Epileptic Seizures Detection Using Deep Learning Techniques: A Review

Afshin Shoeibi; Marjane Khodatars; Navid Ghassemi; Mahboobeh Jafari; Parisa Moridian; Roohallah Alizadehsani; Maryam Panahiazar; Fahime Khozeimeh; Assef Zare; Hossein Hosseini-Nejad; Abbas Khosravi; Amir F. Atiya; Diba Aminshahidi; Sadiq Hussain; Modjtaba Rouhani; Saeid Nahavandi; Udyavara Rajendra Acharya

Source-linked AI summary

Epileptic Seizures Detection Using Deep Learning Techniques: A Review

Afshin Shoeibi, Marjane Khodatars, Navid Ghassemi, Mahboobeh Jafari, Parisa Moridian, Roohallah Alizadehsani, Maryam Panahiazar, Fahime Khozeimeh, Assef Zare, Hossein Hosseini-Nejad, Abbas Khosravi, Amir F. Atiya, Diba Aminshahidi, Sadiq Hussain, Modjtaba Rouhani, Saeid Nahavandi, Udyavara Rajendra Acharya

arXiv:2007.01276v3Jul 2, 2020cs.LGeess.SPstat.ML

TL;DR

Automated epileptic-seizure detection is reviewed against the labor, expertise, and feature-engineering demands of conventional approaches. The paper synthesizes deep-learning methods using EEG and MRI, along with datasets, hardware, cloud deployment, rehabilitation systems, challenges, and future work, reporting representative accuracies across reviewed models.

Problem
Manual EEG screening is time-consuming and susceptible to human error, while conventional machine learning depends on handcrafted feature and classifier selection.
Method
The paper comprehensively reviews deep-learning models for automated seizure detection across EEG and MRI, datasets, rehabilitation systems, hardware, cloud computing, and future challenges.
Results
Reviewed studies report 85.3% accuracy for the S2I-DenseNet base model, 93.92% accuracy in one autoencoder study, and 90.60% training and 86.57% test accuracies for a modified CNN-LSTM approach.
Takeaways & Limitations
The review identifies deep-learning models and practical hardware or cloud deployment as prominent directions for automated seizure-detection applications.

Abstract

from arXiv · show

A variety of screening approaches have been proposed to diagnose epileptic seizures, using electroencephalography (EEG) and magnetic resonance imaging (MRI) modalities. Artificial intelligence encompasses a variety of areas, and one of its branches is deep learning (DL). Before the rise of DL, conventional machine learning algorithms involving feature extraction were performed. This limited their performance to the ability of those handcrafting the features. However, in DL, the extraction of features and classification are entirely automated. The advent of these techniques in many areas of medicine, such as in the diagnosis of epileptic seizures, has made significant advances. In this study, a comprehensive overview of works focused on automated epileptic seizure detection using DL techniques and neuroimaging modalities is presented. Various methods proposed to diagnose epileptic seizures automatically using EEG and MRI modalities are described. In addition, rehabilitation systems developed for epileptic seizures using DL have been analyzed, and a summary is provided. The rehabilitation tools include cloud computing techniques and hardware required for implementation of DL algorithms. The important challenges in accurate detection of automated epileptic seizures using DL with EEG and MRI modalities are discussed. The advantages and limitations in employing DL-based techniques for epileptic seizures diagnosis are presented. Finally, the most promising DL models proposed and possible future works on automated epileptic seizure detection are delineated.

I. INTRODUCTION

The introduction motivates automated epileptic-seizure detection by reviewing limitations of manual and conventional machine-learning approaches, then defines the review’s coverage of deep-learning models, datasets, modalities, and future challenges.

Motivation: Epileptic seizures affect around 60 million people worldwide and can cause physical injury, cognitive disorders, and emotional distress.The paper links early detection with potential improvements in patients’ quality of life.
Motivation: EEG is widely preferred because it is economical, portable, and exposes clear frequency-domain rhythms, but manual screening is time-consuming and vulnerable to human error.The paper presents computer-based diagnosis as a possible way to alleviate these difficulties.
Motivation: Conventional machine learning depends on trial-and-error feature and classifier selection, requiring signal-processing and data-mining expertise and often working well only with limited data.The introduction motivates deep learning as data availability increases.
Scope and goals: The review examines automated seizure detection using various deep-learning models and modalities, while also providing information on available EEG datasets.Its stated goals include reviewing models across modalities and identifying available datasets.
Scope and goals: The paper analyzes the best-performing models for different data modalities and introduces future challenges in epileptic-seizure detection.The review also covers implementation-related hardware and cloud-computing developments.

II. EPILEPTIC SEIZURE DETECTION BASED ON DEEP LEARNING TECHNIQUES

The reviewed deep-learning CAD pipeline accepts neuroimaging or electrophysiological inputs, preprocesses signals, trains models, and evaluates detection performance. Dataset choice and overfitting control are recurring considerations.

Deep-learning CAD systems can accept EEG, MEG, ECoG, fNIRS, PET, SPECT, or MRI inputs before preprocessing and model evaluation.Performance is evaluated with accuracy, sensitivity, and specificity.
Popular seizure-detection datasets include Freiburg, CHB-MIT, Kaggle, Bonn, Flint-Hills, Bern-Barcelona, Hauz Khas, and Zenodo.The signals may be recorded intracranially or from the scalp of humans or animals.
The Bonn dataset is the most frequently used dataset among the reviewed research.
EEG preprocessing commonly includes noise removal, normalization, and signal preparation for deep-network deployment.Filtering, z-score normalization, and time, frequency, or time-frequency preparation are described.
Deep networks require methods to learn many parameters while avoiding overfitting, especially as network depth increases.Pooling, batch normalization, dropout, early stopping, and regularization are among the described approaches.
Convolutional networks reduce trainable parameters through filters convolved with input patterns.

1) Convolutional Neural Networks (CNNs):

The review covers 2D and 1D CNNs for seizure detection, including transformed EEG representations, direct one-dimensional processing, and several established architectures. Reported studies include both specialized designs and adapted image-classification networks.

Convolutional Neural Networks (CNNs):: CNNs are widely used for EEG seizure detection in both two-dimensional and one-dimensional architectures.Two-dimensional approaches transform one-dimensional EEG signals, whereas 1D-CNNs process signals in their original form.
Convolutional Neural Networks (CNNs):: Reviewed 2D-CNN studies include SeizNet, spectral-temporal EEG models, HFO diagnosis, AlexNet, VGG, GoogLeNet, and ResNet approaches.
Convolutional Neural Networks (CNNs):: One AlexNet-based study diagnosed focal epileptic seizures using feature extraction followed by softmax classification and achieved 100% accuracy.
Convolutional Neural Networks (CNNs):: 2D-CNN approaches use representations such as signal-as-image, spectrograms, and outputs from one- or two-layer 1D-CNNs.
Convolutional Neural Networks (CNNs):: VGG-based studies applied the architecture to facial images and sEEG signal plots, with additional 1D-CNN and LSTM components reported in one approach.
Convolutional Neural Networks (CNNs):: S2I-DenseNet achieved a best accuracy of 85.3% with an average of 70 epochs in the reviewed 2D-CNN research.
Convolutional Neural Networks (CNNs):: 1D-CNNs use fewer parameters and preserve one-dimensional signal information, avoiding possible information loss from transforming EEG into 2D representations.
Convolutional Neural Networks (CNNs):: Reviewed 1D-CNN work includes adapted LeNet, AlexNet, VGGNet, ResNet, and DenseNet models, feature extraction, EMD-CNN combinations, and temporal convolutional networks.The CHB-MIT study segmented each channel into 4-second intervals and also accepted overlapping segments.

2) Recurrent Neural Networks (RNNs):

RNNs address the sequential nature of EEG signals, while LSTM and GRU gating mechanisms help preserve relevant information across long sequences. Reviewed studies applied these architectures to seizure detection and reported accuracies up to 98%.

RNN motivation: RNNs are used for sequential signals such as EEG because simple deep-learning methods are poorly suited to variable-length data.RNNs are widely applied in biomedical signal processing.
RNN limitations: Simple RNNs suffer from short-term memory and vanishing-gradient problems when transporting information across long sequences.LSTM gates regulate information flow by preserving necessary data and discarding undesired information.
LSTM: LSTM architectures were used for feature extraction and classification, including a 3-layer model that achieved 96.82% accuracy.Other reviewed studies evaluated 3- and 4-layer LSTM architectures with softmax classification.
GRU: GRU combines the input and forget gates into an update gate, using reset and update gates to determine which information reaches the output.A 4-layer GRU system using spectrogram windows achieved 98% accuracy.

3) Autoencoders:

Autoencoders learn compressed latent representations by reconstructing their inputs and are used mainly for feature extraction or dimensionality reduction. Reviewed seizure-detection systems included standard, denoising, sparse, and stacked variants, with reported accuracies around 93% and above.

Standard autoencoders: An autoencoder reconstructs its input through an encoder, latent code, and decoder, coupling compression and decompression in one neural network.Autoencoders are commonly used for feature extraction or dimensionality reduction.
Detection pipelines: Reviewed systems combined autoencoder-based feature extraction with classifiers including genetic algorithms, SVMs, and softmax.One method used stacked autoencoders for feature extraction before SVM classification.
Autoencoder variants: Autoencoder variants include denoising, stacked, and sparse models designed to learn more robust or constrained representations.Denoising autoencoders reconstruct corrupted inputs, while stacked and sparse variants target deeper or sparse representations.
Reported results: 93.92% accuracy was achieved by the Wave2Vec approach using stacked autoencoders with softmax classification.A separate stacked sparse denoising autoencoder study obtained 93.64% accuracy.

4) Deep Belief and Boltzmann Networks:

Deep belief networks are built by stacking restricted Boltzmann machines and provide unsupervised probabilistic representations for seizure-related EEG analysis. The supplied passages also cover CNN-RNN hybrids, which combine spatial pattern extraction with temporal processing and report varied accuracies.

Deep belief and Boltzmann networks: Restricted Boltzmann machines are building blocks of deep belief networks, which are unsupervised probabilistic generative models with multiple latent layers.Convolutional DBNs extend this approach to high-dimensional data while using nearby spatial information.
DBN applications: A DBN study identified epileptic spikes in EEG using three hidden layers and achieved 96.87% accuracy.Other DBN work was reported as producing promising results, with further details summarized in Table VI.
CNN-RNN hybrids: CNN-RNN architectures combine convolutional layers for nearby spatial patterns with recurrent layers suited to temporal EEG processing.The reviewed literature describes CNN-RNN as a combination for predicting and diagnosing seizures from EEG signals.
CNN-RNN results: A ChronoNet CNN-GRU study achieved 90.60% training accuracy and 86.57% test accuracy.Other reviewed CNN-RNN systems included one-dimensional and three-dimensional CNN-GRU architectures.
CNN-RNN results: An Inception-V3 model fine-tuned with an ST-GRU CNN achieved 77.30% accuracy for seizure detection.Another approach used a multi-scale 3D-CNN whose output was passed to an RNN for temporal feature extraction and classification.

6) CNN - AEs:

CNN-AE systems combine convolutional processing with autoencoder-based representation learning for seizure detection. The reviewed medical-imaging literature largely focused on localization, making detection a secondary aim in those settings.

CNN-AE rationale: CNN-AE combinations support unsupervised feature extraction while convolutional layers identify nearby patterns and reduce parameters.This combination is presented as suitable for epileptic seizure detection.
CNN-AE method: A CNN-AE method extracted unsupervised features from input signals and supervised features from signal spectrograms before softmax classification.The approach used two deep autoencoders and a 2D-CNN for feature extraction.
Medical imaging: The reviewed imaging studies used MRI, structural MRI, functional MRI, resting-state fMRI, and PET, with or without EEG signals.These models were reported to outperform conventional models for automatic detection and monitoring.
Medical-imaging scope: In medical-imaging settings, models were mostly practiced for seizure localization because detection was not the main aim.The passage attributes this emphasis to the nature and difficulties of using imaging methods.
Imaging-related applications: CNN-based FCD detection from MRI achieved a 50% detection rate despite progress in MRI-scan analytics.Other reviewed work applied deep learning to EEG-fMRI, ECoG functional mapping, and fNIRS seizure detection.

IV. HARDWARE AND SOFTWARE USED FOR THE EPILEPTIC SEIZURE DETECTION

Deep learning has supported practical epileptic-seizure detection systems spanning software, cloud, and specialized hardware. However, model selection and real-world deployment remain constrained by dataset suitability, clinical data availability, computational demands, and only fair reported accuracies.

Hardware and software applications: Neuromorphic hardware combined with deep learning can support wearable, real-time, always-on, patient-specific seizure warning with low power consumption.The cited system is described as offering reliable long-term performance.
Deployment pathways: The review identifies graphical interfaces, cloud computing, and handheld or wearable devices as routes toward functional tools usable beyond research settings.Cloud servers may perform model computations for mobile or wearable applications and support alert messages.
Model selection: No deep learning structure is universally superior; the appropriate model depends on dataset and problem characteristics, including real-time requirements and minimum acceptable accuracy.The review states that reported detection accuracies are fair and recommends careful structure selection.
Deployment challenges: Real-world deployment is limited because many datasets use selected EEG segments, clinical datasets are often unavailable publicly, datasets are difficult to combine, and deep learning requires substantial computation.These constraints reduce suitability for real-time signals, limit usable training data, and restrict accessibility of computational resources.
Hardware and software applications: Deep learning has enabled commercial and practical epileptic-seizure detection products, including software, hardware, and cloud-supported systems.Reviewed systems include EEG-based applications, cloud components, and hardware implementations.

APPENDIX A

The appendix compiles reviewed deep learning methods for automated epileptic-seizure detection across clinical, CHB-MIT, Bonn, TUH EEG, and imaging datasets. Entries record preprocessing, model architecture, implementation tools, output functions, and reported evaluation values where available.

Appendix overview: Table X summarizes deep learning methods employed for automated epileptic-seizure detection.The appendix presents the reviewed works and their associated method details.
EEG results: Reported EEG evaluation values range from 79.34 for a TUH EEG 1D-CNN entry to 100 for several Bonn-dataset methods.The appendix lists 79.34 for the TUH EEG 1D-CNN and 100 for Bonn LSTM, DSAE, SAE, SSAE, and related entries.
EEG methods: The EEG entries include 1D-CNN, recurrent, autoencoder, belief-network, and hybrid architectures across clinical, CHB-MIT, Bonn, and TUH EEG datasets.Examples include 1D-CNN, LSTM, GRU, DBN, AE, CNN-AE, and CNN-LSTM systems.
Implementation tools: Implementation tools across the reviewed studies include Keras, TensorFlow, PyTorch, MATLAB, Theano, Lasagne, Scikit-Learn, and open-source toolkits.The appendix associates these tools with different datasets, preprocessing pipelines, and network architectures.
PET methods: The appendix also lists PET-based models, including 2D-VGGNET, 2D-Inception V3, and 3D-SVGG-C3D, with preprocessing involving OSEM, down-sampling, augmentation, and radiomics features.These entries use TensorFlow or Keras and include PET-related imaging workflows.