As shown in Figure 3, the remainder of this survey is organized as follows. Section 2 introduces the key properties of EEG signals and describes the main types of technologies used and technological innovations. Section 3 presents automatic detection technologies for epileptic EEG signals based on nonlinear dynamics, including methods based on Lyapunov exponents (LE), fractal dimensions (FD), entropy measures, and deep learning approaches. Section 4 introduces epileptic EEG signal prediction models based on nonlinear dynamics, including chaos-based, LSTM-based, and hybrid prediction models. Section 5 provides a comprehensive summary of the main epilepsy databases. Finally, Section 6 discusses the key challenges and future directions in epilepsy detection and prediction research.

EEG is a technique that records the electrical activity of neurons in the cerebral cortex via electrodes, reflecting the synchronized discharge patterns of brain neurons. EEG signals exhibit three fundamental characteristics: non-stationarity, nonlinearity, and multi-scale dynamics (Andrzejak et al., 2012). First, the non-stationarity of EEG signals manifests as temporal variations in their statistical properties, particularly during epileptic seizures, where significant fluctuations occur in both frequency and amplitude (Ahmad et al., 2024). These fluctuations render traditional steady-state analysis methods inadequate for direct application. Second, the nonlinearity of EEG signals arises from the complex interactions among large populations of neurons, which introduce nonlinear effects through synaptic connections and feedback mechanisms (Qi et al., 2024). This results in complex dynamic behaviors that cannot be adequately captured by simple linear models. Finally, the multi-scale nature of EEG signals is evident in their encoding of information across multiple time scales, ranging from millisecond-level fast oscillations to second-level slow wave activities. Consequently, analysis methods must account for dynamic changes at various temporal scales to fully characterize these signals.

Nonlinear dynamics establishes rigorous mathematical frameworks for EEG analysis through differential topology and ergodic theory, transcending the limitations of linear methods bound by superposition principles. This mathematical formalism enables phase space reconstruction via Takens' embedding theorem, which provides a diffeomorphic mapping of time series data into higher-dimensional manifolds, preserving the essential dynamics of the underlying system. Fractal dimension calculations employing box-counting algorithms quantify the self-similar geometry of attractors, while entropy measurements derived from the Kolmogorov-Sinai formalism characterize information generation rates in dynamical systems. These methodologies derive their mathematical superiority from the capacity to rigorously quantify critical phenomena. Lyapunov exponents measure exponential divergence of neighboring trajectories in tangent space, correlation dimensions reveal the scaling properties of attractor distributions, Hurst exponents capture long-range dependence through fractional Brownian motion models, and Kolmogorov entropy estimates the rate of information loss. Table 1 summarizes the descriptions, advantages, limitations, and clinical applications of nonlinear dynamics parameters in epileptic EEG analysis. As formalized in Table 1, these parameters constitute functionals of the time series of EEG X(t), with the maximal Lyapunov exponent λ defined through the limit operator λ=limt→∞t-1ln (||δX(t)||/||δX(0)||), mathematically encoding the sensitivity of the system to initial conditions. Such constructs transform epileptic transitions into quantifiable instabilities within nonlinear differential equations, bridging abstract dynamical theory with clinical pathophysiology (Vignesh et al., 2025).

The dynamic changes in EEG signals during epileptic seizures represent a highly nonlinear neurophysiological process that traditional linear analysis methods often fail to characterize adequately. Nonlinear dynamical indicators provide more accurate characterization of critical transitions from pre-ictal to ictal states by quantifying chaotic properties, complexity, and synchronization patterns in EEG signals. During the pre-ictal phase, significant decreases in approximate entropy and sample entropy reveal the gradual loss of complexity in neural electrical activity (Giannakakis et al., 2013; Shen et al., 2013; Song et al., 2012), while declining LLE reflects weakened chaotic characteristics in neural networks (Aarabi and He, 2017). These changes collectively indicate the system's transition toward a highly synchronized ictal state. Simultaneously, the reduction in correlation dimension suggests decreased degrees of freedom in the neural system, and increased phase synchronization index quantifies enhanced functional coupling between brain regions (Lehnertz et al., 2001).

These nonlinear indicators not only detect subtle dynamic changes undetectable by traditional linear methods, but also provide reliable early warning signals minutes to tens of minutes before seizure onset, significantly exceeding the prediction window of linear methods such as power spectral analysis. Crucially, nonlinear indicators can reveal critical slowing phenomena preceding epileptic seizures–decreased system resilience before state transitions–representing sensitivity to stability changes that linear analysis completely lacks. As demonstrated in Table 2, nonlinear indicators substantially outperform traditional linear methods (e.g., power spectral density, band energy) in identifying critical state transitions by capturing the evolution of chaotic properties and complexity in neuronal populations. They achieve superior sensitivity and specificity compared to linear feature-based models, particularly demonstrating enhanced adaptability to interindividual variability and nonstationary EEG signals.

The LE characterizes the chaotic properties of EEG signals by quantifying their sensitivity to initial conditions. Variations in LE reflect dynamic evolutions of neuronal activity. Studies show that Lyapunov exponents of EEG signals decrease significantly during epileptic seizures, indicating enhanced neuronal synchronization and increased system orderliness (Iasemidis et al., 1990). This phenomenon was first identified by Iasemidis et al. in 1990 through the analysis of intracranial EEG (iEEG) signals from epilepsy patients. They observed that the divergence rate of neighboring trajectories in phase space (that is, the maximum Lyapunov exponent, LLE) gradually decreases as seizures approach, signifying a transition from chaotic to ordered brain activity (Iasemidis et al., 1990).

Building on this foundational understanding, Lyapunov exponents have been widely adopted for early seizure detection. (Acharya et al. 2013) combined Lyapunov exponents with support vector machines (SVM) to achieve high classification precision in the Bonn dataset, validating the efficacy of this metric in the analysis of epileptic EEG signals. Recent advances have further enhanced the application of Lyapunov exponents in epilepsy research. (Zhao et al. 2024) pioneered the integration of neurophysiological brain models with nonlinear Kalman filtering, achieving precise detection of focal epileptic seizures through time-evolving Lyapunov spectrum analysis, while revealing that external inputs serve as key drivers of chaotic degree alterations in epileptogenic zones. Complementing this theoretical advancement, (Brari and Belghith 2022) developed a noise-robust PLLE algorithm by refining Wolf's LLE computation method, which effectively preserves essential chaotic signal characteristics while significantly mitigating noise interference, ultimately achieving 100% classification accuracy in epilepsy detection using the Bonn EEG dataset.

FD serves as a critical metric for quantifying the complexity and self-similarity of EEG signals, with its value dynamically reflecting alterations in brain states. Research indicates that self-similarity in EEG signals increases markedly during epileptic seizures, with significant differences in FD observed between healthy individuals and patients (Janjarasjitt, 2014; Silalahi et al., 2021; Lahmiri, 2018).

Several computational approaches have been developed to quantify fractal dimensions in EEG analysis. Commonly used computational methods include the Higuchi fractal dimension (HFD), Petrosian fractal dimension (PFD), and detrended fluctuation analysis (DFA). Among these, HFD is preferred for its computational efficiency and low error rates (Silalahi et al., 2021). PFD quantifies signal fluctuation and self-similarity while outperforming fast Fourier transform (FFT) in temporal resolution (Yang et al., 2020). DFA reveals self-similarity and fluctuation trends in non-stationary time series via the Hurst exponent (Lahmiri, 2018). Additionally, line length (LL), a simplified form of FD, exhibits robust performance in detecting the abruptness of seizure-related signals (Ma et al., 2021; Anuragi et al., 2021).

Building upon these established methods, recent developments have further refined fractal dimension analysis for epilepsy detection. (Adda et al. 2025) proposed an enhanced detrended fluctuation analysis (ADFA) method that precisely quantifies long-range temporal correlations (LRTC) and FD in EEG signals. This approach not only validated that FD values during epileptic seizures were significantly higher than during interictal periods (p < 0.001), but also achieved a 12.3% improvement in classification accuracy compared to conventional DFA methods, substantially enhancing the detection precision of epileptic seizures through FD analysis.

Entropy, a vital metric for quantifying signal irregularity and unpredictability, encompasses common variants such as approximate entropy (ApEn), sample entropy (SampEn) and multiscale entropy (MSE). In epileptic EEG signal analysis, entropy variations reflect dynamic signal properties. Studies reveal significant reductions in entropy during epileptic seizures, indicating increased signal predictability.

Early approaches to entropy-based epilepsy detection focused on combining different entropy measures. Oxidative Medicine and Cellular Longevity (2024) used entropy-based features—including sample entropy (SampEn), permutation entropy (PermEn), and fuzzy entropy (FuzzyEn)—individually and in combination to form three-dimensional feature vectors. Their findings demonstrate that combining SampEn, PermEn, and FuzzyEn yields the highest accuracy and recall rates. However, this approach lacks integration of temporal or frequency-domain information. To address this limitation, (Fei et al. 2017) developed a more comprehensive approach by adopting the fractional Fourier transform (FrFT), adaptive maximum Lyapunov exponents and energy characteristics to capture chaotic and frequency domain characteristics of epileptic EEG signals. Similarly, Zhang J. et al. (2024) adopted a more comprehensive methodology by extracting features via the discrete wavelet transform (DWT), power spectral density (PSD), standard deviation, band energy and fuzzy entropy. While this method effectively aggregated diverse feature information, it also introduced redundancy and increased computational complexity.

Recognizing the need for feature optimization, (Zhang et al. 2018) generated 2,794 features per sample using multiple extraction techniques, which were subsequently evaluated through sequential selection algorithms (IE, VARA, IRFE, and BackF) to filter critical features. Building on these multi-feature approaches, (Chen et al. 2019) used eight entropy algorithms, DWT-based ANOVA, FSF, and LS-SVM to differentiate ictal and interictal EEG states, ultimately providing a compact seizure diagnostic tool with exceptional classification precision, sensitivity, and specificity.

Deep learning, a key branch of machine learning, has gained prominence in the detection of epileptic EEG signals, with convolutional neural networks (CNNs) emerging as the dominant architecture due to their automated feature extraction capabilities. In the early development of CNN-based approaches, (Ahmedt-Aristizabal et al. 2018) pioneered the use of one-dimensional CNN (1D-CNN) for the direct classification of raw EEG signals, achieving an average accuracy of 88.7%. Their design eliminates manual feature extraction, substantially streamlining workflows. Building upon this foundation, (Wang et al. 2021) further employed stacked 1D-CNN architectures to achieve over 99% accuracy on two public datasets, though concerns about the reliability of the results arose due to missing test sets.

To enhance data utilization efficiency beyond 1D approaches, two-dimensional CNNs (2D-CNNs) have been extensively explored. (Avcu et al. 2019) surpassed traditional multichannel models using only dual-channel data. Similarly, (Das et al. 2024) validated the superiority of 2D-CNNs through the reshaping of EEG data format. Addressing the challenge of limited data availability, (Pan et al. 2022) developed a lightweight 2D-CNN that integrates the discrete Fourier transform (DFT) and the discrete wavelet transform (DWT), demonstrating strong performance under limited data, although insufficient testing hindered generalization. Advancing further into higher-dimensional architectures, the introduction of three-dimensional CNNs (3D-CNNs) has pushed technical frontiers: (Xu et al. 2024) reduced detection latency by 50% through probabilistic forecasting, significantly improving real-time capabilities.

Beyond CNNs, the exploration of alternative deep learning models is progressing. Recognizing the temporal characteristics of EEG signals, recurrent neural networks (RNNs) and their variants, long-short-term memory networks (LSTMs), demonstrate unique advantages. LSTMs address the vanishing gradient problem in traditional RNNs through gating mechanisms, making them effective tools for processing long-sequence data (Hochreiter and Schmidhuber, 1997). (Najafi et al. 2022) improved the classification accuracy for focal and generalized epilepsy by combining wavelet transforms with LSTM-RNNs. Extending this approach, (Singh and Malhotra 2022) constructed a dual layer LSTM network based on spectral characteristics, verifying performance improvements through complex architectures. Further integrating feature engineering with LSTM architectures, (Goel et al. 2023) optimized the diagnostic efficacy by integrating entropy characteristics with LSTMs. However, these methods focus primarily on frequency domain and temporal information, neglecting the spatial dimensions of EEG signals and limiting the comprehensive utilization of multidimensional features.

While mainstream approaches have focused on established architectures, applications of generative adversarial networks (GANs) and transfer learning remain exploratory, but show promise. (Truong et al. 2019) pioneered GANs for epilepsy prediction, opening new avenues beyond data augmentation. Meanwhile, multimodal approaches, though enabling real-time detection, face hardware, and data limitations. In pursuit of biologically-inspired and energy-efficient solutions, spiking neural networks (SNNs), which mimic biological neuronal dynamics for low-energy processing, have seen training challenges addressed by (Zhang et al. 2025) through pulsed recurrent neural networks. Addressing computational efficiency concerns, (Liu et al. 2017) reduced memory costs by 75% using cosine convolutional networks (COSCNN). Additionally, the capsule networks of (Sabour et al. 2017) overcome CNN limitations in the modeling of feature relationships. These innovations drive models toward efficiency and interpretability.

Recent breakthroughs are primarily hinged on the integration of deep learning with non-linear dynamics features. (Fu et al. 2015) achieved high classification accuracy by combining intrinsic mode functions (IMF) derived from empirical mode decomposition (EMD) with SVMs. Advancing interpretability alongside accuracy, (Khan et al. 2024) proposed a fuzzy deep learning (FDL) framework that integrates locally interpretable models (LIME) and Shapley values (SHAP), improving both accuracy and interpretability. Demonstrating the power of feature fusion, Acharya et al. attained near-perfect classification accuracy on the Bonn data set by combining Lyapunov exponents and fractal dimensions. These studies demonstrate that multidimensional feature fusion comprehensively captures dynamic EEG characteristics, providing clinically reliable insights.

Chaos-based prediction models analyze nonlinear dynamic properties within EEG signals—such as phase space reconstruction and Lyapunov exponent variations—to identify critical state transitions preceding seizures. Phase space reconstruction techniques, for instance, employ embedding dimensions and time delay parameters to construct multidimensional phase spaces, enabling the tracking of attractor structure dynamics. Research indicates a significant reduction in attractor dimensionality within EEG signals during preictal phases, reflecting heightened neuronal synchronization.

To enhance model robustness and adaptability, recent advancements incorporate adaptive parameter optimization algorithms. Specifically, particle swarm optimization-enhanced phase space reconstruction dynamically adjusts embedding dimensions to improve noise resilience. Clinically, these models have enabled long-term pediatric epilepsy monitoring, achieving seizure predictions 10–15 min in advance with sensitivities surpassing 85%. However, chaos models remain constrained by data length dependencies, necessitating sliding window strategies to enhance real-time performance. Building on these foundational chaos-based approaches, (Lu et al. 2024) systematically applied phase space reconstruction, phase synchronization, and state space entropy to investigate nonlinear characteristic disparities in epileptic EEG signals. Their work identifies distinct preictal nonlinear features compared to ictal and interictal states, substantially advancing seizure detection and prediction efficacy.

Long short-term memory (LSTM) networks leverage their gating mechanisms and temporal modeling proficiency to capture long-term dependencies in EEG signals during transitions from normal to epileptic states, solidifying their role as a cornerstone of seizure prediction. In pioneering applications of LSTM to epilepsy prediction, (Tsiouris et al. 2018) first demonstrated LSTM's potential on the CHB-MIT dataset, achieving prediction sensitivities and specificities exceeding 99% and outperforming conventional convolutional neural networks. Building upon this foundational work, subsequent studies expanded the utility of LSTM: (Singh and Malhotra 2022) fused fast Fourier transform-derived spectral power and mean spectral amplitude features to develop a dual-layer LSTM network, improving accuracy through multidimensional feature integration. Similarly, (Tsiouris et al. 2018) incorporated statistical moments, time domain features, and discrete wavelet transforms within an analogous framework, empirically validating LSTM robustness in multimodal feature processing.

Recognizing the potential benefits of combining different architectural strengths, researchers are increasingly adopting hybrid designs. (Lu et al. 2022) proposed an end-to-end multiframe network that combined convolutional neural networks for spatial feature extraction with LSTMs for temporal dependency modeling, surpassing standalone models on both scalp and intracranial EEG data. As the field matured, addressing the interpretability of these complex models has gained prominence. (Gao et al. 2023) devised a multiscale prototype-part network that delivers transparent inference through self-explanatory mechanisms, maintaining state-of-the-art performance while resolving deep learning opacity challenges and providing clinically actionable insights. Advancing hybrid architectures further, (Alharbi et al. 2024) developed a hybrid graph neural network architecture based on Bayesian optimization, which effectively captures spatiotemporal features of EEG signals through CNN-LSTM integration with adaptive skip connections, significantly enhancing the predictive performance of LSTM on imbalanced epilepsy datasets.

Moving toward practical clinical implementation, current research emphasizes LSTM lightweighting and real-time optimization. Architectures integrating multiscale entropy analyze entropy decay patterns across temporal scales to predict seizures 20 min in advance with specificities exceeding 90%. Complementing these temporal approaches, attention-enhanced Transformer-LSTM models prioritize abnormal discharges in critical brain regions such as the temporal and frontal lobes, reducing false positives. Looking ahead, future investigations must prioritize lightweight LSTM deployment in wearable devices while advancing multimodal feature fusion and biologically inspired architectures to improve clinical applicability.

Hybrid models unify non-linear dynamics features with deep learning architectures to harmonize interpretability and predictive performance. A representative example is the CNN-RQA hybrid model, where convolutional neural networks extract localized time-frequency EEG features while recurrence quantification analysis generates dynamic metrics including recurrence rate and laminarity. Combined features are subsequently processed by fully connected layers for prediction. Using the Freiburg data set, this model detects seizures 25 min in advance, demonstrating superior sensitivity to low-frequency oscillatory signals compared to standalone approaches (Sunaryono et al., 2022). Expanding upon this CNN-based approach, alternative hybrid methodologies integrate empirical mode decomposition with gated recurrent units, decomposing EEG signals into intrinsic mode functions for multiresolution nonlinear feature extraction followed by temporal dependency modeling, achieving exceptional precision in the prediction of neonatal seizures (Ma et al., 2023). Building on these foundational hybrid designs, Zhang J. et al. (2024) further advanced this paradigm through a CNN-GRU-AM model employing discrete wavelet transform-based feature extraction and gated attention mechanisms to enhance detection precision.

Addressing the computational challenges inherent in processing long EEG sequences, hybrid models increasingly integrate attention mechanisms and residual networks. The Transformer architecture manages extended EEG sequences efficiently through multi-head attention, reducing temporal complexity (de Santana Correia and Colombini, 2022; Rukhsar and Tiwari, 2023). Simultaneously, residual networks mitigate gradient vanishing through skip connections, enhancing model stability (He et al., 2016). Implementing these architectural advances, (Huang et al. 2024) introduced a temporal convolutional self-attention network for extracting autonomous features, while (Zhu et al. 2024) proposed a SE-TCN-BiGRU model combining channel selection with parameter efficiency. Focusing on real-world deployment constraints, (Zhou et al. 2024) streamlined complexity through a lightweight multi-branch network tailored for real-time monitoring.

Beyond single-modality approaches, multimodal data fusion has emerged as a pivotal factor in improving detection robustness. (Ahmedt-Aristizabal et al. 2018) integrated facial, postural, and EEG features to differentiate mesial and lateral temporal lobe epilepsy. Demonstrating the benefits of physiological signal integration, (Sirpal et al. 2019) combined EEG with functional near-infrared spectroscopy, increasing classification accuracy by 8%. Advancing toward comprehensive brain imaging fusion, (Hosseini et al. 2020) fused resting-state fMRI and EEG for the localization of the epileptogenic zone by edge computing, while (Martini et al. 2021) optimized real-time sensitivity using stereo-EEG and video-EEG with dynamic thresholds.

While these technological advances show considerable promise, generalizability remains a critical challenge that requires rigorous validation. (Aslam et al. 2022) reported exceptional CNN-LSTM performance on specific data sets, but noted variability between data sets. Similarly, Zhang X. et al. (2024) reduced computational demands through temporal attention models, but retained dependency on high-quality annotations. To address these limitations, future efforts must prioritize multicenter validation and explore few-shot learning alongside unsupervised strategies to accelerate clinical translation.

Comprehensive databases are characterized by multimodal data integration and complete clinical information, providing a multidimensional analytical basis for nonlinear dynamics modeling. A prime example is the Siena Scalp EEG Database v1.0.0 (Detti, 2020), which integrates long-term monitoring data (128 h from 14 patients with refractory epilepsy, synchronized acquisition of dual-channel EKG signals using a high sampling rate of 512 Hz, and contains 47 video-verified seizure events. Its significant advantage is that it provides accurate annotations of the onset and end time of seizures (error < 0.5 s), suitable for the time-frequency characterization of seizure detection algorithms.

Another prominent comprehensive database is the CHB-MIT database,¹ which contains continuous scalp EEG recordings (total duration >850 h) from 23 pediatric patients at Children's Hospital Boston. Data were collected using a drug-free clinical monitoring protocol and contained 162 definite seizures (Goldberger et al., 2000). Its characterization of longitudinal data provides important support for the study of cyclical seizure patterns.

The most comprehensive and ambitious database in this category is EPILEPSIAE (Ihle et al., 2012), which is one of the most comprehensive epilepsy databases available. As the flagship epilepsy research platform of the European Union, this repository integrates intracranial/scalp synchronized EEG data from 275 patients (>45,000 h of total recordings), covering 2,662 seizure events and corresponding MRI images (Klatt et al., 2012). Its innovation lies in the construction of a multidimensional data correlation system (electrophysiology-imaging-clinical phenotype), which supports the study of epilepsy network dynamics based on complex systems theory.

In contrast to comprehensive databases, specialized databases, on the other hand, focus on specific types of EEG signals or specific research questions related to epilepsy. These databases typically provide highly specialized data that are applicable to specific research needs.

The premier example of specialized databases is the Freiburg Seizure Prediction Project EEG Database,² which provides data from the Freiburg Seizure Prediction Project through long-term intracranial electrode monitoring (sampling rate 256 Hz) in 21 patients, providing continuous recordings from 50 min before to 30 min after seizure onset (Ihle et al., 2012). Its high spatiotemporal resolution feature (covering 128 deep cortical electrode channels) is particularly suitable for the extraction of non-linear features in the seizure onset zone.

Complementing intracranial data with large-scale clinical recordings, Temple University Hospital EEG Resources (Temple University EEG Corpus—Downloads) contains 26,846 clinical EEG recordings collected between 2002 and 2017 (total data volume >50 TB) using a three-tier quality assessment system (Level I: signal integrity; Level II: clinical diagnostic consistency; Level III: accuracy of event annotation) (Ochal et al., 2020). Its large-scale abnormal EEG samples (37.2% of the total) provide important training resources to build robust seizure detection models.

In addition to the aforementioned databases, there are several other significant databases that serve specific research purposes, such as the Bonn EEG Database (Andrzejak et al., 2001) and the Seizure Detection Challenge database from the UPenn and Mayo Clinics (Temko et al., 2015).

The Bonn EEG Database represents a foundational resource for controlled epilepsy research, comprising five distinct datasets: Healthy Group A/B, Interictal Group C/D, and Seizure Group E. Each data set includes 100 segments of single channel EEG recordings lasting 23.6 s, sampled at 173.61 Hz. The database's rigorously controlled experimental conditions, including open/closed eye states and a clear distinction between seizure periods and interictal intervals, make it an ideal platform for validating the stability of nonlinear features, such as Lyapunov exponents and entropy values.

Extending beyond human subjects to include animal models, the Seizure Detection Challenge database from the UPenn and Mayo Clinics provides data from human and canine epilepsy models, with a seizure-to-interictal time ratio of 1:24. This database employs a double-blind annotation strategy, validated by three independent experts, and incorporates a multiscale evaluation framework that assesses sensitivity, specificity, and delay time to enhance the clinical utility of seizure detection algorithms.

Despite significant advances in the automatic detection and prediction of epileptic EEG signals using nonlinear dynamics, their clinical translation still faces fundamental mathematical obstacles rooted in the incompatibility of probabilistic frameworks and topological data constraints. First, data quality and generalizability remain core bottlenecks. EEG signals are highly susceptible to environmental noise and physiological artifacts. The divergence of patient-specific EEG distributions violates the Lipschitz continuity assumption in the reproducing kernel Hilbert space, leading to exponential growth of the generalization error bounds with distributional bias and significantly degrading model performance across patients or devices. Furthermore, although existing public datasets are large, their controlled laboratory environments differ from real-world dynamics, leading to topological distortions in the feature manifolds and limiting clinical applicability, thus necessitating technical adjustments for clinical deployment (Jemal et al., 2021). Second, the complexity of multimodal data fusion hinders a complete understanding of epileptic pathophysiology (Rai et al., 2025). Insufficient temporal alignment and feature integration of multimodal data currently lack standardized protocols, thus increasing the complexity of model design.

Existing algorithms face challenges in balancing sensitivity and specificity in epileptic prediction tasks. As the seizure prediction window (SOP) expands, the increase in false positive rates (FPR) becomes a non-negligible issue, while the trade-off between precision and recall further limits model practicality. To address these limitations, research should focus on end-to-end process optimization. In the preprocessing stage, Butterworth filters or wavelet packet transforms (WPTs) can effectively remove noise and improve signal-to-noise ratios. Feature engineering needs to retain classic nonlinear features like entropy and correlation dimensions while exploring new features with high inter-class variance, using recursive feature elimination (RFE) or support vector machine recursive feature elimination (SVM-RFE) to select the optimal feature subset, reducing redundancy and improving computational efficiency. In classifier optimization, prioritizing hyperparameter tuning of mature models like SVMs and CNNs, and combining with statistical validation to optimize decision thresholds, rather than blindly developing new architectures, can more effectively enhance performance.

Feature extraction and selection are the core stages of model optimization. (Jemal et al. 2021) found through analyzing the false positive rate and overlap rate that a specific combination of features can significantly reduce the algorithm's dependence on data volume and maintain robustness in handling small samples. (Zhang et al. 2018) used variance analysis (VarA) and backward feature selection (BackFS) to eliminate features with low discrimination, significantly improving classification accuracy. (Savadkoohi et al. 2020) combined t-tests with sequential forward floating search (SFFS) to select the optimal feature set, ultimately achieving 100% precision, sensitivity, and specificity. These cases demonstrate that refined feature engineering can effectively balance model complexity and performance, providing a viable path for algorithm improvement. However, extracting Morse-Smale complexes from EEG fluctuations remains an unsolved problem in geometric analysis.