Artificial Intelligence-Based Epileptic Seizure Prediction Strategies: A Review

Abstract

Epilepsy, a chronic neurological disorder marked by recurrent and unpredictable seizures, poses significant risks of injury and compromises patient quality of life. The accurate forecasting of seizures is paramount for enabling timely interventions and improving safety. Since the 1970s, research has increasingly focused on analyzing bioelectrical signals for this purpose. In recent years, artificial intelligence (AI), particularly machine learning (ML) and deep learning (DL), has emerged as a powerful tool for seizure prediction. This review, conducted by PRISMA guidelines, analyzes studies from 2020 to August 2025. It explores the evolution from traditional ML classifiers toward advanced DL architecture, including convolutional and recurrent neural networks and transformer-based frameworks, applied to bioelectrical signals. While these approaches show promising performance, significant challenges persist in patient generalization, standardized evaluation, and clinical validation. This review synthesizes current advancements, provides a critical analysis of methodological limitations, and outlines future directions for developing robust, clinically relevant seizure prediction systems to enhance patient autonomy and outcomes.

1. Introduction

A main feature of epilepsy is the presence of abnormal electrical activity in the brain, causing a variety of difficult-to-manage symptoms in patients, including inattention, hallucinations, and uncontrollable movements [1,2,3]. Approximately 2.4 million people develop epilepsy each year [4]. It is worth noting that a seizure caused by epilepsy can be divided into three phases, preictal, ictal, and postictal phases, as depicted in Figure 1.

• Preictal phase: This phase occurs prior to the onset of the seizure and is characterized by an increase in neuronal excitability [5,6].
• Ictal phase: This phase corresponds to the seizure event, from the onset to its end. During this phase, hyperexcitation of neuronal activity occurs [5,6].
• Postictal phase: This phase begins immediately after the ictal phase and is marked by a period of recovery and altered neuronal activity [5,6].

It should be noted that the duration of an epileptic seizure varies from patient to patient, as it depends on several factors, such as (1) the patient’s physical condition at the onset of the seizure, (2) the type and localization of the seizures, and (3) the concept of phases used in seizure prediction [6]. However, this idea can be misleading since determining when a phase starts and ends, as well as its transition to other phases, remains a challenging issue [7]. Despite this, the concept of the ictal phase is fundamental for research related to seizure prediction.

Epilepsy is a global disease; however, its incidence is not evenly distributed among countries, as shown in Figure 2. In this regard, about 50% of people with epilepsy live in low- and middle-income countries [8,9]. It is worth noting that epileptic patients in these countries are often untreated or inadequately treated [4]. This global inequality is visualized in Figure 2, which combines epidemiological and economic indicators to illustrate the gap between disease prevalence and healthcare investment. It is important to mention that it was created based on the data provided by the world bank and Beghi et al. [10,11,12]. In Figure 2a, regions with higher epilepsy prevalence are shown in darker red, while lighter shades represent lower rates. Figure 2b displays the percentage of gross domestic product (GDP) allocated to healthcare, where darker blue tones indicate higher expenditure. Together, these maps reveal a clear inverse relationship: countries with the highest epilepsy burden often allocate the smallest fraction of their GDP to health. This disparity is not only epidemiological but also technological. Limited access to continuous neurological monitoring, specialized care, and high-cost equipment restricts early diagnosis and follow-up in resource-limited settings. Consequently, the development of low-cost, AI-assisted seizure prediction systems based on bioelectrical signals emerges as a feasible strategy to reduce this inequity. These tools could help decentralize epilepsy management, enabling earlier interventions and improving patient safety, particularly in regions where traditional clinical infrastructure is insufficient. The data used to construct Figure 2 can also be accessed at Zenodo under DOI: https://doi.org/10.5281/zenodo.14247540.

The data mentioned above, in Figure 2, highlight the importance of timely seizure prediction, as it can significantly impact three key areas:

• Early intervention. If a seizure is predicted early, implanted neurostimulators can act in advance, preventing the seizure or reducing its severity [13]. This can improve the quality of life for people with epilepsy by minimizing the impact of the condition on daily activities [14,15]
• Safety. Accurate seizure prediction can help individuals take appropriate safety measures, such as avoiding potentially hazardous activities during high-risk periods [16].
• Economic impact. Forecasting epileptic seizures can help reduce healthcare costs associated with emergency room visits and hospitalizations, as people with epilepsy have a 9% probability of requiring hospitalization due to seizure-related injuries [16].

In this context, due to the importance of epileptic seizure prediction, a significant number of methodologies based on the analysis of signals from different organs (e.g., brain, heart, among others) have been proposed since the 1970s [17]. Despite the progress made with these methods, certain limitations remain. Consequently, it has been found that multiparadigm approaches can enhance seizure prediction and increase forecast accuracy in most real-life situations. Over the past ten years, several investigations have focused on analyzing various physiological signals and applying advanced processing techniques, e.g., AI, to develop solutions capable of predicting epileptic seizures in diverse real-life scenarios while maintaining patient comfort.

This paper conducts a review of recent trends in employing machine learning (ML) and deep learning (DL) for predicting epileptic seizures, providing a summary of the developments and methods used in applying these techniques to bioelectrical signals. The review specifically aims to showcase the latest developments in AI-based predictions of epileptic seizures from 2020 to August 2025, a time that has experienced significant advancements in deep learning models, the integration of various types of data, and the application of wearable technologies. This time frame is essential because the area is changing rapidly, with notable advancements in hybrid models and computational efficiency recently emerging, based on foundational research from earlier years. Additionally, a narrative overview is provided of the sources of bioelectrical data, datasets, devices, and techniques utilized for epileptic seizure prediction. Hence, this review combines the latest trends, methods, and performance standards and a general vision of epileptic seizure prediction to offer a current overview of the latest advancements.

This document is structured as follows: Section 2 introduces the foundational concepts of epileptic seizure prediction, such as a general diagram of seizure prediction, bioelectrical signals commonly used, commercial devices, the dataset, and an overview of processing and classifiers employed in epileptic seizure prediction. Section 3 outlines the sequential methodology for conducting the proposed review consultation. Section 4 presents the analysis of current trends and models, and a conscious discussion of some works of ML and DL, leading to Section 5, which offers a discussion of the future tendencies of epileptic seizures according to the findings. Finally, Section 6 details the key results and findings of this review.

2. Background General

A review of the literature reveals a consistent three-phase framework for epileptic seizure prediction, as illustrated in Figure 3. This framework comprises (1) signal acquisition, (2) signal processing, and (3) classification for epileptic seizure prediction. This section provides an overview of each phase to establish the foundational knowledge necessary for understanding research in the field.

• Phase 1, signal acquisition. This initial phase is detailed by first outlining the bioelectrical signals commonly used for seizure prediction. This is followed by a discussion of some available datasets. Finally, a brief discussion of devices developed for seizure prediction is presented.
• Phase 2, signal processing. This phase covers the core components of processing acquired signals, specifically feature extraction techniques and feature selection methods, which are critical for optimizing classifier performance.
• Phase 3, classification. This final phase presents a concise discussion of ML and DL algorithms applied to seizure prediction, including a comparative analysis of their respective advantages and disadvantages.

2.1. Phase 1, Signal Acquisition

2.1.1. Bioelectrical Signals Commonly Used for Seizure Prediction

The data used in the studies for predicting epilepsy can be obtained through different sensors placed on specific parts of the human body, such as the brain and heart, among others, as seen in Figure 4. The signals produced by these organs are known as physiological signals, and the primary challenge for signal processing algorithms is dealing with noise corruption [18]. Power line interference, baseline wander, movement artifacts, equipment artifacts, and biological signals from surrounding organs are common sources of noise [19]. In this regard, by eliminating or mitigating these interferences, the measured signals can be properly analyzed to uncover diagnostic information.

The scalp electroencephalogram (sEEG) is a recording of the electrical signals generated by the coordinated action of brain cells measured from the scalp [19]. sEEG amplitudes range from 1 to 200 µV in awake subjects but may increase tenfold during epileptic episodes [19]. The 10–20 electrode system is commonly used for sEEG recordings. However, sEEG signals can be affected by muscle or heart artifacts, eye movements, electromagnetic fields, and poor contact between the electrode and the scalp, as these interferences are often registered simultaneously.

In contrast, an intracranial electroencephalogram (iEEG) is obtained using depth electrodes. This method is invasive and is typically performed in clinical settings at specialized hospitals. The placement of the electrodes is determined clinically based on the patient’s condition, and once implanted during surgery, the electrodes cannot be repositioned. The amplitude of iEEG signals is approximately ten times higher than that of sEEG, and noise artifacts are generally minimal in these recordings [20].

Another signal widely used in seizure forecasting is the electrocardiogram (ECG). ECG is a noninvasive method for detecting various diseases, with the 12-lead ECG being the most common type, as it uses anatomical landmarks to position the leads [21]. ECG recordings can be affected by muscle contractions, electrode contact issues, and power line interference. It is essential to highlight that the bioelectrical signals collected from ECG can be valuable because epilepsy can lead to changes in autonomic functions that result in cardiovascular manifestations [22]. In addition to ECG signals, the use of other types of signals in the prediction of epileptic seizures has been reported, including photoplethysmography (PPG), oxygen saturation (SpO2), electrodermal activity (EDA), accelerometry (ACC), arm surface electromyography (asEMG), and behind-the-ear EEG (beEEG) [23,24,25,26,27].

Table 1. Broad characteristics of the most employed bioelectrical signals for epilepsy seizure prediction.

Characteristics	iEEG	sEEG	ECG
Origin	Brain neuron activity.	Brain neuron activity.	Action potentials of heart muscle cells.
Frequency range (Hz)	0.1–100	0.1–100.	0.5–100
Amplitude (µV)	5–1000	1–100	100–3000
Invasive?	Yes (intracortical electrode)	No (surface electrode)	No (surface electrode)
Is it affected by power line interference noise?	Yes, it is caused by electrical wiring, which results in frequency interference in the range of 50 to 60 Hz, leading to amplitude distortion; this is less prone to environmental noise than sEEG due to its placement inside the skull. The iEEG can reach a signal–noise ratio of around 30–50 dB, depending on the setup and conditions.	Yes, it is caused by electrical wiring, which results in frequency interference in the range of 50 to 60 Hz, leading to amplitude distortion. Depending on the setup and conditions, the sEEG can reach signal noise ratio of around 10–20 dB.	Yes, it is caused by electrical wiring, which introduces interference with a frequency in the range of 50 to 60 Hz and an amplitude of up to 5 mV, distorting the waveform and potentially masking important features such as the T-wave, QRS complex, and P-wave. The ECG can reach signal–noise ratio of around 10–30 dB.
Is it affected by baseline wander noise?	Yes, but less than sEEG because the electrodes are placed directly within brain tissue, minimizing interference from scalp, muscle, and environmental noise. The result in iEEG is 5–10 times lower baseline than sEEG.	Yes, it is caused by head movement, electrode contact, and sweat on the scalp. These lead to alterations in frequencies below 1 Hz with amplitudes of 200 to 300 mV, resulting in an sEEG signal–noise ratio of around 10–20 dB.	Yes, it results from body movements and inadequate electrode contact, and high impedance between the electrode and skin. This results in alterations in frequencies from 0.05 to 1 Hz, causing distortion in QRS complex and ST segment and low-frequency components, resulting in an ECG signal noise ratio of around 5–10 dB.

presents the main features of the primary signals used for epilepsy prediction, namely iEEG, sEEG, and ECG.

Previously, it described the characteristics of the three types of signals currently used to predict epileptic seizures, as well as the signals being analyzed for potential use. Understanding the characteristics of each signal type is crucial because seizure prediction research faces a trade-off between sensitivity and specificity, while also considering the invasiveness of the signal acquisition technique. While the iEEG is cleaner than the sEEG signal, comparative studies show that iEEG provides a 5 times higher signal-to-noise ratio (50 dB vs. 10 dB for sEEG) [20]. However, iEEG requires invasive implantation, whereas sEEG has a non-invasive nature. Nevertheless, investigations such as those by Eltrass and Ghanem have developed a system based on filtering techniques to reduce noise and artifact suppression in sEEG recordings [28]. Other types of techniques can be employed in signals to reduce the noise effect. Additionally, the incorporation of various signal types (ECG, PPG, SpO2, EDA, ACC, asEMG, beEEG) is currently being investigated to reduce false negatives. It is crucial to mention that, accordingly, false negatives pose a risk to patient safety, with a 1.7 times higher probability of injury derived from a seizure [26]. As a result, the importance of minimizing false negatives is emphasized.

In summary, according to that observed in this section, although the choice of signal type is fundamental for addressing the trade-off between sensitivity, specificity, and invasiveness in seizure prediction, the actual development and validation of predictive algorithms ultimately depend on the availability of high-quality and representative datasets. Comprehensive and well-annotated databases are indispensable to ensure reproducibility, enable comparative analysis across studies, and support the generalization of AI-based methodologies. Therefore, the subsequent section focuses on some available datasets that constitute the empirical foundation for advancing research in seizure prediction.

2.1.2. Available Datasets

The choice of a specific signal type and its inherent trade-offs directly inform the next critical step: acquiring the appropriate data for algorithm development. The performance and generalizability of any AI-based seizure prediction model are fundamentally contingent on the quality, diversity, and structure of the datasets used for its training and validation.

To be truly useful for seizure prediction research, a dataset should comprehensively document several key attributes: (1) the number of recordings; (2) the type of subjects, as some databases provide information about both human and canine patients, however in this review only present databases of human; (3) details about the type of seizure, i.e., whether it is of focal, generalized, or unknown origin; (4) data continuity, which can be either continuous data recorded consistently over specific time lapses or non-continuous data, which often leads to random segmentation without time relationships; (5) the recording length per segment; (6) the time of seizure occurrence for each patient; (7) the sampling frequency; and (8) the balance in the distribution of preictal, ictal, and interictal classes. This information is summarized in

Table 2. Features of the open-access databases.

Database	No. of Patients	Signal Type	No. of Channels	Data Continuity	Recording per Segment (s)	Balance Class	Sampling Frequency (Hz)
Melbourne-neuroVista seizure trial [29]	15	iEEG	16	Noncontinuous	Average of 107	No	400
Kaggle-Melbourne-University AES-MathWorks-NIH [30]	3	iEEG	16	Noncontinuous	600	No	400
Freiburg [31]	21	iEEG	128	Short-term continuous	Average of 3000	No	256
Bern Barcelona [32]	5	iEEG	More or less than 64	No data	20	No data	512
CHB-MIT Scalp EEG [1]	22	sEEG	23–26	Continuous	Average of 36,000	No	256
Neurology and Sleep Centre Hauz Khas [33]	10	sEEG	1	Noncontinuous	5.12	Yes	200
TUH EEG Epilepsy Corpus (TUSZ) [34]	200	sEEG	23–31	Short-term continuous	3600	No	Least 250
Helsinki University Hospital EEG [35]	79	sEEG	19	Short-term continuous	Average of 4440	No	256
Siena Scalp EEG [36]	14	sEEG	20–29	Short-term continuous	Differing	No	512
Postictal Heart Rate Oscillations in Partial Epilepsy [37]	5	ECG	1	Short-term continuous	Differing	No	200
SeizelT1 [38]	82	sEEG/ECG	25/1	Continuous	Average of 36,000	No	250
PEDESITE: Personalized Detection of Epileptic Seizure on the Internet of Things (IoT) Era [39]	1200	sEEG, ECG, PPG, SPO2, EDA, 3D-ACC and asEMG	-----	-----	-----	-----	-----

, which presents the features of various databases. The epilepsy databases offer a wide variety of data, including iEEG and sEEG, ECG, and other bio signals. The iEEG datasets, such as Melbourne NeuroVista and the Kaggle American Epilepsy Society (KAES), are notable for their noncontinuous data and moderate channel counts, while sEEG datasets, such as Children’s Hospital Boston-Massachusetts Institute of Technology (CHB-MIT) and Temple University Hospital (TUH), provide continuous long-term recordings. The diversity in sampling frequencies (ranging from 200 Hz to 5000 Hz) and the number of channels (ranging from 1 to 128) reflects different experimental setups and target applications. This variety is crucial for research into seizure prediction, detection algorithms, and other epilepsy-related studies, offering flexibility for different analytical approaches depending on the dataset’s structure.

Future efforts in dataset creation should prioritize standardized protocols for data collection, long-term continuous monitoring, and annotated seizure metadata (e.g., focal versus generalized onset) to enhance reproducibility. Emerging trends may include the use of IA in database creation to pool decentralized data or for synthetic data generation to mitigate class imbalance. Addressing these gaps will be crucial in developing techniques, methodologies, and systems that are more robust in seizure prediction.

Although the availability of open-access databases has greatly facilitated reproducibility, several limitations persist in data quality and annotation standards. Many datasets exhibit inconsistent labeling of seizure events, incomplete metadata, and uneven representation of demographic and clinical variables, which may introduce biases in model generalization and hinder cross-study comparability. To address these gaps, emerging strategies, including federated learning for decentralized training, and differential and generation of synthetic EEG data to balance underrepresented classes, are progressively adopted. These developments reflect the field’s growing commitment to harmonizing data quality, privacy protection, and methodological rigor in AI-based seizure prediction research.

In this context, while the optimization of datasets through standardized protocols and the integration of synthetic data generation strategies provide the necessary foundation for advancing algorithmic performance, the translation of these advances into clinical practice increasingly relies on the development of devices. These devices operate the insights gained from diverse bio signals and AI-based approaches into wearable and user-friendly technologies capable of supporting real-time seizure prediction. Consequently, the following section examines representative examples of commercial solutions that leverage physiological signals and AI-based algorithms to bridge the gap between research and practical seizure management.

2.1.3. Devices for Epileptic Seizure Prediction

The devices represent the translational application of AI methodologies, directly leveraging physiological signals to provide real-world monitoring and alerting capabilities. The performance and clinical viability of these devices are fundamentally dependent on the robustness of the data and algorithms that power them. In recent years, several wearable devices to meet the need for seizure prediction and alert caregivers of possible seizure occurrences have been developed. These devices use various physiological signals [40,41,42], such as ECG, sEEG, body temperature, and movement, among others. In this regard, the Embrace2 is a food and drug administration-cleared medical device that includes an EDA sensor, an ambient temperature sensor, a 3-axis-ACC, and a gyroscope. As a medical device, it requires a valid prescription for use. It employs machine learning (ML)-based algorithms to detect seizures lasting more than 20 s. It is important to note that seizures without movement cannot be detected. The Embrace2 has a bracelet-like design and is water-repellent [40]. On the other hand, the Epihunter Pro, a portable sEEG headband with dry gold-plated copper electrodes, is classified as a Class I medical device software with European compliance. It uses electrodes, which are positioned at FP1 and F7, and an AI-based algorithm is used to predict a seizure 30 s before it occurs [41]. The Epicare at Home Sensor Dot and its accessories are European standard-compliant medical devices that also require a prescription. This device is a wearable beEEG with an ACC that records vital signs, including heart rate, breathing rate, and activity index [42]. It is important to clarify that studies involving wearable or real-world monitoring devices were not systematically identified, and only representative examples such as Embrace2, Epihunter Pro, and Epicare were described; these examples were selected to illustrate the translational potential of AI-assisted wearable technologies rather than to bias the evidence base toward any specific device or study design.

It is crucial to mention that many enterprises, foundations, and researchers are currently working on developing wearable devices to detect and predict epileptic seizures. For instance, the study performed by Wu and Li [43] focuses on two main aspects: (1) developing a real-time early warning system for seizure prediction using sEEG, employing wireless, portable, and real-time transmittable devices, and (2) evaluating a model with the Bonn dataset. This model, based on six features (i.e., detrended fluctuation analysis, Petrosian fractal dimension, Fisher information, Hjorth fractal dimension, and singular value decomposition entropy), combined with Adaptive Boosting, achieved 100% accuracy in predicting a seizure 23.5 min prior to onset. Xiong et al. [44] proposed a model based on heart rate data collected through a mobile app on a smartwatch, recording 125 seizures over an average of 562 days to estimate seizure risk for a future time window of 60 days. This prospective approach was applied to evaluate forecasting algorithms for seizure prediction, with long-term data being retrained weekly or after each seizure occurred. As a result, 9 out of 13 participants obtained an average area under the curve (AUC) of 0.73. When evaluating subject-specific forecasts with forward-looking data, the mean AUC is 0.77. Zambrana-Vinaroz et al. [24] developed a seizure prediction tool based on beEEG, ECG, and PPG signals obtained from 10 epilepsy volunteers with a total recording time of 6.66 h. The tool achieved an accuracy of 85.4% for predicting epileptic seizures 60 s before the onset of the seizure.

On the other hand, Stirling et al. [45] developed a seizure prediction model where the data provided by smartwatches is used, including heart rate, sleep intervals, and number of steps. Participants manually reported these data points into a smartphone application. The results yielded an average AUC of 0.74. Similarly, Nasseri et al. [46] used EDA, ACC, temperature, and blood volume. As a result, the average AUC was 0.75, with 5/6 participants performing better than chance. Meisel et al. [47] utilized the commercial Empatica E4 device to collect data on EDA, ACC readings, blood volume PPG, and temperature from 64 patients, resulting in the documentation of 452 seizures over 2311 h of recording.

As mentioned in this section, recent wearable technologies demonstrate promising seizure prediction capabilities through the use of machine learning algorithms, offering noninvasive monitoring and early warnings. However, critical challenges remain, such as limited sensitivity to non-motor seizures and interpatient variability in model performance. However, for these devices to reach patients, they must have FDA or CE certifications, a requirement that several experimental devices lack, as they require compliance with a series of minimum standards. For example, the FDA-approved Embrace2 device underwent a multicenter clinical evaluation (NCT03299842), achieving real-world viability but showing variability in predictive performance across patient profiles. Similarly, CE-compliant devices, such as Epihunter and Epicare, highlight the importance of transparent algorithms and post-marketing surveillance for ongoing safety validation. These experiences demonstrate that algorithm customization, harmonized validation protocols, and regulatory alignment are essential to bridge the gap between laboratory performance and clinical implementation. Therefore, future advances should emphasize adaptive retraining schemes, large-scale multicenter trials, and standardized reporting frameworks to ensure both technical robustness and regulatory acceptance.

2.2. Phase 2, Signal Processing

As highlighted in Section 2.1, the diversity of bio signals, datasets, and emerging wearable technologies provides a solid foundation for advancing seizure prediction systems. However, the effectiveness of these approaches is ultimately determined by the ability to process the acquired signals in a manner that enhances their predictive value. In this regard, Phase 2 focuses on signal processing, encompassing both feature extraction and feature selection techniques. These steps are critical for mitigating noise and artifacts, capturing relevant temporal, spectral, and nonlinear characteristics of physiological signals, and ensuring that only the most discriminative features are employed to optimize classifier performance in the subsequent phase.

2.2.1. Features Extraction Techniques

As mentioned above in Section 2.1.1, physiological signals are often contaminated by various noise components, and this can be generated by power line interference, baseline wander, movement artifacts, equipment artifacts, and biological signals produced by surrounding organs, among others. Consequently, in most scenarios, the domain and frequency content are heavily altered, which can lead to the loss of the signal [18]. In this context, a preprocessing step is necessary before feature extraction. Generally, the preprocessing stage involves using a filtering scheme to remove most of the noise present in the signal, where finite or infinite impulse response filters are typically employed [48]. Additionally, the introduction of adaptive filters has been proposed to improve noise removal further. Although notable results have been achieved, the choice of technique is an important aspect to consider, as it directly impacts the modal performance. If an incorrect algorithm is used, the filtered signal may undergo undesired frequency modifications, generating spurious components that lack physical meaning [48]. Consequently, several approaches for selecting and optimizing preprocessing strategies have been proposed [46].

Once the signal is filtered, the features it contains can be extracted. These features are classified according to the domains to which they belong, with the most common being (a) time domain (TD), (b) frequency domain (FD), (c) time-frequency domain (TFD), (d) Nonlinear features (NF), and (e) High-order spectral features (HOSF), which are briefly explained below.

(a)

TD features are known for having the lowest computational burden since they do not require any transformation domain signal transformation [49]. TD techniques can capture temporal patterns and signal variations by analyzing waveforms over specific time intervals. The most employed techniques include (1) statistical methods (e.g., variance, mean, standard deviation, among others), (2) energy-related features, (3) the number of zero-crossings, (4) histogram analysis, (5) the number of slope sign changes, and (6) waveform patterns [49,50]. It should be noted that these techniques reveal the internal dynamics of the signal over time [50]. TD features help distinguish between preictal and interictal states to identify seizure-related features [51,52].

(b)

FD features are obtained from the signal in the frequency domain, which is generated by applying the Fourier Transform (FT), or one of its variants, to the signal in the time domain [53]. Once the transformation is carried out, spectral power analysis is performed to extract features, providing information about the signal’s energy distribution and power characteristics. FD techniques are known for enhancing predictive capabilities by assessing a broader range of signal properties, consequently raising the reliability and accuracy rates of epilepsy prediction in bioelectrical signals [52,54].

(c)

TFD algorithms allow for the analysis of physiological signals that exhibit both time-variant and transient characteristics [52], as their mathematical framework is well-suited for such applications [55]. In general, TFD algorithms enable the following: (1) a comprehensive analysis of dynamic behavior over time and across different frequency bands, (2) improved time resolution for detecting suspicious activities, and (3) effective analysis of transient events or changes in signal dynamics occurring over short time intervals; features are typically obtained using techniques that represent time and frequency simultaneously, such as the Wavelet Transform and its variants, the Short-Time Fourier Transform (STFT), and empirical mode decomposition along with its different versions. These algorithms have been used to develop methods capable of predicting epileptic events based on bioelectrical signals with reasonable accuracy [55,56,57,58].

(d)

NF were applied to signals that exhibit a nonlinear and chaotic nature, characterized by patterns that repeat across different scales [59,60]. To capture these complex dynamics and patterns, it is necessary to employ algorithms specifically designed for this purpose [60]. The most commonly used NF techniques include Lyapunov exponents and fractal estimation algorithms, such as Higuchi’s method, Box dimension, and detrended fluctuation analysis [60]. These techniques provide a more comprehensive and detailed analysis of physiological signals compared to linear methods, enabling the development of more effective classification schemes [61].

(e)

HOSF analysis is a nonlinear method that can handle higher-order data and provide comprehensive signal characterization, as it preserves both the phase and magnitude of the frequency components [62]. Moreover, this technique generates smoother spectral lines, allowing it to be effectively applied to weak and noisy signals [5]. These techniques have been employed in sEEG signals to predict an epileptic seizure [5,63]. In particular, HOSFs offer a robust framework for analyzing complex and nonlinear bio signals, allowing meaningful characteristics from signals such as sEEG, iEEG, ECG, EMG, and general bio signals to be extracted. These characteristics provide deeper insights into physiological processes and the high-order spectral features in extracting valuable information from bio signals that traditional methods (e.g., FT, statistical features, cross-correlation, among others) may not capture [64].

Once features are extracted, it is important to acknowledge that not all of them contribute equally to the discrimination between preictal, ictal, and interictal states. The presence of redundant or irrelevant features may degrade classifier performance, increase computational cost, and reduce generalizability. Therefore, the subsequent step involves feature selection, which ensures that only the most informative and discriminant features are retained for model training, while simultaneously addressing the curse of dimensionality and enhancing the robustness of seizure prediction algorithms.

2.2.2. Feature Selection

Feature selection techniques play a paramount role in seizure prediction processing, as they are used either to select the most discriminant features for training a classification model or to perform dimensionality reduction [65,66,67,68]. The first approach focuses on methods that assess the relevance of features to identify the most important ones, thereby improving the classification algorithm’s accuracy. Examples of such methods include the analysis of variance and the Kruskal–Wallis test, among others [65,66]. The second approach involves strategies for dimensionality reduction while preserving essential information; commonly used methods include principal component analysis, singular value decomposition, principal component singular value decomposition, and independent component analysis, among others [67,68].

The process of feature extraction and selection establishes the foundation of algorithms for epileptic seizures. High-dimensional, redundant, or irrelevant features may not only increase computational burden but also compromise generalizability, leading to suboptimal classifier performance. Conversely, carefully selected features enhance the discriminative capacity of models, improving their robustness and predictive accuracy across heterogeneous patient data. Therefore, once an optimal feature set is obtained, the subsequent step involves the design and implementation of classification strategies for epileptic seizure prediction. In this context, both ML and DL approaches have been extensively investigated for epileptic seizure prediction, each offering distinct advantages and limitations.

2.3. Phase 3, Classification

The thorough processes of signal acquisition and preprocessing culminate in the critical stage of feature selection, which establishes the foundational input for classification algorithms. The quality of this feature set is paramount; high-dimensional or redundant features can increase computational burden and compromise model generalizability, whereas a refined, discriminative set directly enhances robustness and predictive accuracy. Consequently, with an optimal feature set defined, the final and decisive phase involves selecting a classification strategy to predict an epileptic seizure. This stage is now dominated by AI, and with the development of novel AI-based strategies, the field has witnessed remarkable advances. These algorithms are generating enhanced tools for prevention, diagnosis, and treatment [69]. AI-based strategies for epileptic seizure prediction must be analyzed. Broadly speaking, classification strategies fall into two main types of algorithms:

• Supervised algorithms: These require labeled data, which are employed during the training and validation stages to develop the classification strategy.
• Unsupervised algorithms: These do not require labeled data; instead, the algorithm clusters data with similar features during its training stage.

Recently, algorithms based on both ML and DL have been developing AI-based strategies for seizure prediction. To illustrate the main differences between these two frameworks, Figure 5 provides a comparison.

From this figure, several observations can be made regarding the procedures used:

• Data required for the training stage: DL techniques usually require more data for training than ML methods due to the supervised training algorithms typically employed.
• Computational load: ML techniques generally require fewer computational resources compared to DL techniques. Therefore, when the computational load becomes a critical factor in algorithm selection, a good balance between computational demand and resulting accuracy should be achieved.
• Training time: DL algorithms usually require more training time since they process a large amount of data to achieve optimal results. Conversely, ML algorithms can be trained in less time. However, the selection criteria often depend on the presence of noise in the data.

In the following subsections, the most recent AI-based proposals are discussed and compared to provide insights into their accuracy and computational burden [70,71].

In the following subsections, ML and DL techniques are presented. ML techniques are applied once the optimal feature set is determined. ML algorithms constitute the first approach for classification. These methods have been widely employed in seizure prediction due to their ability to handle structured features, moderate computational requirements, and relatively high interpretability. Section 2.3.1, therefore, presents some applied ML classifiers for epileptic seizure prediction.

2.3.1. ML-Based Algorithms for Epileptic Seizure Prediction

After selecting the extracted features, the development of classifiers can be appropriately carried out. This section analyzes the classification algorithms that are most frequently employed.

(a)

The Support Vector Machine (SVM) algorithm is a well-known classification strategy that aims to separate two different classes using hyperplanes. During the training stage, the algorithm determines the parameters of two hyperplanes that maximize the separation between the classes, typically represented as a linear boundary [72,73]. However, when the data cannot be linearly separated, the algorithm applies a kernel function to map the data into a higher-dimensional space where linear separation becomes feasible. For this purpose, radial basis function, polynomial, and linear kernels are commonly employed [74].

(b)

The K-nearest Neighbors (KNN) algorithm has been extensively utilized in numerous studies by researchers around the world to predict epileptic seizures [75,76,77]. This algorithm is a simple yet effective ML classifier that relies on recent training examples. KNN assigns multiclass labels based on two factors: the number of nearest neighbors (K) to the data point being classified, and the selection of K [78,79]. By calculating the distance between the new data vector and all existing vectors, the model is approximated [80].

(c)

Decision Tree (DT) is an effective classifier that provides a straightforward and adaptable implementation, as it can be programmed using a series of if-else rules [81]. Reasonable accuracy can be achieved if the feature sets do not exhibit a significant degree of overlap [82,83]. These tests are repeated until a terminal node (leaf node) is reached [82]. Once a leaf node is reached, the tree predicts the associated outcome, completing the classification. In other words, the classifier operates by taking an object described by a set of properties as input, which are used to build a classification tree model, where decisions at each stage are determined by previous branching operations.

Although diverse ML classifiers have been successfully implemented, their performance varies considerably depending on dataset size, feature quality, and computational resources. For this reason, it is important to know the advantages and disadvantages to guide the appropriate selection of models for seizure prediction tasks.

Advantages and Disadvantages of the ML Classifiers

ML classifiers differ in terms of accuracy, computational cost, scalability, and interpretability. To select the most appropriate classifier for a given application, it is essential to understand these characteristics. Some classifiers might offer high precision but require significant computational resources, while others may be faster and more efficient but less accurate. The decision on which classifier to use should consider these trade-offs to match the specific needs of the task. The main advantages as well as disadvantages of the ML classifiers presented in this article are condensed in

Table 3. Advantages and disadvantages of diverse ML classifiers.

ML Classifier	Advantages	Disadvantages
SVM	Effective in high-dimensional spaces. Can handle nonlinear data (i.e., sEEG, iEEG, ECG, etc.), employing kernel. functions. High accuracy, especially for small to medium datasets with clear class separation.	Computationally intensive for large data (i.e., large databases). Requires careful turning of kernel functions and hyperparameters. Not easily scalable for large datasets due to memory and computation requirements. Difficult to interpret, especially with nonlinear kernels.
NB	Easy to interpret due to its probabilistic nature. Fast computation even with large data. Highly scalable; performs well with large datasets. Low computational cost; very fast training and prediction. Moderate accuracy performs well with categorical data.	Assumes strong feature independence (this is supposed to be a problem in EEG data that may not always hold). Sensitive to rare events in data (may perform poorly with highly correlated features or noisy data. Characteristics that are commonly in bio signals).
KNN	Easy to understand and implement. No training phase required. Can adapt to new data in an online setting.	Requires careful turning of k and hyperparameters. Computationally expensive for large datasets. Sensitive to noisy data and irrelevant features. Low scalability: computational cost increases significantly with dataset size.
DT	Minimal data processing required. Easy to interpret and visualize. Can handle with categorical and continuous data. Easily scalable.	Prone to overfitting with noisy data. Requires pruning to improve generalization. Small changes in data can lead to significant model variations. Depth and size can become issues with large datasets.

[84,85].

Despite the effectiveness of traditional ML approaches, their reliance on handcrafted features often constrains performance in large-scale or highly heterogeneous datasets. To overcome these limitations, DL methods have emerged as an alternative, offering automated feature extraction and improved modeling of nonlinear dependencies. Section 2.3.2 introduces these DL-based approaches for seizure prediction.

2.3.2. DL-Based Algorithms for Epileptic Seizure Prediction

As an alternative approach to overcoming the challenges faced in conventional ML methods, a new branch of ML known as DL has been explored in seizure prediction. DL methods, which are part of AI algorithms, offer a distinct advantage over traditional ML approaches. One key difference is that in DL, the feature extraction process is automated, unlike in ML-based strategies, where hand-crafted features are often required [86,87]. The main advantages of DL-based methods include their capability of handling large and complex datasets, as well as modeling intricate relationships. However, these benefits come with certain drawbacks, such as high computational costs, a risk of redundancy, and limited interpretability [86,87]. The following section briefly describes some of these DL classifiers.

(a)

Convolutional Neural Networks (CNNs) are bioinspired algorithms capable of extracting relevant features without requiring human assistance [86,88]. A CNN classifier can be developed as follows: (1) the selection and number of convolutional layers must be determined, as they set the dimensionality of the input layers; next, (2) the selection of kernel size, number of filters, stride, padding, and the number of pooling layers must be made to reduce dimensionality and computational complexity while retaining essential features; after that, (3) it is necessary to define the activation functions, and (4) the fully connected layers that define the classifier output [70,86,88]. It should be pointed out that the selection of the filters used in the convolutional layers determines the classifier’s accuracy [86]; hence, they must be carefully determined.

(b)

Recurrent Neural Networks (RNNs) are characterized by being a type of neural network that is well-suited for analyzing data with temporal patterns [89]. The stages of this classifier are (1) the input layer, whose size is determined by the time-series sequences, (2) the number and size of the recurrent layers, which define the classifier’s ability to capture the temporal dependencies of the training data, (3) the use of dropout layers between recurrent layers to prevent overfitting, and (4) a fully connected layer added to obtain the classification result. An important aspect to highlight is that the selection of the number of hidden layers, neurons, and activation functions in these layers plays a crucial role in determining classification accuracy [90,91].

(c)

Transformer-based methods (TBMs) models combine the strengths of recurrent architecture with attention mechanisms to enhance the specificity and sensitivity of the resulting models. One key advantage of TBMs is their capability to perform parallelization during training, which makes the process faster and more efficient. Additionally, TBMs mitigate the vanishing gradient problem, resulting in easier training and the development of classifiers with a higher resistance to uncertainty [92]. The process begins by passing the inputs through a positional encoding layer, which includes a multi-head self-attention mechanism. Then, in step two, a dropout and normalization layer are applied to enhance generalization capabilities. Step three involves a fully connected feed-forward network, followed by step four, where a decoder layer, similar in structure to the encoder, incorporates a multi-head attention mechanism that attends to the encoder output, ensuring that the result depends only on the known outputs [93,94]. Various TBM variants have emerged, including encoder-only transformers, bidirectional encoder representations from transformers, decoder-only transformers, star-transformers, BigBird, and generative pre-trained transformers, among others.

While DL models provide substantial improvements in accuracy and scalability, their deployment is associated with significant challenges, including high computational cost, the need for large, annotated datasets, and limited interpretability. Consequently, an evaluation of their advantages and disadvantages is necessary to contextualize their applicability in seizure prediction.

Advantages and Disadvantages of the DL Algorithms

DL algorithms differ in terms of accuracy, computational cost, scalability, and interpretability. It is essential to understand these characteristics in order to select the most appropriate DL model for a given application. Some DL models may offer high accuracy but require significant computational resources and specialized hardware, while others might be more computationally efficient but less accurate. Additionally, DL models tend to be less interpretable, making it challenging to understand the decision-making process. The choice of which DL model to use should consider these trade-offs to meet the specific needs of the task. The main advantages and disadvantages of the DL algorithms presented in this article are condensed in

Table 4. Advantages and disadvantages of diverse DL algorithms.

DL Algorithm	Advantages	Disadvantages
CNNs	Excellent for spatial features extraction (i.e., sEEG and iEEG). Effective in large-scale datasets. High accuracy in image-based and spatial pattern recognition tasks. Highly scalable using GPUs and parallel processing.	Requires large, labeled datasets for training. Struggles to capture temporal dependencies. Computationally expensive. Low interpretability due to the complexity of layers and parameters.
RNNs	Suitable for sequential and time-series data. Can capture long-term temporal dependencies in data. Effective for time series data. Good accuracy for sequential and temporal data. Moderate interpretability.	Prone to disappearing or exploding gradient issues over long sequences. Slow training times and resource-intensive. Requires large amounts of labeled data. High computational cost, especially with long sequences due to vanishing gradients. Limited scalability for long sequences.
TBMs	Handles long range dependencies are better than RNNs and CNNs. High accuracy for sequential tasks. Highly flexible for capturing complex patterns in data. Scalable and parallelizable.	Requires considerable computational resources and memory. A training in a large amount of data is required. Complex model tuning and hyperparameter optimization. Low interpretability due to complex architecture makes it challenging to understand.

[72,95].

3. Review Methodology

This section outlines the methods for searching and identifying studies investigating the use of ML and DL methods in the epileptic seizure prediction using bioelectrical signals. This review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines [96]. The search was restricted to articles published between 2020 to August 2025, written in English; preprints, reviews, articles retracted, and letters to the editor were excluded. The search was conducted in PubMed, Scopus, and Web of Science databases. The following key terms were combined using Boolean operators:

TS = ((“epileptic seizure prediction” OR “seizure forecasting” OR “seizure prediction”) AND (“artificial intelligence” OR “machine learning” OR “deep learning”) AND (eeg OR electroencephalogram OR electroencephalography OR ecg OR electrocardiogram OR photoplethysmography OR ppg OR bioelectrical signals)) AND PY = (2020 OR 2021 OR 2022 OR 2023 OR 2024 OR 2025) AND LA = (English)

This search resulted in the identification of 642 articles (Scopus = 272, PubMed = 89, WOS = 281).

Following this, the duplicated articles were removed through a process involving manual cross-referencing of Digital Object Identifiers (DOIs), resulting in 361 unique articles.

Studies were analyzed by the following criteria for inclusion in the present review.

• Population: Human participants diagnosed with epilepsy.
• Objective: Application of ML or DL algorithms for the prediction of epileptic seizures (not limited to ictal detection/classification).
• Data: Utilization of bioelectrical signals (e.g., iEEG, sEEG, ECG, PPG, or multimodal approaches).
• Outcomes: Reporting of at least one performance metric (e.g., accuracy, sensitivity, specificity, AUC) and the time of prediction (seizure prediction horizon (SPH)).

Subsequently, after applying these exclusion criteria, 130 articles were included in the qualitative synthesis. The full selection process is summarized in Figure 6.

For each selected article, a data extraction procedure was employed to collect key information encompassed: the source of the data, the type of bioelectrical signals used, the ML or DL algorithms implemented, the SPH, and reported performance metrics.

It is crucial to clarify that to mitigate publication and reporting bias, it included studies that reported negative or inconclusive findings, provided they met the eligibility criteria. Throughout the review, we avoided selectively emphasizing and interpreting performance claims in the context of both favorable and unfavorable evidence.

Furthermore, to ensure transparency and reproducibility, the inclusion process was reverified up to August 2025. Each eligible paper was screened according to the predefined criteria. Ambiguous cases, such as studies applying ML or DL for seizure detection without a clearly stated prediction horizon, were excluded to maintain methodological consistency. This approach prevents the misclassification of detection studies as predictive works and reinforces the reliability of the final synthesis.

4. ML and DL in Epilepsy Seizure Prediction

The search strategy described in Section 3 was designed to capture both classical ML approaches and emerging DL paradigms, using specified keyword combinations. The relative representation of ML versus DL reported here reflects the results returned by the searches for 2020–2025. It is important to clarify that they did not result from any weighting or rebalancing by the authors. We did not impose measures to equalize counts across methodological families, as doing so would bias the evidence base. Consequently, the higher prevalence of DL reports observed in this time window is presented as an empirical outcome of the search rather than a curatorial choice and shows the tendency of employing these types of algorithms.

4.1. ML in Epilepsy Seizure Prediction

4.1.1. SVM

Among ML techniques, the SVM remained one of the most widely adopted classifiers, with 17 studies identified during this period that implemented it either as a standalone model or in conjunction with other classification algorithms [97,98,99,100,101,102,103,104,105,106,107,108,109,110,111,112,113,114]. Of these 17 investigations, 13 studies utilized sEEG signals, either independently or combined with additional modalities such as functional near-infrared spectroscopy, ECG, or heart rate variability (HRV), for seizure prediction. In these studies, the SPHs ranged from a few seconds to approximately one hour, and sensitivity varied between 90% and 99%. Such results demonstrate a high capability for detecting preictal events, suggesting a potentially reliable warning performance under controlled clinical conditions.

Nevertheless, elevated sensitivity does not necessarily correspond to a low false-positive rate (FPR). This observation underscores the lack of standardization in performance metrics across seizure prediction research, as each investigation reports distinct combinations of indicators, thereby complicating inter-study comparison and interpretability.

Among the 13 sEEG-based SVM studies [97,99,100,102,103,105,106,107,108,109,110,111,113,114], only two explicitly reported FPR values. Specifically, Kill et al. [106] achieved a sensitivity of 96.11% with an FPR of 0.48 h⁻¹, enabling prediction approximately one hour before seizure onset. Conversely, Qureshi and Kaleem [103] reported a sensitivity of 94.9% with a markedly lower FPR of 0.138 h⁻¹, corresponding to 30 min of SPH. These findings indicate that higher sensitivity levels may occur at the expense of increased false alarms, emphasizing the necessity for harmonized reporting standards and balanced evaluation frameworks that jointly assess sensitivity and FPR to ensure clinical applicability and reproducibility.

Furthermore, five additional investigations employed ECG signals, either alone or in combination with other modalities such as EEG, HRV, respiratory activity, or ACC [98,99,101,104,112], to predict seizure onset. The corresponding prediction intervals extended from several seconds to 24 h. It is noteworthy that longer SPHs were generally associated with a decline in performance metrics; for instance, Ding et al. [101] reported a sensitivity of 64.4% for a 24 h prediction. Although this performance represents a limited detection capability, unsuitable for clinical implementation in high-risk patients, it constitutes a valuable contribution in terms of non-invasiveness in data acquisition, given the exclusive use of ECG, respiration, and ACC signals. The following subsection provides a focused discussion of some SVM-based studies on seizure prediction.

For example, Giannakakis et al. [112] utilized an SVM combined with 18 nonlinear methods, including time-domain features (e.g., envelope, mean heart rate, standard deviation and mean of RR intervals) and frequency-domain features (e.g., power in high- and low-frequency bands), to predict seizures using HRV from ECG signals. They achieved an accuracy of 77.1%, predicting seizures 21.8 s prior to their onset. Similarly, Perez-Sanchez et al. [113] applied a fractal dimension and discrete wavelet transform (DWT) in combination with an SVM for predicting an epileptic seizure by means of sEEG signals. An accuracy of 93.3% in forecasting a seizure 30 min before its onset is reported by the authors. On the other hand, Saadoon et al. [100] proposed a highly effective hybrid framework combining a pre-trained EfficientNet-B0 model for deep feature extraction with an ensemble of six SVM classifiers. A voting mechanism was used to aggregate the predictions from the SVMs. This approach, tested on the CHB-MIT dataset using EEG signals segmented into 10 s windows, achieved outstanding accuracies of 96.12%, 94.89%, and 94.21% for predicting seizures 10, 20, and 30 min before onset, respectively. The sensitivity values were similarly high at 95.21%, 93.98%, and 93.55%. This work demonstrates the potent synergy between advanced deep learning feature extractors and robust classical classifiers like SVM.

4.1.2. K-Nearest Neighbors (KNN)

Among ML techniques, k-NN appeared only sporadically, with three studies during the review period employing k-NN either as a primary classifier or as a point of reference [105,106,115]. Of these, two investigations used sEEG chiefly to contrast k-NN against other traditional ML or DL models, while one study implemented k-NN as the principal algorithm on ECG signals [115], reporting 20 min of SPH and an accuracy of 93.25% as the sole quantitative metric. The absence of sensitivity, specificity, and especially FPR in this report constrains clinical interpretability, as accuracy alone does not characterize the alarm burden or the reliability of preictal detection within a defined time-to-warning window. Taken together, these findings are consistent with a declining emphasis on classical ML approaches for seizure prediction in the recent literature, as deep learning methods increasingly dominate and underscore the need for standardized reporting (time-to-warning, sensitivity, specificity, and FPR per hour) to enable robust cross-study comparisons and credible assessment of clinical viability.

4.1.3. Decision Tree (DT)

Among ML classifiers, DT-based approaches (including single trees and ensemble derivatives such as Random Forest, Gradient Boosting, XGBoost, and LightGBM) alone or in conjunction with other techniques were identified in 13 studies [97,102,105,116,117,118,119,120,121,122,123,124,125] during the review period. Of these 13 investigations, 12 studies employed sEEG as the primary signal with only sparse mentions of auxiliary physiological channels, and only one employed iEEG [120] was documented in this subset. Reported SPHs for DT models spanned from seconds to approximately 75 min, reflecting heterogeneous preictal definitions and validation protocols. Sensitivity was explicitly stated in several reports, but with variable magnitude between 83 and 99%, whereas FPR was only declared in the study of Coşgun & Çelebi [124], who reported an FPR = 0.041 h⁻¹ for predicted a seizure 33.23 min before onset. This singular dataset limits the balanced assessment of the alarm in DT-based algorithms. Notably, Chen et al. [122], focusing on sEEG, applied DT algorithms to classify epileptic states and achieved 87.9% accuracy with a 30 s lead time before seizure onset, exemplifying short-horizon feasibility under controlled conditions. This allowed for inter-study comparison and credible evaluation of clinical readiness.

It is important to mention that throughout this section, focusing on the area of seizure prediction, several researchers have proposed different signal processing techniques that can predict an epilepsy episode seconds or hours prior to its occurrence. However, it is essential to compare their efficacy using concepts such as true positive, false negative, false positive, and true negative, as these are the theoretical basis for defining sensitivity, specificity, accuracy, FPR, and other related metrics [126]. To provide a deeper understanding of the prediction time before a seizure occurs, as well as the advantages and opportunities of research of various ML-based approaches for seizure prediction,

Table 5. Main advantages and opportunities of ML-based methodologies presented for seizure prediction.

ML-Based Algorithms	Proposal Advantages	Opportunities of Research	Time Prediction	Application
SVM	The application reduces noise and isolates critical features, preserving essential frequency components associated with seizure activity. Feature selection and dimensionality reduction streamline the classification process, enabling the model to handle complex, high-dimensional data while maintaining computational efficiency. The classifier employed is robust and offers strong generalization capabilities.	The application is computationally demanding, posing challenges for real-time seizure prediction. The reliance on extensive pre-processing and feature extraction steps introduces the potential for overfitting.	23 min y 36 seg	Altaf et al. [114]
KNN	This application allows for a detailed examination of time and frequency characteristics, providing a comprehensive understanding of the underlying patterns preceding a seizure. The method used for feature selection ensures that the most relevant and statistically significant features are retained, enhancing the model’s ability to identify seizure precursors accurately. Additionally, the classification method employed is well-suited for recognizing patterns in the data, enabling reliable seizure prediction that is essential for timely intervention.	This approach is computationally demanding, potentially limiting its application in real-time scenarios where quick processing is essential. While useful, the focus on specific features may lead to the omission of other relevant patterns, reducing the model’s overall robustness. Additionally, the approach may be sensitive to parameter selection and could struggle with handling imbalanced datasets, impacting the accuracy and reliability of seizure predictions. Reliance on distance measures in the classification process can also present challenges.	20 min	Perez-Sanchez et al. [115]
Automatic threshold	This application effectively isolates relevant features. The feature selection process is robust, focusing on statistically significant features that enhance the model’s predictive accuracy. The classification strategy is straightforward and efficient, allowing for quick and reliable seizure prediction.	This may be limited by its reliance on specific features that might not fully capture the complexity of the pre-seizure state, potentially reducing predictive robustness. The application may be vulnerable to noise and artifacts in the data, which could impact the accuracy of its predictions. While the simplicity of the classification strategy is beneficial for efficiency, it may result in a less nuanced analysis, potentially leading to a higher rate of false positives or missed seizures.	Variant up to 60 maximum minutes	Mbarek et al. [127]
DT	This application effectively isolates relevant frequency bands from data. The feature selection process is rigorous, ensuring that only the most statistically significant features are included, which helps reduce the data’s dimensionality while maintaining predictive accuracy. The classification strategy is interpretable and straightforward, allowing for transparent decision making, which is critical in a clinical setting.	The approach may be limited by its sensitivity to noise and artifacts in the data, which could impact the reliability of the predictions. Although the classification strategy is easily interpretable, it might not fully represent the complexity of seizure precursors, leading to a higher risk of false positives or missed predictions.	8 h	Saboo et al. [123]

presents a summary of the different ML algorithms employed for this task.

4.2. Epileptic Seizure Prediction Algorithms Using DL Methods

4.2.1. CNNs

Among DL techniques, CNNs were the most frequently employed architecture in the 2020–2025 window, with 36 studies implementing either standard CNNs or deep variants such as 1D-, 2D-, and 3D-CNN, DenseNet, ResNet, VGG, and Temporal CNN [102,107,111,128,129,130,131,132,133,134,135,136,137,138,139,140,141,142,143,144,145,146,147,148,149,150,151,152,153,154,155,156,157,158,159]. This breadth reflects the field’s convergence toward convolutional feature extraction for preictal pattern modeling while allowing architectural flexibility to accommodate diverse datasets and experimental protocols.

Of the 36 CNN investigations, 33 studies utilized sEEG as the primary signal, either alone or in combination with auxiliary modalities such as ECG or iEEG [102,107,111,128,129,130,131,132,133,134,135,136,137,138,139,140,141,142,143,144,145,146,147,148,149,150,151,152,153,156,157]. Reported SPHs for sEEG-based CNNs ranged from a few seconds to several hours. Within this subset, sensitivities spanned 80% to 99%, a variability that likely reflects differences in preictal labeling, validation strategy, and dataset characteristics. FPR was reported in some studies, but where available, it ranged from 0.017 to 0.07 h⁻¹, indicating a low alarm burden compatible with future clinical monitoring. Nevertheless, the affirmation that CNN-based methods always give low FPR values as a result is incorrect due to the limited and non-heterogeneous FPR reporting in studies, which precludes definitive cross-study comparisons and complicates assessments; additionally, this variation in range implies that the result of low FPR does not depend on the classifier employed to predict alone. On the other hand, a smaller subset of five CNN studies employed iEEG alone or in combination with sEEG [136,152,153,154,155]. These works reported SPHs in the 5 to 50 min range, with sensitivities between approximately 87% and 100% and low FPR where available, with a minimum of 0.25 h⁻¹. The intracranial setting, typically characterized by reduced artifact contamination and higher signal-to-noise ratio, is expected to favor the extraction of localized preictal dynamics; however, many iEEG studies omitted explicit FPR and occasionally other key metrics, limiting head-to-head comparability with sEEG-based approaches. CNNs and their variants constitute the dominant paradigm for seizure prediction in the recent literature. The sEEG studies offer broad validation across public datasets and a rich spectrum of architecture and preprocessing pipelines, while the iEEG subset demonstrates promising sensitivities and feasible warning times under cleaner intracranial conditions. The following provides a focused discussion of some CNN-based studies on seizure prediction.

Shen et al. [156] used the STFT in combination with Google Net CNN to predict epileptic seizures, employing sEEG data provided by the CHB-MIT database. They reported that their proposal can predict seizures 9.85 s before onset, achieving notable results with an accuracy of 97.74%. Mohankumar et al. [157] created a combined model that merges CNN and Long Short-Term Memory (LSTM) layers to effectively capture spatial characteristics and temporal relationships in EEG segments. Their improved model reached an accuracy of 98.5%, a recall of 98.7%, and an F1 score of 98.5% for a prediction of 5 s. Liu et al. [133] also used CHB-MIT data, combining it with nonlinear features (i.e., Higuchi fractal dimension, sample entropy, approximate entropy, and fuzzy entropy). The mean Max-relevance and Min-redundancy method was used to select the best features. Next, the authors classified the epileptic features by employing fused multidimensional structures, which are a combination of a pseudo-3D CNN and bidirectional convolutional LSTM (BiLSTM) 3D. With these combined techniques, the authors specified that an epileptic seizure can be predicted 15 min before onset with an accuracy of 98.13%, precision of 98.30%, sensitivity of 98.03%, and specificity of 98.23%. Meanwhile, Xu et al. [138] employed a dynamic functional connectivity neural network to predict seizures, achieving 88% accuracy and 87.3% specificity for predicting a seizure with an average prediction time of 0.207 s.

4.2.2. Recurrent Neural Networks (RNNs)

Among DL techniques, RNNs, including LSTM, GRU, and bidirectional variants, were identified in 30 studies during the review period [46,47,97,105,135,136,137,140,142,148,150,151,160,161,162,163,164,165,166,167,168,169,170,171,172,173,174,175,176]. This family was primarily applied to model temporal dependencies in EEG dynamics for preictal detection. Of the 30 investigations, 26 studies explicitly used sEEG as the main modality in combination with iEEG in some cases. Within these studies, sensitivities ranged from 41% to 99.34%, while accuracies ranged from 41% to 98.84%, and specificities from 41% to 99.34%. The SPHs presented in the records span from minutes to hours. In this search, only Lu et al. [148] reported an FPR of 0.017 h⁻¹ for predicting a seizure 30 min before onset with a sensitivity of 98.40%; consequently, although RNNs can achieve high sensitivity in certain configurations, the lack of FPR reporting limits inter-study comparability and clinical interpretability. On the other hand, four studies employed iEEG alone or in conjunction with sEEG report the times of prediction and metrics similar to the sEEG studies. As this aspect is of special interest, the paper realized by Meisel et al. [47] employed wristbands that continuously record physiological parameters, including EDA, body temperature, blood volume pulse, and actigraphy to predict a seizure with 10 to 50 min before a seizure, and they identified better-than-chance predictability in 43% of the patients, opening a door to the possibility that these signals can be used in conjunction with AI techniques to predict epileptic seizures without the need for a device that limits the patient’s mobility. Next, more studies that employed RNNs variants are discussed.

Salhi and Namma [176] compared an LSTM and a deep feedforward network (DFN) architecture in their investigation to predict seizures from 5 to 50 min before onset using sEEG signals. They reported that a DFN is more accurate for predicting an epileptic seizure prior to its onset. Using a DFN, they achieved the following accuracies for prediction times of 50, 30, 10, and 5 min: 94.83%, 96.98%, 98.99%, and 99.49%, respectively. With the LSTM, the results were 82.66%, 88.88%, 97.26%, and 98.62% accuracy in predicting a seizure 50, 30, 10, and 5 min before onset. Zhang et al. [174] employed sEEG data in combination with the DWT and nonlinear features, and a classifier based on multiclass feature fusion with a CNN-gated recurrent unit-attention mechanism for seizure prediction. Using this combination of techniques, the authors predicted a seizure 5 min before onset with 95.16% accuracy, 95.47% sensitivity, and 94.93% specificity. Lee et al. [175] utilized sEEG data from the CHB-MIT and Seoul National University Hospital (SNUH) datasets. In their work, the authors used the STFT to convert data into spectrogram images, followed by a residual neural network in combination with LSTM to extract image features and classify them into ictal or preictal states. They reported that their method could predict an epileptic seizure with 91.9% accuracy, 89.6% sensitivity, and 94.2% specificity, 15 min before seizure onset using the CHB-MIT dataset. For the SNUH database, they achieved the following results for predicting a seizure 30 min before onset: 78.24% accuracy, 87.3% sensitivity, and 82.78% specificity.

4.2.3. Transformer-Based Methods (TBMs)

TBMs, including Vision Transformer (ViT) and related attention-based models, were identified in 24 studies [52,130,132,133,135,145,150,152,153,177,178,179,180,181,182,183,184,185,186,187,188,189,190,191]. Their adoption reflects growing interest in long-range temporal/context modeling for preictal state identification. Of these, 23 studies employed sEEG as the primary signal or a combination with iEEG. The metrics reported sensitivities ranged from 91% to 99%, accuracy from 97% to 99%, and specificities from 97% to 99%. These are for the time of prediction in the range of minutes. The FPR for the work that was reported varied between 0.0064 and 0.18 h⁻¹. These results are promising due to their low false alarm rate. As for the two iEEG works, alone or in conjunction with sEEG, these reported prediction times range from a few seconds to a few minutes, with sensitivity around 90%. Consequently, despite the strong discriminative signals, the alarm burden and time-to-warning cannot be evaluated for all TBMs, limiting clinical interpretability and cross-study synthesis. Standardized reporting that jointly specifies the warning window, sensitivity/accuracy/specificity, and FPR (h⁻¹) are essential for establishing the clinical readiness of TBMs seizure prediction systems. Next, some studies that employed TBMs variants are discussed.

Hussein et al. [135] used sEEG data from the CHB-MIT and KAES seizure prediction datasets. The authors employed a continuous wavelet transformer to convert sEEG signals into scalograms, which were fed into the multi-channel vision transformer. In particular, the authors observed that the proposed model predicted seizures with an accuracy of 99.80%, a sensitivity of 99.76%, and a specificity of 99.92% using sEEG data 10 min before seizure onset. Conversely, they noted that the model achieved 90.28% accuracy for iEEG data 10 min before seizure onset. Yan et al. [178] used sEEG data from the CHB-MIT database. They applied an STFT for estimating the time-frequency features, and three transformer tower models were utilized for fusing and classifying the features of the sEEG signals. With their approach, the authors achieved a seizure prediction sensitivity of 96.01% and a specificity of 96.23%, predicting seizures up to 3 min before onset. Yuan et al. [186] used sEEG data from the CHB-MIT database. They applied an STFT to convert sEEG signals into time-frequency matrices. The hybrid model, which combines Dense Net for local feature extraction and a vision transformer for global pattern recognition, successfully predicted seizures. Their model achieved a prediction accuracy of 93.65% and a sensitivity of 93.56%, with a 3 min of SPH. Ma et al. [184] proposed a parallel double-branch fusion network. This model effectively merges the strengths of transformers and CNNs to capture both long- and short-term dependencies in EEG signals. Using Mel-frequency cepstral coefficients as input features, their approach incorporates a dedicated fusion module to evaluate the significance of time, frequency, and channel information within the transformer branch. Evaluated through rigorous leave-one-out cross-validation on the CHB-MIT database, their method achieved high performance, with an accuracy of 95.76%, sensitivity of 95.81%, specificity of 95.71%, and precision of 95.71% for predicting seizures 60 min before onset. Similarly, Dong et al. [162] introduced a multi-scale spatio-temporal attention network based on a Swin transformer backbone to extract hierarchical spatial features from EEG spectrograms. Their architecture includes a spatial pyramid module for multi-scale feature extraction and a sequential aggregation module using the LSTM to model temporal dependencies. To address class imbalance, the network was trained with a combination of triplet loss and focal loss. The authors demonstrated that using the CHB-MIT database for validating their proposal, it can reach a sensitivity of 96.27% for 60 min seizure prediction.

Table 6. Summary of DL-based methodologies for seizure prediction.

Classifier	Advantages	Opportunity of Research	Time Prediction	Application
Pseudo-3D CNN–BiLSTM 3D, Attention3D	This application captures the signal’s complexity and irregularity through advanced entropy measures and fractal analysis, providing a rich set of highly informative features for seizure prediction. The feature selection process ensures that the most relevant and least redundant features are retained, optimizing the model’s predictive power. The classification model combines spatial and temporal information with attention mechanisms. This complex architecture is particularly well-suited for capturing the nuanced dynamics of seizure development, offering high predictive accuracy and robustness.	The approach is computationally intensive, particularly in the feature extraction and classification stages, which may pose challenges for real-time applications. While powerful, the complexity of the model increases the risk of overfitting, particularly if not carefully tuned and validated across diverse patient datasets. The model’s sensitivity to variations in the input data could lead to a higher incidence of false positives or missed predictions.	15 min	Liu et al. [158]
BiLSTM	This application offers a comprehensive data analysis by combining linear and nonlinear features, capturing a wide range of signal characteristics relevant to seizure prediction. Advanced feature fusion techniques enhance the representation of spatial features, leading to a more robust model that can accurately identify pre-seizure patterns. Integrating an attention mechanism further refines the feature selection process, enabling the model to focus on the most critical aspects of the data, thereby improving predictive accuracy. The classification model is well-suited for handling temporal dependencies in the signals.	This approach’s complexity may pose challenges regarding computational demands, particularly during the feature extraction and classification stages, which could limit its applicability in real-time scenarios. While enhancing predictive performance, advanced fusion techniques and attention mechanisms can lead to overfitting, especially with limited data, if not properly managed. Additionally, the intricate nature of the model reduces interpretability.	40 min	Ahmad et al. [135]
1-D CNN	This application effectively enhances the model’s stability and convergence by normalizing the input data during pre-processing, which leads to improved training efficiency and predictive accuracy. The classification model employs a one-dimensional CNN. Combining techniques allows for precise and reliable detection of relevant signal features, enabling accurate predictions. The model’s ability to handle large amounts of data and detect subtle changes in the signal.	The approach may face challenges related to computational resource requirements, particularly during the DL model’s training phase, which could limit real-time applicability. Additionally, while batch normalization improves training stability, it may not fully address the variability and noise inherent in data, potentially affecting the model’s performance. Though effective for sequential data, the one-dimensional CNN might struggle with capturing more complex temporal patterns and interactions in the signals, which could lead to reduced predictive performance.	60 min	Saeizadeh et al. [159]
Parallel Dual-Branch Fusion Network	This application leverages sophisticated feature extraction and classification methods to provide a detailed analysis of signals. Using an advanced fusion network for feature selection and classification enhances the model’s ability to integrate and analyze multiple aspects of the data simultaneously, leading to improved predictive accuracy. The parallel architecture allows for the efficient processing of complex signals, making it well-suited for handling large datasets and diverse patient profiles, thereby increasing the model’s robustness and generalizability.	The approach’s complexity and reliance in advanced neural network architectures can lead to significant computational demands, potentially limiting its feasibility in real-time or resource-constrained environments. The intricate nature of the model also increases the risk of overfitting, mainly if the training data are not sufficiently diverse or abundant.	60 min	Ma et al. [184]
Transformer deep model	This application used correlation-based feature extraction, allowing for the identification of strong, relevant signal patterns indicative of pre-seizure states. The integration of positional encoding enhances the model’s ability to capture temporal dependencies and contextual information within the data. With its advanced architecture, the classification model can learn complex patterns and long-range dependencies in the data, resulting in high predictive accuracy.	The approach involves significant computational complexity, particularly with DL models and large-scale data processing, which can pose challenges for real-time implementation. The model’s complexity also increases the risk of overfitting, especially if the training data is not sufficiently comprehensive or diverse. The advanced architecture may also reduce interpretability.	60 min	Lih et al. [191]
CNN	The application effectively streamlines the data by pre-processing it to focus on the most relevant information, enhancing the model’s ability to detect pre-seizure patterns with higher accuracy. The use of DL classification models allows for extracting intricate features and patterns from the processed data, which is crucial for identifying subtle changes indicative of impending seizures. The reduced data volume after pre-processing helps accelerate the training process and improves computational efficiency.	In this approach, the pre-processing steps may result in the loss of potentially important information, such as filtering, down-sampling, and undersampling can omit significant signal details critical for accurate seizure prediction. Additionally, the DL model’s complexity may lead to high computational demands, potentially hindering real-time applications. While effective at feature extraction, the CNN model may struggle with generalizing across diverse patient data or varying seizure types, potentially leading to reduced performance in different scenarios.	15 min	Saeizadeh et al. [192]

provides a summary of the different DL-based methodologies used for seizure prediction, highlighting their advantages and opportunity of research. Additionally, it offers a deeper understanding of the prediction time before a seizure occurs, as well as the strengths and limitations of the approaches employed for this task.

5. Future Perspectives

Since the 1970s, seizure prediction research has progressed from short, single-modality recordings to richer, longer-term datasets. Looking ahead, clinically useful systems will require multimodal acquisition (integrating iEEG/sEEG with physiological streams, e.g., ECG, PPG, to improve robustness in real-world settings) together with larger, well-annotated cohorts and harmonized labeling to reflect patient heterogeneity.

Methodologically, advances should prioritize patient-independent, multi-site evaluation with preregistered protocols, fixed SPH, and harmonized reporting of sensitivity, specificity/accuracy, and FPR/h. To curb overfitting, studies must be used to report methods to prevent overfitting. For personalization, transfer learning, meta-learning, and ensembles are promising for inter-patient variability and faster adaptation, with published work already indicating improved stability and efficiency [129,173].

Deployment demands resource-aware models suitable for wearables and embedded platforms (tinyML) while meeting constraints on latency, power, and on-device privacy directions already explored in adjacent domains such as safety systems, construction, and procedural healthcare [193,194,195,196]. Equally important are human factors (alarm-fatigue minimization, user adherence), interoperability with clinical and implantable ecosystems, and regulatory-grade documentation (risk management, software lifecycle, cybersecurity) supported by prospective, intended-use studies and post-market performance monitoring. In addition, to align with recent regulatory expectations for transparent and reproducible IA workflows, studies should explicitly document subject-level training/validation/test partitions and leakage-prevention safeguards (e.g., subject-disjoint splits, nested model selection), and report learning curves and variance across repeated runs to assess overfitting.

In summary, the path from methodologies to clinic application needs the following: (1) validated generalization across centers and devices, (2) transparent, comparable metrics including FPR/h for a pre-specified horizon, (3) rapid, safe personalization, and (4) efficient, secure execution on patient-worn hardware. Standardized reporting of sensitivity, specificity/accuracy, FPR/h, and the predefined SPH will further enable fair, regulation-ready comparisons across studies. Meeting these milestones will translate technical progress into reliable, patient-centered tools that are feasible for clinical application.

6. Conclusions

From the review carried out in this work, several conclusions about methodologies that employed ML/DL for epileptic seizure prediction have been noted in general.

6.1. Advantages

• The observed trend is for models to learn nonlinear and multiscale structures directly from signals, which reduces the reliance on manually created features and improves discriminative performance.
• The architectures discussed in this review jointly capture spatial/spectral structure and temporal dependencies, improving the detection of subtle preictal patterns.
• It is observed that integrating sEEG/iEEG with other physiological signals (e.g., ECG, PPG) improves the robustness and validity of the application in real-world conditions.

6.2. Opportunity Areas

• Future studies could more systematically characterize out-of-distribution performance by prioritizing patient-independent, multi-site evaluations.
• The seizure prediction studies would benefit from a core, consistently reported metric set so that results can be compared fairly across studies and settings.
• To strengthen the methodological rigor of the studies centered in seizure epilepsy prediction, authors that reported works in this area might consider subject-disjoint splits, nested model selection, and clear documentation of preprocessing and hyperparameter search, together with an explicit description of the techniques used to mitigate overfitting.
• Providing learning curves, variance across repeated-run studies would help disentangle architectural contributions from data-centric effects and clarify how models perform in the area.
• Routine disclosure of computing budgets, memory, and training/inference time would enable more transparent and equitable comparisons among methods and facilitate deployment planning.
• Due to clinical utility hinges on timely warnings, it would be helpful for all studies with “seizure prediction” in their title to consistently report the SPH alongside performance metrics to avoid mixing with works that realized classification or detection of epileptic seizure models, increasingly learn nonlinear, multiscale representations directly from bio signals, and jointly model spatial/spectral and temporal structure, yielding stronger discrimination and better detection of subtle preictal patterns.
• Multimodal integration combining sEEG/iEEG with physiological streams (e.g., ECG, PPG) enhances robustness and ecological validity, supporting performance in real-world conditions.

Based on the findings of this review, the proposed models in the timeframe of 2020–2025 increasingly learn nonlinear, multiscale representations directly from bio signals and jointly model spatial/spectral and temporal structures. This approach yields stronger discrimination and better detection of subtle preictal patterns. Furthermore, the integration of multimodal data, combining sEEG/iEEG with physiological streams (e.g., ECG, HRV), enhances robustness. However, to further improve the field, it would be advisable to adopt the following measures: when discussing early prediction, studies should demonstrate the prediction time, report computational cost and memory footprint, include clinically important metrics such as the FPR, detail cross-validation methods, and describe techniques to avoid overfitting.