Machine and Deep Learning-Based Seizure Prediction: A Scoping Review on the Use of Temporal and Spectral Features

Abstract

Epileptic seizures result from abnormal brain activity, posing significant health risks due to their sudden and unpredictable nature. Accurate seizure prediction is crucial for improving patient outcomes and enabling timely interventions. Recent advancements in artificial intelligence (AI), particularly in machine learning (ML) and deep learning (DL), have significantly enhanced seizure detection and prediction. This review provides a comprehensive overview of seizure prediction models that integrate temporal and spectral features as inputs or enhanced representations for ML and DL models. Emphasizing convolutional neural networks (CNNs) and other deep architectures, we explore the role of time-domain and frequency-domain features, such as wavelet transforms, short-time Fourier transforms, and spectrogram representations, in improving model performance. Additionally, the review discusses common challenges, including feature interpretability, generalizability across datasets, and computational efficiency. By highlighting recent advancements and limitations, this study provides insights into optimizing spectral and temporal feature integration for seizure prediction, paving the way for more robust and clinically viable AI-based solutions.

1. Introduction

Epilepsy is a chronic neurological disorder that affects millions of people worldwide and is characterized by recurrent and unpredictable seizures [1]. Although it can affect individuals at any age, its prevalence is higher among newborns and the elderly. Seizures are typically caused by sudden disruptions in the brain’s electrical activity, often originating from the frontal cortex and potentially spreading throughout the body. In severe cases, individuals may experience hundreds of seizures per day, leading to progressive cognitive and neural damage. Thus, early and accurate seizure prediction is essential to improving patient outcomes and quality of life [2]. Among the available neuroimaging modalities, the electroencephalogram (EEG) is the most widely used for seizure detection and prediction due to its low cost, portability, and high temporal resolution [3]. EEG captures brain activity by recording voltage fluctuations resulting from ionic currents in cortical neurons. These signals typically reflect patterns in both the time domain and frequency domain, which are critical for identifying abnormal brain states associated with epilepsy [4]. However, EEG data are often high-dimensional, noisy, and affected by various artifacts such as muscle tremors, electrode shifts, and environmental interference, making accurate interpretation a significant challenge. To overcome these limitations, artificial intelligence (AI) techniques, particularly machine learning (ML) and deep learning (DL), have been increasingly applied to automate and enhance seizure prediction from EEG signals [5,6,7]. These methods can assist clinicians by offering fast, consistent, and scalable diagnostic tools. Classical ML models such as decision trees [8], support vector machines (SVMs) [9,10], k-nearest neighbors (KNNs) [11], and random forests [12] were among the first approaches used to classify epileptic events based on engineered features extracted from EEG. More recently, DL models have demonstrated superior performance due to their ability to automatically extract complex patterns from raw or minimally preprocessed signals. Recurrent neural networks (RNNs), especially long short-term memory (LSTM) networks, are particularly effective in capturing temporal dependencies in EEG signals, making them suitable for modeling the dynamic nature of seizures [13,14,15]. Convolutional neural networks (CNNs), on the other hand, are adept at extracting spatial and spectral features from EEG spectrograms and have been successfully employed in seizure recognition tasks [16]. Despite the growing volume of research, the role and impact of temporal (e.g., statistical time-domain features, autocorrelation) and spectral (e.g., power spectral density, wavelet coefficients) features remain underexplored in many reviews [17,18,19,20]. These two categories of features are essential for understanding seizure dynamics, as they capture the timing and frequency characteristics of abnormal brain activity. Temporal features offer insight into signal variation over time, while spectral features decompose EEG into frequency bands that are often linked with seizure onset and propagation. By focusing on the interplay between ML/DL models and temporal/spectral EEG features, this review aims to provide a comprehensive and comparative understanding of how different algorithms leverage these features to predict seizures effectively.

This scoping review categorizes existing approaches based on the type of learning model and the nature of the features used, as shown in Figure 1, identifies key trends and limitations, and discusses future directions. In doing so, it seeks to highlight not only the strengths of current methods but also the potential for innovation in seizure prediction based on feature representation and the learning strategy. The primary objectives of this study are threefold:

• To systematically analyze deep learning models used for seizure prediction based on the nature of the temporal (e.g., amplitude, entropy, statistical time features) and spectral (e.g., Fourier, wavelet, or filter bank-based) features they utilize.
• To identify and critically assess common limitations in these approaches, including issues related to generalizability across datasets, the robustness of feature extraction methods, and computational complexity in clinical environments.
• To offer practical insights into deep feature-based classification strategies, highlighting how the integration of temporal and spectral representations can improve predictive accuracy and inform future directions in clinically viable seizure detection systems.

2. Related Reviews and Current Contribution

Several review articles have addressed the problem of epileptic seizure prediction (ESP) using signal processing, ML, and DL techniques. However, these studies differ in scope, emphasis, and depth, and most have not systematically analyzed the role of deep feature learning from temporal and spectral EEG characteristics.

Mosheni et al. [17] compared various traditional signal processing methods for seizure detection, including nonlinear time series analysis, entropy measures, and spectral transforms, finding that entropy features could achieve up to 100% accuracy on specific datasets. However, the generalizability of these results across diverse real-world EEG recordings remains questionable. Kamini et al. [18] provided a comprehensive overview of ML-based seizure prediction techniques, presenting available code, features, and evaluation metrics, but did not delve into deep learning architectures or the benefits of learned features over hand-engineered ones. Xu et al. [19] focused on recent advancements in portable ESP systems and signal processing pipelines, emphasizing the importance of each stage from preprocessing to classification. While their study introduces trends and challenges in ESP, it does not explore the performance or architecture of DL models such as CNNs and LSTMs. Shoeibi et al. [5] offered a broad review of automatic seizure detection using DL methods across neuroimaging modalities, including EEG and MRI. Although valuable, their work emphasized DL-based rehabilitation and cloud-based systems more than EEG-specific feature extraction for prediction tasks. Craik et al. [20] presented an in-depth review of deep learning methods applied to EEG classification tasks. Their work highlighted the shift from traditional handcrafted feature extraction to end-to-end learning approaches and emphasized key challenges in model interpretability, generalization, and performance across different EEG datasets. The following describes how this review fills this gap:

• This review categorizes seizure prediction approaches based on the type of deep features extracted, distinguishing between time-domain, frequency-domain, and hybrid feature representations.
• Unlike prior reviews that assess generic DL models, this study emphasizes CNN-based architectures, which have shown promising results for capturing spatial and temporal EEG patterns.
• Finally, this study outlines potential future research directions, particularly focusing on self-supervised learning, Transformer-based models, and interpretable DL approaches for seizure prediction.

3. Methodology of Literature Search and Selection

This scoping review explores DL-based approaches for epileptic seizure prediction, emphasizing how temporal and spectral features extracted from EEG signals are represented across different model architectures. The aim is to synthesize recent advancements, compare feature representation strategies, and identify methodological gaps that could inform future clinically deployable systems.

3.1. Search Strategy and Data Sources

A comprehensive literature search was conducted across five major academic databases: PubMed, Scopus, IEEE Xplore, Web of Science, and Google Scholar. The search covered peer-reviewed journal articles and conference proceedings published from January 2000 to April 2025.

Three thematic categories guided the search query design:

• Disorder-related terms: “epilepsy”, “epileptic seizures”, “seizure detection”, and “seizure prediction”;
• Modeling techniques: “deep learning”, “CNN”, “convolutional neural network”, “LSTM”, “RNN”, “Transformer”, “graph neural network”, and “self-supervised learning”;
• Feature extraction approaches: “temporal features”, “spectral features”, “wavelet transform”, “STFT”, “Fourier transform”, “EEG spectrogram”, and “power spectral density”.

Boolean operators (AND/OR) and wildcard symbols were used to refine the search and accommodate terminology variation. A full search string for each database is documented in Appendix A.

3.2. Inclusion and Exclusion Criteria

The following inclusion criteria were applied:

• The use of deep learning architectures (e.g., CNN, LSTM, RNN, Transformer) for seizure prediction (not merely detection);
• The use of EEG as the primary input modality;
• The incorporation of temporal, spectral, or hybrid features, either through explicit feature engineering or learned representations;
• Having reported the details of the experimental setup, including

  −

  Dataset source (e.g., CHB-MIT, Bonn, TUH);

  −

  Preprocessing steps (e.g., filtering, segmentation);

  −

  Evaluation strategy (e.g., cross-validation, patient-specific testing);

  −

  Performance metrics (e.g., accuracy, sensitivity, AUC).

• Studies that offer a comparative analysis of features or models.

The exclusion criteria were

• Studies focused solely on seizure detection without predictive modeling;
• Non-EEG-based studies or multimodal approaches where EEG was not the core signal;
• Articles lacking sufficient methodological transparency or evaluation rigor;
• Review papers, editorials, non-English texts, or studies without accessible full-text.

3.3. Screening Workflow and Reviewer Agreement

The study selection followed the PRISMA-ScR guidelines. As illustrated in Appendix A, an initial pool of 3164 records was retrieved. After duplicate removal, titles and abstracts were screened for relevance. The remaining full-text articles were independently reviewed by two researchers using the above criteria.

Discrepancies were resolved through discussion, and a third reviewer was consulted for unresolved conflicts. Inter-reviewer agreement was calculated using Cohen’s Kappa, resulting in a value of 0.83, indicating strong agreement.

A final set of 74 studies was included for detailed analysis.

3.4. Data Charting and Extraction

A structured data extraction template was developed and piloted. The following data items were charted:

• Study metadata: authorship, year, publication type;
• Model type and architecture;
• EEG features (temporal, spectral, hybrid);
• Dataset(s) used and data size;
• Preprocessing pipeline;
• Evaluation metrics and validation strategy;
• Key findings and comparative insights.

This information was compiled in tabular format and used to support the narrative synthesis.

3.5. Classification and Synthesis Strategy

Studies were grouped along two axes: feature domain (temporal, spectral, hybrid) and model architecture (CNN, RNN/LSTM, Transformer, hybrid). This matrix structure enables

• Horizontal comparisons: the performance of different architectures using similar features;
• Vertical comparisons: the impact of different feature types within the same model family.

Additionally, studies offering explicit comparisons between learned and engineered features were highlighted to examine interpretability, generalizability, and clinical utility.

4. DL Models

In an AI-powered seizure prediction system using EEG input (Figure 2), traditional machine learning (ML) models, such as support vector machines (SVMs), decision trees (DTs), and random forests (RFs), require a preprocessing stage where raw EEG signals are transformed into engineered features (e.g., statistical, spectral, or nonlinear). These models then classify EEG segments based on these manually extracted features. While relatively efficient, they are limited by their dependence on an expert-driven feature design.

In contrast, deep learning (DL) approaches [18,19,20], especially convolutional neural networks (CNNs) and recurrent neural networks (RNNs), including long short-term memory (LSTM) variants, have revolutionized seizure prediction by directly learning from raw EEG data. CNNs are widely adopted due to their capacity to capture spatial and spectral patterns [21], particularly when EEG signals are represented as time–frequency spectrograms. CNN architectures typically include convolutional, pooling, and fully connected layers, allowing for automatic feature extraction and efficient classification between seizure and non-seizure states.

LSTMs, on the other hand, are tailored to sequential data and are effective at modeling temporal dependencies across EEG time windows, making them ideal for predicting transitions into ictal states. Hybrid models combining CNNs and LSTMs have also been proposed to leverage both spatial and temporal dynamics in EEG data.

5. ML Model with Time-Domain Input

Initial studies emphasized handcrafted temporal features such as peak-to-peak amplitude, duration, and spike frequency for seizure prediction [22,23,24,25]. These descriptors were often dataset-specific and lacked robustness across individuals. Srinivasan et al. [26] demonstrated the utility of Approximate Entropy (ApEn) with Elman and Probabilistic Neural Networks, achieving perfect classification accuracy on selected datasets. Ahammad et al. [27] extracted energy-based metrics (SEN, IQR, MAD) and achieved 84.2% accuracy using statistical classifiers on CHB-MIT.

Martis et al. [28] introduced Intrinsic Time-Scale Decomposition (ITD) to compute the sample entropy and fractal dimension, reaching 99% sensitivity and 99.5% specificity. Enhancements to entropy were proposed by Fadlallah et al. [29], who developed Weighted Permutation Entropy (WPE), incorporating both temporal order and amplitude. Yang et al. [30] used a 6 s EEG window to extract statistical features like skewness, kurtosis, and RMS, reporting 90.2% sensitivity with a low false prediction rate.

Li et al. [31] employed Scale-Dependent Lyapunov Exponents (SDLEs) with SVM and Random Forest classifiers, achieving 92.2% sensitivity and predicting seizures with up to a 0.12/h false positive rate. Brari [32] used a simplified classifier based on Correlation Dimension (CD), maintaining high accuracy with minimal computational load. Soomro et al. [33] utilized Canonical Correlation Analysis (CCA)-based features in a multilayer perceptron neural network (MLPNN), achieving 92.6% accuracy.

Other works applied statistical distributions and selection methods for robustness. Abbaszadeh et al. [34] evaluated Kruskal–Wallis and IQR features, while Memarian et al. [35] incorporated mutual information and minimum redundancy–-maximum relevance (mRMR) techniques in multimodal EEG contexts, achieving 95% accuracy.

With increasing computational power, raw EEG inputs have been used directly in DL models. Zhang et al. [36] applied wavelet-decomposed Common Spatial Pattern (CSP) features to CNNs, achieving 90.2% sensitivity and low false prediction rates. Wei et al. [37] used 3D CNNs to learn spatial–temporal EEG structures, surpassing 90% accuracy, while Xun et al. [38] implemented temporal autoencoders to reduce EEG into compressed embeddings with a 22.93% error rate. Yang et al. [39] reported 92.07% accuracy using a combination of CNN, LSTM, and self-attention. Transformer-based models such as EpilepsyNet [40] integrated temporal attention while reducing computational complexity (85% accuracy). Saadoon et al. [41] combined EfficientNet-B0 with an SVM to tackle inter-subject variability and achieved 96.12% accuracy, demonstrating the power of DL-based feature fusion.

Real-time EEG prediction frameworks are gaining traction. Qiu et al. [42] applied cloud- and edge-based DL pipelines to achieve 93.4% accuracy. Salafian et al. [43] maintained 91.4% accuracy in real-time contexts using federated learning and privacy-aware fuzzy models. Zhou et al. [44] optimized LMA-EEGNet for neonatal seizure detection, delivering 95.7% accuracy.

Several studies have focused on generalized modeling across subjects and sessions. Cui et al. [45] and Darvishi-Bayazi et al. [46] used joint-probability discrepancy transfer learning with features from 0–40 Hz wavelet packets, achieving 86.31% accuracy in cross-subject classification. Zarei Eskikand et al. [47] applied neural mass models to analyze time-domain EEG dynamics and predict seizures up to 1 h in advance. Kong et al. [48] used Particle Swarm Optimization (PSO) to enhance both model performance and feature selection strategies, achieving 99.3% accuracy.

Recent innovations have also integrated spatial and temporal dependencies using graph-based learning. Qiao et al. [49] and Xiang et al. [50] implemented Graph Convolutional Networks (GCNs) and Spatial–Temporal Graph Attention Networks (STGATs), reaching sensitivity levels of 98.5%. Wei et al. [51] and Sun et al. [52] leveraged directed graph models to emphasize inter-channel temporal connectivity, reporting 98.15% classification accuracy.

Finally, time-domain modeling beyond scalp EEG has been explored. Aldana et al. [53] decoded motor activity from electrocorticography (ECoG) using Partial Least Squares (PLS), achieving >99% sensitivity. Dümpelmann et al. [54] identified high-frequency oscillations (HFOs) using Kruskal–Wallis tests, and Wei et al. [55] utilized DL models to detect interictal epileptiform discharges (IEDs), achieving 95.1% accuracy.

Table 1. Comparison of performance of ML models with time-domain input.

Ref	Temporal Feature Representation	Model Type	Dataset	Performance
Srinivasan et al. [26]	Approximate Entropy (ApEn)	Elman NN, Probabilistic NN	EEG	100% overall accuracy
Ahammad et al. [27]	Energy, SEN, IQR, MAD	Statistical Classifiers	CHB-MIT	84.2% overall accuracy
Martis et al. [28]	Energy, Fractal Dim., Sample Entropy (via ITD)	Decision Tree	EEG	95.67% average classification accuracy
Fadlallah et al. [29]	Weighted Permutation Entropy (WPE)	Not specified	EEG	Not specified
Yang et al. [30]	Min, Max, Mean, Variance, Skewness, Kurtosis, RMS	Supervised Model	EEG	Not specified
Zhang et al. [36]	Wavelet-decomposed CSP Time Features	CNN	EEG	90.2% sensitivity, 0.096/h FPR
Li et al. [31]	Scale-Dependent Lyapunov Exponents (SDLE)	SVM, Random Forest	EEG	92.2% sensitivity, 0.12/h FPR
Brari [32]	Correlation Dimension (CD)	Simplified Classifier	EEG	Accuracy 100%
Soomro et al. [33]	CCA-based features	MLPNN	EEG	Accuracy 92.583%
Abbaszadeh et al. [34]	IQR, Kruskal-Wallis	Supervised ML	EEG	-
Memarian et al. [35]	mRMR, mutual info	ML classifiers	Multimodal EEG	Accuracy 95%
Wei et al. [37]	Raw EEG (Temporal Modeling)	3D-CNN	EEG	Accuracy > 90%
Xun et al. [38]	Temporal Autoencoding	Sparse Autoencoder	EEG	Error rate 22.93%
Yang et al. [39]	Temporal Dependency	CNN-LSTM, Self-attention	EEG	Accuracy 92.07%
Bundy et al. [56]	Time-domain EEG	PLS Regression	EEG	Sensitivity > 99%
Dümpelmann et al. [54]	Temporal HFOs	Kruskal-Wallis test	EEG	-
Wei et al. [55]	IED classification	DL Temporal Models	EEG	Accuracy 95.1%
Kavitha et al. [57], Abbaszadeh et al. [58]	Peak-to-peak, Variance, IQR, Energy	SVM, DT, KNN, RF	Bonn, Senthil	Alert 75 min before seizure
Qiao et al. [49], Xiang et al. [50]	Temporal-Spatial Features	GCN, STGAT	Multichannel EEG	Sensitivity 98.5%
Wei et al. [51], Sun et al. [52]	Time-domain Connectivity	Directed Graph Models	EEG	Accuracy 98.15%
Qiu et al. [42]	Real-time EEG Monitoring	Cloud/Edge DL	EEG	Accuracy 93.4%
Salafian et al. [43]	Real-time EEG	Federated, TSK-Fuzzy	EEG	Accuracy 91.43%
Zhou et al. [44]	Neonatal EEG Temporal	LMA-EEGNet	EEG (Neonatal)	Accuracy 95.71%
Cui et al. [45], Darvishi-Bayazi et al. [46]	Mean amplitude, std, median, kurtosis, skewness of 0–40 Hz WPD	CEJT Transfer	Cross-subject EEG	Accuracy 86.31%
Zarei Eskikand et al. [47]	Neural Mass Time-Domain	Neural Models	EEG	1 hr before seizure
Kong et al. [48]	PSO, Correlation	Optimized ML Models	EEG	Accuracy 99.32%
Lih et al. [40]	Temporal Transformers	EpilepsyNet	EEG	Accuracy 85%
Saadoon et al. [41]	EEG Variability	EfficientNet-B0 + SVM	EEG	Accuracy 96.12%

provides a summary of the key approaches described above.

6. DL Models with Spectral-Domain Input

Spectral-domain features play a critical role in capturing frequency-specific patterns that characterize epileptic EEG signals [59,60,61,62,63,64,65,66,67]. These features enable models to detect and interpret abnormal neural oscillations commonly associated with seizure activity. EEG recordings are often decomposed into standard frequency bands—Delta (0.5–4 Hz), Theta (4–7 Hz), Alpha (8–12 Hz), Sigma (12–16 Hz), and Beta (13–30 Hz)—to isolate the relevant rhythms that may signal the onset of a seizure [68,69], as illustrated in Figure 3. Subdividing these bands can enhance discrimination accuracy. For instance, Perez et al. [70] proposed a refined beta sub-band to improve classification resolution, while Tsipouras [71] demonstrated that low-frequency bands, particularly those below 7 Hz, achieved a seizure classification accuracy of 90%.

To convert EEG signals into a format suitable for spectral analysis, many studies have applied STFT, which yields time–frequency spectrograms. These spectrograms are commonly used as inputs to CNNs, which can capture localized spectral features across temporal windows. Truong et al. [59,64] implemented this approach using 30 s EEG segments with 50% overlap, normalizing the frequency axis to prevent bias from low-frequency dominance. Their CNN-based model achieved 81.4% sensitivity and a false prediction rate (FPR) of 0.06 per hour.

In a similar context, Wang et al. [67] designed a 3D CNN by stacking STFT-transformed EEG data into a three-dimensional tensor, where each axis corresponded to time, frequency, and EEG channels. By integrating dilated convolutions, they enhanced the model’s temporal perception and reported 85.8% sensitivity and 80.5% accuracy on the CHB-MIT dataset.

Ramos-Aguilar et al. [66] expanded on conventional spectrogram use by combining them with spectral descriptors, such as K-means clustering and local ternary patterns. Their approach, evaluated on the Bonn dataset, achieved the perfect classification of healthy versus epileptic EEG signals. Usman et al. [72] employed STFT for initial feature extraction, followed by a hybrid CNN-SVM classification model, achieving 92.7% sensitivity and 90.8% specificity. In a related effort, Hu et al. [65] proposed the Mean Amplitude Spectrum (MAS) as an alternative spectral descriptor, generating MAS maps for each EEG channel and applying them to a CNN-SVM framework, which reached 86.25% accuracy on CHB-MIT recordings.

Incorporating self-attention and residual mechanisms has shown further promise in spectral learning. Yang et al. [39] developed a dual self-attentive residual CNN trained on STFT-based spectrograms and reported 92.07% accuracy, 89.33% sensitivity, and 93.02% specificity across 13 patients in the CHB-MIT dataset. Liu et al. [73] proposed a multi-view CNN architecture that combined both temporal and spectral features, achieving 93% sensitivity and 71% specificity. Singh and Lobiyal [74] introduced a CNN-LSTM hybrid trained on spectrogram images, which achieved 94.5% accuracy and an F1-score of 0.9376, with an FPR of 0.055 per hour, across various prediction intervals.

Further innovations have leveraged spectral connectivity features. Wang et al. [75] used the Directed Transfer Function (DTF) to measure frequency-domain causal flow between iEEG channels, transforming these relationships into channel-frequency maps processed by CNNs. The model achieved 90.8% sensitivity with an FPR of 0.08 per hour. Romney et al. [76] applied Empirical Ensemble Mode Decomposition (EEMD) in conjunction with neural networks to isolate relevant frequency components, reporting 86.7% sensitivity and 89.5% specificity using 23 CHB-MIT subjects. Qi et al. [77] introduced a hybrid 3D–2D CNN (HyCNN) that preserved spectral depth while optimizing spatial representation, achieving 98.43% accuracy, 98.58% sensitivity, and 96.86% specificity.

Recent advancements have explored the potential of Transformer-based architectures for modeling long-range spectral–temporal dependencies [39,40,72,74,77,78,79,80,81,82,83]. Li et al. [81] proposed EpilepsyNet, a Transformer model that used Pearson Correlation Coefficients (PCCs) as input features, achieving an accuracy of 85%. Zhu et al. [80] built upon this model by combining Transformer layers with LSTM and GRU modules, which yielded a sensitivity of 98.24% on the CHB-MIT dataset. These models underscore the superiority of attention-based mechanisms in modeling complex dependencies beyond the scope of CNNs alone. Li et al. [84] further enhanced this direction by integrating STFT inputs into a Transformer-guided CNN (TGCNN), which achieved 91.5% sensitivity, 93.5% AUC, and an FPR of 0.145 per hour. Assali et al. [79] incorporated a Stability Index (SI) metric alongside STFT features and trained a CNN classifier, achieving accuracies between 90.1% and 94.5%, depending on whether the preictal interval was 30 or 60 min.

Lightweight models leveraging spectral features have also shown great potential for real-time and mobile healthcare environments. Li et al. [81] utilized entropy-based spectral features in combination with traditional classifiers such as SVMs and decision trees, achieving high classification accuracy even with limited EEG data. This suggests suitability for resource-constrained clinical settings. Urbina Fredes et al. [82] focused on alpha and beta band features, pairing them with SVMs to enable accurate and efficient seizure detection, with minimal computational overhead suitable for wearable systems. Amer and Belhaouari [83] introduced a novel Forward–Backward Fourier Transform (FBFT), which enhanced the clarity of EEG frequency components. When paired with a CNN classifier, their method demonstrated excellent classification accuracy across various brain disorders, including epilepsy.

Hybrid deep learning architectures that integrate spectral and temporal information have also gained momentum. CNN-LSTM models, as discussed in recent work [85], have demonstrated the capability to capture a frequency-specific structure through convolution while modeling sequential evolution through recurrent layers. These models are especially effective when applied to spectrally transformed EEG data. Furthermore, multi-domain integration strategies are emerging, as shown in studies like [80], where spectral-domain features are combined with Transformer architectures and graph neural networks (GNNs). These systems benefit from the precise frequency resolution of spectral methods, the attention-based abstraction of Transformers, and the topological modeling of inter-channel relationships enabled by GNNs.

Table 2. Comparison of performance of DL models with spectral-domain input.

Ref	Feature Representation	Model Type	Dataset	Performance Metrics
Truong et al. [59,64]	STFT (spectrogram)	CNN	CHB-MIT	81.4% Sensitivity, FPR 0.06/h
Wang et al. [67]	STFT (3D Tensor)	3D CNN with Dilated Conv	CHB-MIT	85.8% Sensitivity, 80.5% Accuracy
Ramos-Aguilar et al. [66]	Spectrogram + Descriptors	MLP, K-means, LTP	Bonn	100% Accuracy
Usman et al. [72]	STFT	CNN-SVM	CHB-MIT	92.7% Sensitivity, 90.8% Specificity
Hu et al. [65]	Mean Amplitude Spectrum (MAS)	CNN-SVM	CHB-MIT	86.25% Accuracy
Yang et al. [39]	STFT	Self-attentive Residual CNN	CHB-MIT	92.07% Accuracy, 89.33% Sensitivity, 93.02% Specificity
Liu et al. [73]	Temporal + Spectral	Multi-view CNN	CHB-MIT	93% Sensitivity, 71% Specificity
Singh & Lobiyal [74]	Spectrogram	CNN-LSTM	CHB-MIT	94.5% Accuracy, F1-score 0.9376, FPR 0.055/h
Wang et al. [75]	DTF (Spectral Flow)	CNN + Moving Avg	iEEG	90.8% Sensitivity, FPR 0.08/h
Romney et al. [76]	EEMD	Neural Networks	CHB-MIT	86.7% Sensitivity, 89.5% Specificity
Qi et al. [77]	Spectral Depth	3D+2D HyCNN	CHB-MIT	98.43% Accuracy, 98.58% Sensitivity, 96.86% Specificity
Lih et al. [40]	PCC	Transformer	EEG	85% Accuracy
Zhu et al. [80]	Spectral Features	Transformer + LSTM/GRU	CHB-MIT	98.24% Sensitivity
Li et al. [84]	STFT	TGCNN	CHB-MIT	91.5% Sensitivity, AUC 93.5%, FPR 0.145/h
Assali et al. [79]	STFT + SI	CNN	CHB-MIT	Accuracy 90.1%–94.5%
Li et al. [81]	Entropy Spectral Features	SVM, Decision Trees	EEG	High Accuracy, Low Data Requirements
Urbina Fredes et al. [82]	Alpha/Beta Bands	SVM	EEG	High Accuracy, Real-Time Suitability
Amer & Belhouari [83]	FBFT	CNN	EEG	High Accuracy

presents a summary of these models and their performance.

7. Discussion

The application of deep learning (DL) has significantly advanced the prediction and classification of epileptic seizures, particularly in accurately identifying ictal events. In contrast to traditional machine learning (ML) approaches that depend on handcrafted features, DL models can automatically learn complex patterns from raw EEG data, including both temporal and spectral dynamics. This capacity to discover high-level representations directly from input signals has proven especially valuable for detecting seizure onset, where subtle fluctuations in signal properties may escape manual analysis. The performance of DL-based seizure prediction systems is closely tied to the nature and quality of feature representation, with spectral-domain inputs demonstrating particular promise in capturing ictal-specific signatures.

7.1. Impact of Feature Representation on Prediction Performance

Temporal features such as signal amplitude, entropy measures, and waveform variability have historically underpinned seizure classification in ML pipelines using models like SVM and RF [31,32]. These features are computationally efficient and often interpretable, making them suitable for real-time applications. However, their capacity to model the nonlinear dynamics and abrupt transitions characteristic of epileptic seizures is limited. While methods employing Approximate Entropy [26] and Scale-Dependent Lyapunov Exponents [31] have demonstrated encouraging results in distinguishing preictal from ictal phases, their generalization to diverse patient populations and recording conditions remains suboptimal.

Spectral-domain features have emerged as a more expressive alternative, offering deeper insight into seizure-related frequency anomalies. Time–frequency decomposition techniques such as STFT, wavelet transform, and EEMD have been extensively adopted to convert EEG signals into rich spectral representations. These inputs enable CNNs to extract hierarchical features that encode both spatial and frequency-related seizure dynamics [59,60,76]. CNNs trained on STFT-based spectrograms, for example, have demonstrated superior performance, with some achieving over 92% sensitivity and 90% specificity in detecting ictal activity [72].

Hybrid and attention-based architectures have further improved spectral learning. The Transformer-based model EpilepsyNet [40], which integrates Pearson correlation coefficient features into self-attention mechanisms, achieved robust results across varying seizure contexts. Building on this, Zhu et al. [80] incorporated recurrent units such as LSTM and GRU into the Transformer pipeline, boosting sensitivity beyond 98% on the CHB-MIT dataset. Similarly, models such as the 3D–2D hybrid CNN [77] and dual self-attentive residual CNN [39] have delivered high classification accuracy by learning multiscale temporal–spectral dependencies.

7.2. Challenges and Limitations

Despite these advancements, several persistent challenges limit the broader applicability of current models. One of the most pressing issues is generalizability. Many high-performing DL models are developed and validated using patient-specific datasets, which tend to overfit to the unique signal properties of a limited cohort. This restricts the ability of such models to be scaled across diverse populations. Cross-patient variability in EEG patterns, coupled with inconsistencies in acquisition protocols, electrode placement, and sampling frequencies, further undermines reproducibility and external validity [27,73,79].

Another major limitation involves the scarcity of preictal data, which leads to class imbalance and reduced model robustness during training. Although data augmentation techniques, including generative adversarial networks (GANs) [86] and synthetic EEG generation, have been proposed to mitigate this issue [36], their effectiveness is highly dependent on dataset characteristics and model architecture.

High-resolution models such as 3D CNNs and Transformers also present significant computational challenges. These architectures, while powerful, often require large memory footprints and high processing speeds, limiting their deployment in real-time or embedded systems. Lightweight alternatives, such as LMA-EEGNet and model pruning strategies, offer potential solutions [74], but more work is needed to balance classification accuracy with computational efficiency.

Interpretability remains a major barrier to clinical adoption. Although temporal features offer a degree of transparency, spectral representations learned by deep models often function as black boxes. Emerging explainable AI methods, including saliency maps, SHAP values, and attention-based visualization tools, have attempted to clarify model reasoning [65]. However, these approaches are still in their early stages and often fall short of providing clinically actionable explanations.

7.3. Future Directions

Future research must address these challenges by emphasizing multimodal feature fusion, model generalization, and explainability. Integrating temporal and spectral features into unified architectures will likely yield models that are both accurate and resilient to signal variability. Techniques such as transfer learning, federated learning, and domain adaptation hold promise for developing generalized models that can operate effectively across heterogeneous EEG datasets and clinical environments.

The standardization of EEG datasets is also essential. Harmonizing sampling rates, electrode configurations, and annotation protocols across public databases would enable fairer benchmarking and facilitate model replication. Initiatives aimed at curating large-scale, labeled, and demographically diverse EEG datasets will be crucial in advancing the field.

Ultimately, the performance of DL-based seizure prediction systems is shaped by the feature domain from which models learn. Spectral-domain features, particularly when combined with advanced architectures such as Transformers and hybrid CNN-LSTM models, substantially improve the detection of ictal states. However, continued progress depends on addressing unresolved issues related to generalization, interpretability, and deployment feasibility. The future of seizure prediction lies in the development of adaptive, transparent, and clinically reliable models that can accurately forecast seizures and support real-time interventions.

8. Conclusions

This review synthesized recent progress in DL-based seizure prediction, with an emphasis on how different model architectures and feature representations affect the ability to distinguish between normal, preictal, and ictal EEG states. The analysis revealed that leveraging frequency-domain transformations and combining them with temporal features yield models that are more adept at capturing the complex dynamics associated with seizure onset.

Rather than reiterating architectural superiority, this review underscored the importance of aligning feature representation strategies—particularly temporal and spectral domains—with the neurophysiological characteristics of seizures. The choice of domain not only influences predictive accuracy but also determines a model’s generalizability and interpretability. DL models incorporating spectral–temporal features, especially those using STFT, wavelets, or hybrid CNN–Transformer mechanisms, have emerged as effective tools in capturing ictal transition dynamics.

Despite these technical advances, key issues persist, including model overfitting to patient-specific patterns, insufficient cross-dataset robustness, and the interpretability of learned features. This review emphasized the need to rethink preictal window design, improve dataset standardization, and integrate explainable AI (XAI) methods that can map model outputs back to meaningful EEG biomarkers.

Going forward, the convergence of transfer learning, federated learning, and edge computing will offer promising avenues to translate high-performing models into real-world clinical tools. Importantly, future efforts should not only aim for higher performance metrics but also prioritize transparency, adaptability, and clinical relevance. Achieving these goals will mark a significant step toward making seizure prediction systems both scientifically rigorous and practically deployable.