISAmI 2022, V. Julián et al.

(Eds.), LNNS 603, pp. 229–238, Springer Nature

Switzerland AG. DOI: 10.1007/978-3-031-22356-

3_22

Suryakala, S. V., Vidya, T. R. S., & Ramakrishnan, S. H.

(2024). Federated Machine Learning for Epileptic

Seizure Detection using EEG. International Journal of

Advanced Computer Science and Applications, 15(4),

1048, DOI: 10.14569/IJACSA.2024.01504106

A. Aminifar, M. Shokri, and A. Aminifar. (2024). Privacy-

Preserving Edge Federated Learning for Intelligent

Mobile-Health Systems. arXiv:2405.05611 [cs.LG],

Available: https://doi.org/10.48550/arXiv.2405.05611

Data.world. (last accessed 2023/05/21)

https://data.world/uci/epileptic-seizure-recognition,

Advancements in Personalized Federated Learning for Epileptic Seizure Detection

231