Prediction of Epileptic Seizures Using Deep Learning: A Brief Review of Current Methods and Emerging Trends

Today, epilepsy is treated with methods such as medication and surgical procedures that remove the part of the brain causing seizures. However, surgery is only an option for a small number of patients. Approximately 30%-40% of epilepsy patients cannot be treated with medication because they have a drug-resistant form of epilepsy [1]. The global prevalence of epilepsy is estimated to be around 1%, and a significant proportion of these patients are classified as untreatable. The prediction of uncontrolled epileptic seizures has been a focus of research for many years due to the potential risks associated with such seizures. EEG data is commonly used for predicting epileptic seizures, with studies in this area dating back to the early 1970s [2]. Although signal analysis methods were frequently used in the early stages of research, faster and more efficient results have been achieved over time with advanced machine learning techniques. In recent years, deep learning methods have gained significant popularity and have been extensively utilized due to their ability to efficiently process vast amounts of data, including EEG data, which is crucial for the accurate prediction of epileptic seizures.

Table 1. EEG datasets

This study aims to provide a detailed description of epileptic seizure prediction methods and to compare the techniques used at each step of the process. Currently, there is no single study that comprehensively combines and describes all these methods in detail. By reviewing both past and present studies, this research offers an overview of the evolution of seizure prediction. Table 1 presented highlight various rates and criteria, with the goal that the advantages and disadvantages of the methods will serve as a guide for future research.

When examining studies on epilepsy, it is evident that most research focuses on diagnosing epilepsy to provide the necessary treatment. In contrast, there are fewer studies on seizure prediction, which aims to alert the patient a certain time before an upcoming seizure. The expectation for a seizure prediction study is that it functions in a way that minimizes interruption of the patient’s life, provides sufficient time for medical intervention, and has the lowest possible number of false predictions. One of the most critical aspects of seizure prediction is the selection of the appropriate approach. Two different approaches are used to predict epileptic seizures. The first approach, known as preictal-interictal classification, involves labeling a specific periods preictal and interictal. A specific period before the onset of a seizure, such as 15, 30, 60, or 120 minutes, is considered preictal, while the period before the preictal phase and after the end of the seizure is considered interictal. Typically, a moving window is used to temporarily characterize these interictal and preictal stages, where a linear or nonlinear measure is calculated and ranges from 5 to 50 seconds. These windows may overlap to a certain extent. Classifier is then applied to distinguish between these phases. The second approach, which is rarely studied, involves the threshold methodology. In this method, increasing and decreasing values are detected during the preictal period, and an alarm is activated when a certain threshold is reached. An example of the second approach is Usman and Hassan’s research [4]. He worked exclusively with seizure data and determined threshold values based on various statistical calculations, such as variance, entropy, skewness, complexity, and kurtosis. Using these values, he performed classification by applying machine learning algorithms. However, this method is not preferred because preictal signals are often unclear and may overlap with artifacts. Another important consideration in predicting epileptic seizures is whether the study is patient-specific or patient-independent. Although patient-specific studies often achieve high success rates, the limited number of samples can pose challenges. In such studies, each patient is classified according to their own data. Patient-independent studies, on the other hand, are more applicable to real-world scenarios but have a lower success rate and a more complex classification structure, as EEG data vary significantly from patient to patient. In these studies, data from all patients are used for training and testing is conducted on any individual patient. Upon reviewing the studies for the prediction of epilepsy attacks, it was observed that the process shown in Figure 1 was generally applied.

1.png

Figure 1. A common process in the prediction of epileptic seizures

In reviewing the literature, it would be logical to examine epileptic seizure prediction studies in two distinct periods:

(1) The early period and traditional machine learning: Until the 1990s, linear methods were frequently used, especially in the feature extraction phase. While these approaches offer computational advantages, they have low sensitivity to artifacts. Since EEG signals are multidimensional and chaotic, these methods have yielded limited success. From the 1990s onward, methods based on nonlinear dynamics theory started to be employed. Studies conducted in the early 1990s initially focused only on the preictal period; however, Lehnertz and Elger [5] compared the preictal and interictal periods using the correlation dimension method, a nonlinear approach for time series analysis. This method enables the prediction of epileptic seizures minutes before they occur by detecting signal deterioration during the preictal stage.

Given the uncertainty and irregularity of EEG signals, nonlinear methods, which are chaotic and dynamic, offer higher accuracy in predicting seizures and are relatively preferred despite requiring more computational costs than linear methods [6]. As chaotic methods, the Large Lyapunov Exponent [7] and Correlation Intensity [8], which are time-domain analyses providing information about the brain's dynamic stability, have been used. Iasemidis et al. [7] analyzed the chaotic state before the onset of an epileptic seizure using the moving window method and reported that it decreased. In their later work, Iasemidis et al. [9] developed a real-time statistical algorithm that continuously calculates the maximum Lyapunov exponent and monitors the T-index curves of this calculation. The measurement calculated by the algorithm generated a warning when it exceeded a specified threshold. The study, which predicted seizures 45.3 minutes in advance, achieved an average sensitivity of 81% and a specificity of 78%. However, these methods cannot provide fast analysis due to their computational complexity on a long time scale.

Since the EEG signal is not static and linear, EEG signal analysis of some mathematical models in the time domain is difficult. To overcome this difficulty, some studies have used frequency domain methods. Frequency Domain, also known as Spectral Analysis, which is more sensitive to the presence of artifacts than Time Domain statistical methods, is another method used in the feature extraction phase in the traditional method. The most commonly used frequency domain property is the Power Spectral Density (PSD). Bandarabadi et al. [10], Direito et al. [11] and Zhang and Parhi [12], used Spectral Power Analysis as their feature extraction method. They windowed the EEG data and calculated the spectral power property of the data in each window. They reported that this method is very cost-effective in terms of calculation and therefore suitable for portable warning systems. They also found that Spectral Power Analysis could show promising results for seizure prediction, especially in high gamma frequency bands.

An example of the use of Time-Frequency (TF) techniques is the research by Gadhoumi et al. [13]. They used the Continuous Wavelet Transformation (CWT) analysis method for preictal and interictal periods. In this method, wavelet energy and entropy in different frequency bands were calculated using a two-second non-overlapping sliding window. They used LDA for classification and achieved an average sensitivity of 85% in the data of seventeen patients. However, CWT conversion increases file size, and therefore working with large files can lead to hardware and time issues.

In this period, in 2002, the first international workshop on epileptic seizure prediction took place. In 2007 and 2009, the International Workshop on Seizure Prediction (IWSP3) organized two competitions using iEEG datasets from dogs and humans. Although these competitions contributed to significant progress in seizure prediction, the overall performance remained low, and no standardized methods were established.

A small number of data sets were used in the Early Stage and these data sets were of low quality and short time. There is no controlled and standardized evaluation method. The transition to nonlinear chaotic feature extraction methods and the classification process by comparing preictal-interictal periods instead of preictal period analysis alone is noteworthy. in the Early Stage, linear approaches that are frequently used have a computational advantage, but on the other hand, the artifact sensitivity is low. They are also able to identify certain changes that occur before the seizure. Although it was stated in early studies that epilepsy attacks could be predicted between 20-90 minutes before, the performance of these methods was not evaluated in long and high-quality data. Although traditional signal processing methods with phases such as pre-processing, feature extraction, and feature selection provide good accuracy in Machine Learning-based studies conducted in the following years, process complexity, length of time, lack of a generalized model, and constraint in big data can be listed as disadvantages. Since most of the operations are done manually, it can cause data loss.

(2) Deep learning: Due to the development of technology for fully automated systems to overcome the constraints of traditional methods, Deep Learning techniques have been used in the prediction of epileptic seizures since the second half of the 2010s. Deep Learning, a subset of Machine Learning, models human learning ability in computer science and is the multi-layered version of Artificial Neural Networks (ANN) consisting of neurons. In seizure prediction studies, CNN and RNN models, along with their derivatives, are commonly utilized as deep learning algorithms. CNN was the first method employed due to its success in image processing and automatic feature extraction, as proposed by Troung et al. [14]. Troung and colleagues initially transformed the Freiburg and CHB-MIT datasets using STFT and achieved a sensitivity of 89.8%, FPR of 0.17 with CNN classification. Subsequent studies applied various CNN models, such as 1D, 2D, and 3D. Ra and Li [15] utilized a 1D CNN model for feature extraction. However, while this model demonstrates high performance in certain datasets, it can be restrictive in some application areas. The LSTM model, derived from RNN, was first employed by Tsiouris et al. [16], where they developed different LSTM models, addressed the overfitting problem, and achieved a 15–120-minute early alert with a sensitivity of 99.37% and specificity of 99.6%. Liu et al. [17] proposed the Bi-LSTM model, which captures contextual information from both past and future sequences, for a patient-independent model. Their performance reaching 99.37% and 99.6% for sensitivity and specificity. However, since this model operates bidirectionally, it is inefficient in terms of time consumption and computational efficiency. Additionally, it may pose challenges for real-time applications with continuous data flow. Bhattacharya et al. [18] were the first to propose the use of transformers for epileptic seizure prediction, achieving a sensitivity and false-positive rate per hour (FPR/h) as 98.46%, 94.83% and 0.12439, 0, respectively in their study. Deep learning methods are explained in detail in Chapter 4.

 With automatic feature extraction, data loss can be minimized, and model complexity can be reduced. However, the drawbacks include the requirement for large datasets, the need for powerful hardware to support automatic feature extraction, and consequently, increased energy consumption.

3.1 Preprocessing

This phase involves cleaning the raw EEG data by removing artifacts and organizing the data. High-amplitude signals are typically identified as artifacts. The pre-processing stage aims to reduce noise by detecting and eliminating these artifacts, thereby enhancing the accuracy of the classification process. Although the filtering techniques used may vary based on the type of artifacts present, the methods frequently employed for predicting epileptic seizures are generally divided into two categories:

(1) Use of Simple Filters: This method involves applying uncomplicated digital filters, such as band-pass, high-pass, and low-pass filters, by selecting a specific frequency range. Among these, the band-pass filter is the most commonly used in EEG signal analysis. An example of a band-pass filter that allows signals within a defined frequency range is the 'Butterworth' filter. As the filter's order increases, it offers a wider transition zone compared to other band-pass filters.

(2) Use of Signal Analysis Methods: Simple digital filters become ineffective when signal frequency bands overlap with artifact frequencies. To address this, various techniques from Signal Analysis are employed for filtering. These methods are also used for decomposing data and reducing its size. In signal analysis methods used for preprocessing, PCA, which is a linear transformation, is used to rotate the coordinate system to obtain lower dimensional components. ICA, generally utilized to remove artifacts, the first axis, which corresponds to the first component, that is, the main component, is selected according to the maximum variance of the data in a given direction and rotated in the direction in which the variance of the data is maximum. The next component, the secondary axis, is perpendicular to the first component and shows the next highest data variance. Other components are selected in this way and so on [19].

3.2 Feature extraction

The feature extraction phase involves applying methods to the filtered EEG data. In traditional methods, signal analysis techniques are generally used, while in some cases, studies may employ deep learning approaches that utilize automatic feature extraction. Feature extraction is the process of identifying significant properties from the recorded EEG data and obtaining a feature vector. This step reduces the size of the feature vector while selecting the most relevant features for the classifier [20]. Feature extraction methods for the prediction of epileptic seizures are based on various approaches, including multichannel, single-channel, linear, and nonlinear (chaotic) analysis. As shown in Figure 2, feature extraction methods used in studies can be categorized into four groups: time domain, frequency domain, time-frequency domain (signal analysis), and automatic feature extraction (deep learning techniques).

2.png

Figure 2. Feature extraction methods

In the Time Domain, signal analysis is usually performed with statistical calculations. Unlike the time domain, frequency domain relates to the spectral component of a signal and it does not contain any information about time. Time-frequency domain analysis is the method that characterizes the energy of EEG and other signals over time and frequency and decomposes the signal in a two-dimensional level, one dimension in time and the other in frequency.

3.3 Feature selection

One of the important phases for the correct and effective classification of EEG data is feature selection. The feature selection approach falls into two categories: (1) Feature order-based approach; each feature is ranked according to a specific selection criterion, and the top-ranked features are selected as relevant based on a predefined threshold value. (2) Feature subset selection approach; this method evaluates an independent subset of features using a measure of feature selection, such as correlation, consistency, and so on. It generates possible combinations of subsets using various search methods, such as the Greedy algorithm. The feature ranking method is less computationally expensive and has lower complexity. Although it is not commonly used in deep learning algorithms, it can be preferred by most machine learning algorithms to enhance performance.

3.4 Classification

In the general process for predicting epileptic seizures, classification is performed by making a preictal-interictal comparison at this stage. Classification, a type of learning used in statistics, data analysis, pattern recognition, and machine learning, involves determining the class to which the data belongs based on its characteristics, using information about the input dataset. In classification, each test identifies the class of its instance by combining features and finding patterns from the training data that are common to each class. Classification is simply conducted in two stages: first, a classification algorithm is applied to the training dataset, and then the extracted model is validated against a labeled test dataset to evaluate model performance and accuracy.

Algorithms called 'non-black box' classifiers in the studies of the Machine Learning Period in which each process step can be interpreted because it is visible and understandable and methods that do not offer explanations called 'black boxes' have been used. Examples of Non-Black Box methods are 'Decision Trees', 'Random Forest', 'Naive Bayes', and Black Box methods are 'SVM', 'ANN', 'KNN', and 'LDA'. Although there is a wide range of machine learning methods, from simple to extremely complex computational approaches, SVM is the most popular technique used in the classification of epileptic seizure prediction.

In deep learning methods, while filtering or signal analysis techniques can be used for preprocessing, this step can also be skipped, allowing the raw EEG data to be processed directly with deep learning techniques. Additionally, these techniques can automatically extract features. In deep learning applications, EEG data can generally be processed in two ways: as images or time series. For instance, the Convolutional Neural Network (CNN) technique is used for image processing, while derivatives of Recurrent Neural Networks (RNNs) are used for processing time series data. Upon reviewing recent studies, it is evident that, in addition to those employing the transfer and transformer models, various deep learning algorithms can be utilized in hybrid forms, transfer learning and multiple methods can be integrated through fusion techniques.

3.5 Assessment

The results obtained after classification are evaluated using some criteria. Various methods are used to evaluate these prediction results. The assessments are shown Figure 3. 

3.png

Figure 3. Performance evaluation

4.png

Figure 4. SPH/SOP evaluation

The prediction period before the seizure occurs is another criterion. This involves determining the Seizure Prediction Horizon (SPH) and the Seizure Occurrence Period (SOP). As shown in Figure 4, SOP represents the interval in which the seizure is expected to occur, while SPH indicates the time between the alarm and the start of the SOP.

Predicting a seizure too early can cause discomfort for the patient, and there must be sufficient time for preparation before a seizure occurs; thus, determining the optimal SPH/SOP is crucial. In their study on the optimal SPH/SOP duration, Alaei et al. [21] found that the best times are 7 minutes for SPH and 27 minutes for SOP. The k-of-n method is used in SPH/SOP evaluation, where an alarm is triggered if k out of n predictions are identified as preictal within the specified period.

Research on the prediction of epileptic seizures, which began in the early 1970s with linear approaches such as spectral analysis and pattern recognition using spike rates, later solved many problems through the development of machine learning techniques; however, these efforts were still not sufficient. With an increase in the availability of sample data, machine learning approaches gradually gave way to deep learning methods. Recently, the rise of transformer-based models and the use of deep learning models in hybrid and fusion forms have led to significant progress in addressing existing challenges. Nevertheless, despite these advancements, several issues remain unresolved. These persistent challenges can be outlined as follows:

•It remains unclear which model performs best under specific conditions.

•The most critical issue appears to be related to datasets. The EEG datasets used for seizure prediction are not sufficiently descriptive and suffer from artifacts, data deficiencies, and imbalances.

•Single-channel EEG systems, which are typically used to collect data, negatively impact performance, while multi-channel EEG systems are impractical for widespread application.

•While traditional methods demonstrate low performance, advanced algorithms come with high hardware and computational costs. • In addition to brain signals, seizures are often accompanied by changes in heart rate, muscle contractions, tremors, and blood values. However, most studies focus primarily on EEG signals, neglecting these other symptoms.

•Many of the methods used so far have not been evaluated on real patients.

Precise and effective seizure prediction methods are necessary to improve the quality of life for epilepsy patients who do not respond to medication or surgery. Although current studies have made significant progress, there is still no definitive solution. However, the pursuit of more effective models using new deep learning algorithms continues. It seems that we will encounter algorithms like Transformers more frequently. Similarly, methods such as GANs will be preferred, especially to address the dataset problem. First of all, for adaptation to real life, there is a need for a system that is more practical for obtaining EEG signals, can be applied at any time, and is capable of acquiring artifact-free signals that can be easily utilized in practice. This will facilitate the clinical application of future studies. Additionally, we will likely see that it will incorporate not only EEG signals but also different sensor information through fusion models, which are now frequently used in deep learning. In the future, there will be an increased focus on designing real-time alert systems, particularly within the realm of wearable technologies. For this, models that require less hardware will be needed.