Epilepsy is one of the critical neurological disorders that influence the human brain by causing epileptic seizures. An epileptic seizure is an uncontrollable electrical interruption in the brain. It is undeniable to recognize the epileptic seizure to provide appropriate diagnosis and treatment for epileptic patients. Epilepsy detection is utilized to detect the abnormalities in the brain with the help of an electroencephalogram (EEG). In recent times deep learning technologies are highly utilized EEG for epilepsy detection.
Importance of Deep Learning in Epilepsy Detection: Some of the significant points to developing deep learning-based epilepsy detection are below:
• Deep learning models possess significant advances over conventional machine learning models for diagnosing epileptic seizures.
• Deep learning models potentially facilitate automatic epileptic seizure detection and prediction due to their multiple layers of neural networks with automated feature extraction and huge data handling capability.
• Deep learning models are trained to perform binary and multi-class classification of epileptic segments from EEG and intracranial EEG (iEEG) signals to detect epilepsy.
• Processes involved in epilepsy detection using deep learning models are EEG signal acquisition, pre-processing, automatic feature extraction, and classification and performance analysis.
Powerful Deep Learning Techniques for Epilepsy Detection: Most commonly applied deep learning algorithms for the epileptic seizure detection to support clinical decision making in epilepsy diagnosis are highlighted here:
Convolutional Neural Networks (CNNs):
• CNN is one of the popular deep learning models and is adopted to diagnose diseases using biological signals.
• Currently, both two-dimensional (2D) and one-dimensional (1D) CNN are employed in the epileptic seizure detection research area.
• 2D-CNN effectively learns the structure of seizures by extracting both temporal and spectral characteristics of EEG signals to detect epilepsy.
• 1D-CNN is inherently desirable to process EEG signals to recognize seizure patterns with its more straightforward structure.
• Other CNN-based architectures applied to diagnose epileptic seizures are the Alexnet network, Visual Geometry Group (VGG) model, GoogleNet, and ResNet networks.
Recurrent Neural Networks (RNNs):
• RNN possess the ability to handle sequential characteristic of EEG signals to discover appropriate seizure patterns.
• Long Short-Term Memory (LSTM) is a variant of RNN, which is majorly applied to detect epilepsy to deal with time series and long-range dependencies based on EEG data.
• Gated Recurrent Units (GRU) attain remarkable outcomes for epileptic seizure detection using five to three-layered GRU enabled with a softmax classifier.
Deep Belief Networks (DBNs): DBN networks are exploited to identify the epileptic spikes in EEG data for epilepsy diagnosis with a promising outcome.
Auto Encoders (AEs): AE networks are efficaciously applied for automated epilepsy diagnosis and many extracted features and dimensionality reduction methods.
Hybrid deep learning models: Some recently investigated efficient hybrid deep learning architectures for epileptic seizure detection are CNN-RNNs and CNN-AEs.
Other Physiological Signals for Deep Learning-based Epilepsy Detection: Epileptic seizure detection via computer-aided-diagnosis (CAD) employs deep learning models because of its capability to process the following diagnosing signals includes:
• Magneto-encephalography (MEG) – Recently, advanced deep learning models use MEG with high-frequency oscillation signals to detect lesions and interictal spikes of epileptic patients.
• Functional Near-Infra-Red spectroscopy (fNIRS) – CNN is a suitable deep learning model to recognize epileptic seizures with the help of fNIRS is regarded as a new concept.
• Positron Emission Tomography (PET) – PET images are presently utilized for identifying epileptic seizures under the deep learning strategy with improved accuracy.
• Single-Photon Emission Computed Tomography (SPECT) – Recent study shows that SPECT images effectively localize the epileptic zone for epilepsy treatment.
• Magnetic Resonance Imaging (MRI) – MRI scans is another effective modality to categorize epileptic seizures and offer better outcome while enabled by the deep learning model.
Challenges and Future Scopes for Deep Learning-based Epilepsy Detection: There are various challenges in diagnosing epileptic seizures using deep learning technology, and some of them are listed below:
• The foremost challenge is unattainable datasets with high recording time.
• The partially available datasets for identifying epileptic seizures are difficult due to the finite recording duration.
• The real-time diagnosis of epileptic seizures enabled by long-duration EEG datasets needs to be investigated in the future.
• Optimal performance for the epileptic seizure detection model has not been attained using deep neural networks owing to the lack of reachable datasets.
• In order to achieve consistent detection accuracy, combining available EEG datasets will be utilized for epileptic seizure detection.
• Incorporating semi-supervised and unsupervised methods in epileptic seizure detection will be enabled to conquer EEG dataset size limits.
• Further improvement in the automated epilepsy diagnosis system will be obtained by combining different modalities with EEG modality to detect epileptic seizures more precisely.
• On the whole, deep learning models proficiently work on epilepsy detection, owns the advantage of no need for healthcare clinical experts to design the feature extraction and assist the clinical systems to provide an accurate diagnosis of epilepsy.