A Hybrid Multi-Domain Feature Fusion and Ensemble Learning Framework for Neurodynamic Analysis of Focal and Non-Focal EEG Signals

Abstract

Accurate differentiation between focal and non-focal EEG activity remains a challenging task due to the nonlinear and nonstationary nature of neural signals. This work proposes a novel hybrid framework that combines multi-stage signal filtering, multi-domain feature extraction, and ensemble learning to improve automated neurodynamic analysis. The framework employs a dual-path design: the first path extracts linear, nonlinear, and hybrid signal descriptors, while the second path generates dimensionality-reduced representations that are fused with these descriptors to form a comprehensive, information-rich feature set. Performance was rigorously evaluated across eight machine learning paradigms, including conventional, ensemble, and hybrid models, ensuring methodological rigor and reproducibility. Results demonstrate that fusing nonlinear features with reduced-dimensional embeddings, when processed through a stacked ensemble classifier, achieves superior discrimination between focal and non-focal EEG patterns. The framework's scalability, computational efficiency, and adaptability establish a robust foundation for automated EEG-based neurodynamic analysis, offering strong potential for future studies in cognitive signal processing, neural dynamics modeling, and epilepsy research.

Keywords

EEG signal dynamics multi-domain feature extraction ensemble learning dimensionality reduction focal versus non-focal EEG neurodynamic analysis

1. Introduction

Epilepsy is a chronic neurological disorder characterized by recurrent seizures arising from abnormal and unpredictable electrical activity in the brain. Accurate differentiation between focal and non-focal EEG patterns is critical for understanding seizure dynamics and advancing computational analyses. However, the nonstationary and nonlinear nature of EEG signals, coupled with noise and artifacts, makes automated classification challenging (Chaddad et al., Jul. 01, 2023; Prabhakar & Won, 2023). Traditional visual inspection remains subjective and prone to error, highlighting the need for stable computational frameworks.

Preprocessing is a critical stage in EEG analysis, as noise suppression and compact signal representation are essential for reliable classification. Given the inherent complexity of EEG signals, single-stage preprocessing techniques are often insufficient to simultaneously address broadband noise and preserve diagnostically relevant temporal and spectral features. As a result, preprocessing strategies that integrate complementary mechanisms such as frequency-selective filtering and time–frequency decomposition have gained attention for improving signal fidelity. Similarly, dimensionality reduction techniques that jointly consider variance preservation and class separability enable the construction of low-dimensional representations that remain informative for subsequent analysis.

Earlier investigations have primarily addressed focal and non-focal EEG classification using individual filtering techniques and end-to-end deep learning models, with an emphasis on model-centric performance evaluation and filter ranking (Rai B et al., Nov. 2025a). Separately, hybrid filtering and hybrid dimensionality reduction approaches have been explored to enhance EEG signal representation and class separability at the preprocessing level (Rai B et al., Dec. 2025b). While these studies provide important insights into preprocessing optimization and model selection, they largely treat these components as independent stages and offer limited analysis of how optimized representations interact with feature-level information across multiple domains.

Feature extraction from EEG signals can be broadly categorized into linear, nonlinear, and hybrid approaches. Linear features primarily capture amplitude and frequency characteristics, while nonlinear features reflect complex temporal and dynamic patterns. Hybrid approaches combine both perspectives, improving classification performance. However, most existing studies consider these features independently or apply dimensionality reduction without integrating multiple feature domains, which limits their ability to exploit the complementary information inherent in EEG signals. Recent studies have highlighted the benefits of multi-domain feature integration for EEG classification. A 3D CNN–LSTM hybrid model demonstrated improved spatial-temporal feature representation for motor imagery EEG classification (Hao & Cheng, Dec. 2025). Similarly, a multiresolution feature fusion framework for EEG-based schizophrenia diagnosis showed that integrating features across multiple scales enhances discriminative capability (Ranjan & Sahana, Oct. 2024).

Although deep learning architectures have achieved remarkable accuracy in EEG classification, they typically require large-scale labeled datasets and their black-box nature can limit interpretability. This motivates alternative approaches that balance predictive performance with transparency and robustness (He et al., Dec. 17, 2021). In this context, the proposed framework integrates multi-domain feature fusion with dimensionality-reduced representations to enhance discriminative capability while maintaining interpretability across different feature domains.

The novel contribution of this work is the development of a multi-domain feature fusion framework for classifying focal and non-focal EEG signals. This framework integrates linear, nonlinear, and hybrid features with dimensionality-reduced representations, forming a comprehensive, enriched feature space that captures multiple aspects of the same signal. By leveraging this fused representation, the framework enhances discriminative capability, and interpretability, outperforming methods that rely solely on single-domain features or dimensionality reduction.

Through the systematic integration of optimized preprocessing, dimensionality reduction, and multi-domain feature fusion, the proposed framework establishes a interpretable and scalable computational approach for automated EEG classification. This methodology lays a foundation for future applications in neurodynamic modeling, cognitive studies, and computational support in epilepsy research.

2. Literature Review

EEG-based monitoring of brain electrical activity plays a crucial role in the diagnosis of epileptic disorders. In particular, distinguishing between focal and non-focal epilepsy remains a key challenge in automated EEG analysis. Recent studies have explored various preprocessing, feature extraction, dimensionality reduction, and classification techniques to improve this discrimination.

Traditional filters, such as Butterworth, Chebyshev, and notch filters, have long been employed to reduce artifacts in EEG signals. Among these, Butterworth filters are particularly favored due to their flat passband, smooth frequency response, and minimal signal distortion. EEG signals preprocessed with a Butterworth bandpass filter demonstrated improved quality, enhancing sensitivity and specificity in epilepsy classification (Saemaldahr & Ilyas, Jul. 2023). The effectiveness of Butterworth filtering was also evident in another study, where wavelet decomposition combined with a hybrid CNN model yielded improved classification performance (Hassan et al., 2022a).

Given the inherently non-stationary nature of EEG signals, time–frequency analysis methods, such as Short-Time Fourier Transform (STFT) and Wavelet Transform (WT), have proven highly effective. While conventional bandpass and notch filters remove low-frequency drifts, high-frequency noise, and power-line interference, they are often insufficient in suppressing non-stationary artifacts such as eye blinks and muscle activity (Luján et al., Dec. 01, 2021). In response to these challenges, adaptive techniques like Empirical Mode Decomposition (EMD) have been applied, decomposing signals into Intrinsic Mode Functions (IMFs) and enabling selective reconstruction by excluding noise-dominated components. The data-driven nature of EMD makes it highly suitable for analyzing non-linear and non-stationary signals (Chaddad et al., Jul. 01, 2023). Additionally, dimensionality reduction methods, including Principal Component Analysis (PCA) and t-distributed Stochastic Neighbor Embedding (t-SNE), have been employed to emphasize the most informative features (Alalayah et al., Jun. 2023). Nevertheless, challenges remain in optimizing EMD and dimensionality reduction for real-time processing and multi-channel EEG systems.

Higher-level decomposition methods, such as Independent Component Analysis (ICA) and Wavelet Packet Decomposition (WPD), have further advanced artifact removal by isolating neural activity from noise. ICA effectively eliminates components corresponding to eye blinks or muscle activity, improving signal clarity and facilitating more accurate classification (Jung et al., 2000). Similarly, wavelet-based approaches decompose EEG signals into time–frequency components, allowing selective artifact attenuation while preserving neural signal integrity (Rai et al., Jun. 01, 2025c; Bashashati et al., Jun. 01, 2007). The adaptability of time–frequency filtering has been demonstrated even in low-density electrode arrays (Liu et al., Feb. 2020). However, selecting the optimal decomposition level and threshold for different datasets remains an open problem.

Recent studies have explored hybrid denoising approaches that combine linear and nonlinear methods to achieve superior signal quality. For example, integrating STFT with bidimensional EMD significantly reduces high-frequency noise (Yan & Wu, Nov. 2022), while Hybrid Wavelet Transform (HWT) techniques, combining Discrete Wavelet Transform and Wavelet Packet Transform, effectively minimize ocular artifacts (Ferdous et al., 2021). Additional approaches, including EMD with stacked autoencoders (Sohaib et al., Oct. 2022) and discrete wavelet transform combined with Pisarenko harmonic decomposition and deformable convolutional networks (Sahoo & Mohapatra, 2022), have demonstrated further improvements in EEG denoising. Collectively, these hybrid strategies highlight the growing emphasis on combining complementary methods to enhance signal reliability in EEG-based applications and brain–computer interfaces. Yet, developing standardized hybrid frameworks that generalize across different EEG datasets remains a challenge.

Feature extraction translates raw EEG signals into discriminative representations for classification. Time-domain features such as mean, variance, and zero-crossing rate provide high discriminative power with low computational cost, making them particularly suitable for real-time epilepsy detection (Sharmila & Geethanjali, May 2020). Hjorth parameters — Activity, Mobility, and Complexity capture variance and frequency characteristics, further improving EEG classification (Al-Nafjan et al., Dec. 01, 2017). Wavelet-based feature extraction methods, including Discrete Wavelet Transform (DWT) (al-Qerem et al., Mar. 2020), Multidepth Wavelet Packets (Rafiuddin et al., 2022), and Dual-tree Complex Wavelet Transform (Al-Salman et al., Mar. 2022), have also been widely applied. Advanced wavelet features, such as peak detection and phase-space reconstruction, contribute to enhanced classification performance (Jang & Lee, Aug. 2020). However, selecting optimal features and handling the curse of dimensionality remain significant hurdles for EEG classification. EEG-based epilepsy classification has employed a variety of linear and nonlinear techniques. Radial Basis Function (RBF) models integrating linear and nonlinear features have shown strong performance (Zhou & Li, Jun. 2020), while Gradient Boosting Decision Trees (GBDT) combined with entropy-based features improved seizure prediction (Xu et al., Sep. 2022). Hurst exponent and Adaptive Fractal Analysis, in combination with GBDT and Support Vector Regression, have effectively classified complex EEG signals (Buchanna et al., Feb. 2022). Furthermore, wavelet-based features combined with SVM using RBF kernels and Particle Swarm Optimization for feature selection have demonstrated superior discrimination of epileptic patterns (Hemachandira & Viswanathan, 2022). Fusion frameworks integrating time-domain, frequency-domain, and nonlinear features have further enhanced classification accuracy (Chen et al., 2023). Despite these advances, generalization to unseen patients and cross-dataset validation remains a persistent challenge.

Time-frequency and complexity-based features have also contributed significantly to improved EEG characterization. Techniques such as quadratic time-frequency distributions (Alazrai et al., Aug. 2018), Multiscale Permutation Lempel-Ziv Complexity (MPLZC) (Borowska, Jul. 2021), Multi-scale PCA with AR modeling (Jukic et al., Sep. 2020), and Tunable Q-factor Wavelet Transform (Sadiq et al., 2021) have been proposed to capture subtle signal dynamics. Additionally, derivative-based reconstruction combined with Higuchi Fractal Dimension (HFD) features supports detection of epileptic activity (Brari & Belghith, 2021). Nonetheless, effectively integrating these complex features into real-time systems without excessive computational cost is a challenge.

Machine learning frameworks increasingly leverage hybrid feature extraction and optimization strategies. For instance, DWT combined with Binary Particle Swarm Optimization facilitates dimensionality reduction while enhancing computational efficiency (Tran et al., Nov. 2022). Deep learning architectures such as TF-HybridNet integrate STFT with 1D and 2D convolutions, improving localization of epileptic foci (Sui et al., 2021). Image-based EEG classification using Gramian Angular Summation Field representations, with SIFT and ORB features classified via Random Forests, further highlights the potential of automated EEG analysis (Krishnan et al., 2024). However, interpretability and deployment of such hybrid models in clinical settings remain challenging.

Standard classifiers, including SVM with linear kernels and KNN with City Block distance, provide reliable baselines (Mardini et al., 2020). Deep learning models, particularly 1D-CNNs, outperform ensemble models such as XGBoost, Random Forest, and TabNet in capturing temporal patterns (Hussain, Jun. 2018). For multiclass seizure classification, XGBoost effectively integrates features across time, frequency, and time–frequency domains (Kode et al., 2024). Hybrid CNN-classical frameworks balance deep feature learning with interpretability (Abirami et al., 2024), while metaheuristic optimization techniques, such as Cuckoo Search with Gaussian Mixture Models, enhance performance in complex feature spaces (Hassan et al., 2022b). Dimensionality reduction techniques like DWT, PCA, and t-SNE, when combined with Random Forest and MLP, improve signal separability and computational efficiency (Deepa et al., 2022). EEG-based Brain–Machine Interface studies further demonstrate the effectiveness of LDA, SVM, and bi-objective Invasive Weed Optimization classifiers (Tahernezhad-Javazm et al., Feb. 19, 2018). Advanced preprocessing strategies, including channel selection and deep learning-based data augmentation, continue to improve EEG-based BCI performance, although computational complexity and interpretability remain challenges (Sun & Mou, 2023). Future research should focus on scalable, interpretable frameworks that balance accuracy with computational feasibility.

3. Methodology

The proposed methodology follows a structured, multi-stage pipeline encompassing signal filtering, feature engineering, dimensionality reduction, feature fusion, classification. A schematic overview of the overall framework is presented in Figure 1, while the subsequent subsections provide a detailed description of each stage to ensure clarity and reproducibility of the proposed approach.

3.1 Dataset

The analysis in this work is based on the complete Bern-Barcelona intracranial EEG dataset (Andrzejak et al., Oct. 2012), a curated collection designed to explore the neurophysiological characteristics of focal and non-focal brain activity in epilepsy. This dataset includes a total of 7500 signal pairs—3750 pairs each for focal and non-focal categories—captured from five patients undergoing presurgical monitoring for drug-resistant epilepsy. Each pair comprises recordings from two spatially adjacent electrodes (labeled X and Y), with focal pairs obtained from regions clinically marked as part of the epileptogenic zone and non-focal pairs drawn from areas uninvolved in seizure generation. The EEG segments are 20 s in duration, recorded at a sampling rate of 512 Hz.

All signals represent interictal activity, meaning they were recorded outside of seizure episodes. The inter-electrode distance is maintained at approximately 10 millimeters, allowing the capture of localized neuronal interactions. Due to its balanced class distribution, consistent signal length, and precise anatomical mapping, the Bern-Barcelona dataset provides a strong foundation for training and evaluating EEG-based classification frameworks in epilepsy research.

The temporal characteristics of focal and non-focal EEG signals is shown in Figure 2. The top two plots depict focal EEG signals (channels X and Y), showing high-amplitude bursts and rhythmic activity indicative of epileptogenic regions. The bottom plots show non-focal EEG, characterized by smoother, more consistent waveforms.

Figure 1.

Block diagram of the EEG signal classification framework using hybrid filtering and feature fusion.

Figure 2.

Focal and non-focal EEG signals.

3.2 Preprocessing

EEG recordings are inherently susceptible to various types of noise, such as muscle artifacts, line interference, and background electrical activity. To address this, a hybrid filtering technique is employed, combining Butterworth (BW) filtering with Wavelet Packet Decomposition (WPD). The BW filter is known for its maximally flat frequency response in the passband, making it effective in removing baseline drifts and muscle artifacts without introducing sharp transitions (Savadkoohi et al., Jul. 2020). However, it lacks the ability to capture localized time-frequency patterns. To overcome this, WPD is subsequently applied, enabling the decomposition of EEG signals into multiple frequency sub-bands with fine temporal resolution (Sairamya et al., Jan. 2021; Yuan et al., Aug. 2017). This dual-stage approach enhances the signal's interpretability without compromising its physiological integrity (Rajani Rai & Martis, 2021). A comparative analysis of the raw and filtered EEG signals, as illustrated in Figure 3, indicates a noticeable reduction in high-amplitude noise fluctuations and irregular artifacts after preprocessing. The filtered signals exhibit smoother temporal variations and improved structural consistency while preserving key signal characteristics, suggesting effective noise suppression without distortion of underlying EEG dynamics. This improvement in signal stability enhances the quality of subsequent feature extraction and contributes to improved classification performance. The effects of pre and post filtering is shown in Figure 3.

Figure 3.

Filtered focal and non-focal EEG signal.

3.3 Feature Extraction

The performance of an EEG classification framework critically depends on the discriminative strength of the extracted features. Following the filtering stage, the denoised EEG signals are subjected to feature extraction to obtain a comprehensive set of representative descriptors. This process captures the intrinsic temporal and spectral characteristics of the signals, thereby enhancing the differentiation between focal and non-focal EEG patterns with improved sensitivity and clinical relevance.

In the present work, a comprehensive feature extraction strategy is employed, integrating linear, non-linear, and hybrid approaches to capture the complex dynamics of EEG signals. Linear features include statistical measures, wavelet energy coefficients, Hjorth parameters, and functional connectivity derived from coherence, collectively reflecting both temporal and spectral variations. Given the inherently non-stationary and complex nature of EEG, linear descriptors alone may not fully characterize signal dynamics. To address this, non-linear features are extracted, comprising fractal dimensions (e.g., Higuchi and Petrosian), entropy-based metrics (e.g., sample entropy, Shannon entropy, permutation entropy), graph-theoretic measures (e.g., clustering coefficient, global efficiency), and manifold learning representations via Isomap. Graph-theoretic features are computed by constructing a weighted undirected graph for each EEG segment, where the two EEG channels (X and Y) are treated as nodes and the edge weight is defined using the Pearson correlation coefficient between the two channel signals. No thresholding is applied, and the edge weights are directly obtained from the correlation values computed over the entire signal segment. These non-linear descriptors capture the intrinsic chaotic and topological properties of neural activity. Although the graph consists of two nodes, the computed graph measures are used as functional connectivity descriptors that quantify the strength of interaction between the two channels rather than large-scale network topology. The hybrid feature set synergistically combines linear descriptors, which provide interpretable physiological insight, with non-linear features that improve reliability against variability and noise, thereby forming a rich, multidimensional representation suitable for discriminating focal and non-focal EEG patterns.

Figure 4 provides a visual summary of the types of linear, non-linear, and hybrid features extracted from the pre-processed EEG signals.

Figure 4.

Overview of extracted linear, non-linear, and hybrid EEG features from pre-processed signals.

3.3.1 Linear Features

Linear features, representing the statistical and morphological aspects of EEG signals, are summarized in Table 1.

Table 1.
Linear Features Extracted from EEG Signals.

Features Equation Description

Mean (μ) $μ = \frac{1}{N} \sum_{i = 1}^{N} x_{i}$ Average amplitude; baseline brain activity.

Standard Deviation (σ) $σ = \sqrt{\frac{1}{N} \sum_{i = 1}^{N} {(x_{i} - \bar{x})}^{2}}$ Measures signal variability around the mean.

25^th and 75^th percentile $Q_{p} = x_{p \cdot n}$ Value below which p% of the EEG data points fall; captures signal distribution.

Wavelet Energy $E_{j} = \sum_{k = 1}^{N} | c_{j, k} |^{2}$ Quantifies the signal's energy distribution across wavelet subbands, reflecting spectral content.

Hjorth Parameters $Activity = \frac{1}{N} \sum_{i = 1}^{N} (x_{i} - \bar{x})^{2}$ Measures signal power, mean frequency and frequency variation

$Mobility = \sqrt{\frac{Var (d x / d t)}{Var (x)}}$

$Complexity = \frac{Mobility (d x / d t)}{Mobility (x)}$

Linear Functional Connectivity $Mean Coherence = \frac{1}{f_{2} - f_{1}} \int_{f_{1}}^{f_{2}} \frac{{| S_{x y} (f) |}^{2}}{S_{x x} (f) \cdot S_{y y} (f)} df$ Measures linear synchronization between two EEG signals within a specific frequency band, indicating functional connectivity.

Features	Equation	Description
Mean (μ)	$μ = \frac{1}{N} \sum_{i = 1}^{N} x_{i}$	Average amplitude; baseline brain activity.
Standard Deviation (σ)	$σ = \sqrt{\frac{1}{N} \sum_{i = 1}^{N} {(x_{i} - \bar{x})}^{2}}$	Measures signal variability around the mean.
25^th and 75^th percentile	$Q_{p} = x_{p \cdot n}$	Value below which p% of the EEG data points fall; captures signal distribution.
Wavelet Energy	$E_{j} = \sum_{k = 1}^{N} \| c_{j, k} \|^{2}$	Quantifies the signal's energy distribution across wavelet subbands, reflecting spectral content.
Hjorth Parameters	$Activity = \frac{1}{N} \sum_{i = 1}^{N} (x_{i} - \bar{x})^{2}$	Measures signal power, mean frequency and frequency variation
	$Mobility = \sqrt{\frac{Var (d x / d t)}{Var (x)}}$
	$Complexity = \frac{Mobility (d x / d t)}{Mobility (x)}$
Linear Functional Connectivity	$Mean Coherence = \frac{1}{f_{2} - f_{1}} \int_{f_{1}}^{f_{2}} \frac{{\| S_{x y} (f) \|}^{2}}{S_{x x} (f) \cdot S_{y y} (f)} df$	Measures linear synchronization between two EEG signals within a specific frequency band, indicating functional connectivity.

Note: Where N is the total number of data points, $x_{i}$ is the value of the i^th data point, $\bar{x}$ is the mean of the signal, p is the percentile fraction, n is the total number of sorted data points, $c_{j, k}$ is the wavelet coefficient at level j and index k, $(d x / d t)$ is the first derivative of the signal, Sxx(f)and Syy(f) are the auto-spectral densities of signals x and y, Sxy(f) is the cross-spectral density between x and y, and f1,f2 define the frequency band for mean coherence calculation.

3.3.2 Nonlinear Features

Nonlinear characteristics of EEG signals, capturing their intrinsic complexity and dynamical behavior, are summarized in Table 2.

Table 2.
Nonlinear Features Capturing the Dynamical Characteristics of EEG Signals.

Features Equation Description

Higuchi Fractal Dimension (HFD) $L_{m} (k) = \frac{1}{k} \cdot \frac{N - 1}{| \frac{N - m}{k} | \cdot k} \sum_{i = 1}^{| \frac{N - m}{k} |} | x (m + i k) - x (m + (i - 1) k) |$ Quantifies signal complexity and self-similarity using k-lagged subseries. Effective for capturing chaotic structure in EEG.

$L (k) = \frac{1}{k} \sum_{m = 1}^{k} L_{m} (k)$

HFD = slope of the regression line of log(L(k)) vs. log(1/k)

Katz Fractal Dimension (KFD) $K F D = \frac{\log_{10} (N)}{\log_{10} (\frac{d}{L}) + \log_{10} (N)}$ Measures signal complexity by considering amplitude variations and structural irregularities

$d = max | x_{i} - x_{1} | for\; i = 2, 3, \dots, n$

$L = \sum_{i = 2}^{n} | x_{i} - x_{i - 1} |$

Petrosian Fractal Dimension (PFD) $PFD = \frac{\log_{10} (N)}{\log_{10} (N) + \log_{10} (\frac{N}{N + 0.4 \times Δ})}$ Estimates complexity based on zero crossings of the first derivative, reflecting abrupt signal changes.

Sample Entropy (SampEn) $SampEn (m, r, N) = - \log (\frac{A}{B})$ Evaluates signal regularity by assessing the likelihood that similar subsequences remain similar at longer lengths. Higher values indicate greater irregularity.

Approximate Entropy (ApEn) $ApEn (m, r, N) = ϕ^{m} (r) - ϕ^{m + 1} (r)$ Quantifies unpredictability of time-series patterns. Low ApEn suggests regular signals, high ApEn indicates complex or irregular activity.

$Φ^{m} (r) = \frac{1}{N - m + 1} \sum_{i = 1}^{N - m + 1} \log (C_{i}^{m} (r))$

$C_{i}^{m} (r) = \frac{Number\; of\; j s u c h t h a t d [X_{i}^{m}, X_{j}^{m}] \leq r}{N - m + 1}$

$d [X_{i}^{m}, X_{j}^{m}] = max_{k = 1, \dots, m} | x_{i + k - 1} - x_{j + k - 1} |$

Permutation Entropy (PE) $PE = - \sum_{i = 1}^{n!} p_{i} \log (p_{i})$ Captures complexity by analyzing order relations in time-series segments. Noise tolerant and effective for detecting dynamic EEG changes.

The Average Clustering Coefficient $C = \frac{1}{N} \sum_{i = 1}^{N} \frac{2 e_{i}}{k_{i} (k_{i} - 1)}$ Measures localized connectivity in functional brain networks, indicating modular or synchronized activity

Global Efficiency ( $E_{glob})$ $E_{g l o b} = \frac{1}{N (N - 1)} \sum_{i \neq j} \frac{1}{d_{i j}}$ Reflects overall efficiency of information transfer across brain regions, with higher values showing stronger integration.

Isometric Mapping (Isomap) $Y = Q_{d} Λ_{d}^{1 / 2}$ A manifold learning method preserving intrinsic geometry of high-dimensional EEG data by embedding into a lower-dimensional space.

Features	Equation	Description
Higuchi Fractal Dimension (HFD)	$L_{m} (k) = \frac{1}{k} \cdot \frac{N - 1}{\| \frac{N - m}{k} \| \cdot k} \sum_{i = 1}^{\| \frac{N - m}{k} \|} \| x (m + i k) - x (m + (i - 1) k) \|$	Quantifies signal complexity and self-similarity using k-lagged subseries. Effective for capturing chaotic structure in EEG.
	$L (k) = \frac{1}{k} \sum_{m = 1}^{k} L_{m} (k)$
	HFD = slope of the regression line of log(L(k)) vs. log(1/k)
Katz Fractal Dimension (KFD)	$K F D = \frac{\log_{10} (N)}{\log_{10} (\frac{d}{L}) + \log_{10} (N)}$	Measures signal complexity by considering amplitude variations and structural irregularities
	$d = max \| x_{i} - x_{1} \| for\; i = 2, 3, \dots, n$
	$L = \sum_{i = 2}^{n} \| x_{i} - x_{i - 1} \|$
Petrosian Fractal Dimension (PFD)	$PFD = \frac{\log_{10} (N)}{\log_{10} (N) + \log_{10} (\frac{N}{N + 0.4 \times Δ})}$	Estimates complexity based on zero crossings of the first derivative, reflecting abrupt signal changes.
Sample Entropy (SampEn)	$SampEn (m, r, N) = - \log (\frac{A}{B})$	Evaluates signal regularity by assessing the likelihood that similar subsequences remain similar at longer lengths. Higher values indicate greater irregularity.
Approximate Entropy (ApEn)	$ApEn (m, r, N) = ϕ^{m} (r) - ϕ^{m + 1} (r)$	Quantifies unpredictability of time-series patterns. Low ApEn suggests regular signals, high ApEn indicates complex or irregular activity.
	$Φ^{m} (r) = \frac{1}{N - m + 1} \sum_{i = 1}^{N - m + 1} \log (C_{i}^{m} (r))$
	$C_{i}^{m} (r) = \frac{Number\; of\; j s u c h t h a t d [X_{i}^{m}, X_{j}^{m}] \leq r}{N - m + 1}$
	$d [X_{i}^{m}, X_{j}^{m}] = max_{k = 1, \dots, m} \| x_{i + k - 1} - x_{j + k - 1} \|$
Permutation Entropy (PE)	$PE = - \sum_{i = 1}^{n!} p_{i} \log (p_{i})$	Captures complexity by analyzing order relations in time-series segments. Noise tolerant and effective for detecting dynamic EEG changes.
The Average Clustering Coefficient	$C = \frac{1}{N} \sum_{i = 1}^{N} \frac{2 e_{i}}{k_{i} (k_{i} - 1)}$	Measures localized connectivity in functional brain networks, indicating modular or synchronized activity
Global Efficiency ( $E_{glob})$	$E_{g l o b} = \frac{1}{N (N - 1)} \sum_{i \neq j} \frac{1}{d_{i j}}$	Reflects overall efficiency of information transfer across brain regions, with higher values showing stronger integration.
Isometric Mapping (Isomap)	$Y = Q_{d} Λ_{d}^{1 / 2}$	A manifold learning method preserving intrinsic geometry of high-dimensional EEG data by embedding into a lower-dimensional space.

Note: Where N = number of points in the signal, d = maximum distance from the first point, L is total signal length, Δ = number of zero crossings, $p_{i}$ = probability of i-th permutation, d_ij = shortest path between nodes i,j, Q_d = eigenvector matrix, Λ_d = eigenvalue matrix

3.3.3 Hybrid Features

A hybrid feature set was constructed by integrating linear and nonlinear descriptors to capture both statistical properties and complex dynamical patterns of EEG signals. This composite representation improves the ability of the framework to characterize diverse neural activity. Additional parameters that further enhance discriminability are outlined in the subsequent section. Hybrid feature setare shownm in Table 3.

Table 3
Hybrid Feature Set Integrating Linear and Nonlinear Descriptors for EEG Signals.

Features Equation Description

Phase Locking Value (PLV) $P L V = | \frac{1}{N} \sum_{k = 1}^{N} e^{j (ϕ_{1} (k) - ϕ_{2} (k))} |$ Measures phase synchronization between two EEG channels by quantifying the consistency of their phase differences over time. Strong synchrony indicates functional connectivity.

Wavelet Energy $E_{j} = \sum_{k = 1}^{N} | c_{j, k} |^{2}$ Captures the distribution of signal energy across frequency bands obtained via Discrete Wavelet Transform (DWT). Reflects signal power at multiple scales.

Wavelet Entropy (WE) $p_{j} = \frac{E_{j}}{\sum_{i = 1}^{L} E_{i}}$ Quantifies the complexity or disorder of the EEG signal at multiple frequency scales using DWT coefficients. Higher entropy indicates more irregularity.

$W E = - \sum_{j = 1}^{L} p_{j} \log_{2} (p_{j})$

Manifold Learning (Isomap) $Y = Q_{d} Λ_{d}^{1 / 2}$ Extracts low-dimensional features (Components 1–5) that preserve the intrinsic geometry of the EEG signal. Captures nonlinear structures and improves discrimination of EEG patterns.

Features	Equation	Description
Phase Locking Value (PLV)	$P L V = \| \frac{1}{N} \sum_{k = 1}^{N} e^{j (ϕ_{1} (k) - ϕ_{2} (k))} \|$	Measures phase synchronization between two EEG channels by quantifying the consistency of their phase differences over time. Strong synchrony indicates functional connectivity.
Wavelet Energy	$E_{j} = \sum_{k = 1}^{N} \| c_{j, k} \|^{2}$	Captures the distribution of signal energy across frequency bands obtained via Discrete Wavelet Transform (DWT). Reflects signal power at multiple scales.
Wavelet Entropy (WE)	$p_{j} = \frac{E_{j}}{\sum_{i = 1}^{L} E_{i}}$	Quantifies the complexity or disorder of the EEG signal at multiple frequency scales using DWT coefficients. Higher entropy indicates more irregularity.
	$W E = - \sum_{j = 1}^{L} p_{j} \log_{2} (p_{j})$
Manifold Learning (Isomap)	$Y = Q_{d} Λ_{d}^{1 / 2}$	Extracts low-dimensional features (Components 1–5) that preserve the intrinsic geometry of the EEG signal. Captures nonlinear structures and improves discrimination of EEG patterns.

Where N is the total number of EEG samples, k is the sample index, ϕ1(k)ϕ2(k) are the instantaneous phases of two signals, $c_{j, k}$ is the wavelet coefficient at level j, E_j is the wavelet energy at level j

3.4 Classification

Different classifiers were utilized to evaluate the performance of the extracted and augmented feature sets. The models include traditional supervised approaches, ensemble-based methods, and hybrid combinations that integrate multiple learning strategies. The stacked ensemble classifier was implemented using a cross-validated stacking strategy. The dataset was divided using a stratified train–test split, and feature selection was performed on the training data using Recursive Feature Elimination with Cross-Validation (RFECV) with stratified cross-validation. The stacking classifier internally generates out-of-fold predictions from the base learners to train the meta-learner, ensuring that the meta-learner is trained on predictions from unseen data and preventing data leakage. Hyperparameters of the individual classifiers were selected based on cross-validation during model development to ensure optimal performance and generalization.

Table 4 summarizes each classifier with a brief model description, its corresponding learning type, and the key hyperparameters applied during training.

Table 4.
Summary of Classifiers with Model Descriptions, Learning Types, and Hyperparameters Used.

Classifier Model Description Learning Type Hyperparameters Used

SVM A margin-based classifier that identifies optimal separating hyperplane in high-dimensional space. The RBF kernel was used to model non-linear decision boundary (Savadkoohi et al., Jul. 2020). Supervised (Margin-based) Kernel: RBF; C = 1; Gamma: scale; probability enabled.

Random Forest (RF) An ensemble of decision trees trained on bootstrap samples with randomized feature subsets. RF handles noisy data and limits overfitting (Suárez et al., Mar. 2025) Ensemble (Bagging) Number of trees: 100; bootstrapped sampling; random feature selection.

AdaBoost AdaBoost is a sequential ensemble learning algorithm that combines multiple weak learners by iteratively adjusting sample weights, emphasizing misclassified instances in subsequent iterations. The final model aggregates the weighted predictions of all learners to form a strong classifier (Al-Hadeethi et al., Dec. 2021). Ensemble (Boosting) Estimators: 50; learning rate: 1.0; random state: 42.

XGBoost A highly effective gradient boosting model based on constructing additive models with regularized trees (Panigrahi et al., Feb. 2022). It is very accurate with the ability to resist overfitting. Gradient Boosting Learning rate: 0.1; max depth: 3; estimators: 100.

SVM + RF A mixed classifier that uses the margin maximum characteristics of SVM and then fusing the decision level using Random Forest to enhance reliability (Polat & Nour, Dec. 2020). Hybrid (Margin + Bagging) Soft voting; SVM (RBF kernel); RF (100 trees).

XGBoost + RF Combines the strengths of the regularized boosting of the XGBoost and the diversity of the ensemble to acquire increased generalization (Kode et al., 2024). Hybrid Ensemble XGBoost (depth = 3); RF (100 trees).

XGBoost + AdaBoost A finer integrated composite model which combines gradient boosting and adaptive weighting to focus on hard samples. Boosting Hybrid 50 adaptive iterations; shallow boosted trees.

Stacked Classifier A meta-ensemble framework that combines the output of multiple base learner's using a meta-classifier, while capturing diverse learning patterns across classifiers (Anuragi et al., Jan. 2022). Meta-Learning Ensemble Base: SVM, RF, AdaBoost; Meta: Logistic Regression.

Classifier	Model Description	Learning Type	Hyperparameters Used
SVM	A margin-based classifier that identifies optimal separating hyperplane in high-dimensional space. The RBF kernel was used to model non-linear decision boundary (Savadkoohi et al., Jul. 2020).	Supervised (Margin-based)	Kernel: RBF; C = 1; Gamma: scale; probability enabled.
Random Forest (RF)	An ensemble of decision trees trained on bootstrap samples with randomized feature subsets. RF handles noisy data and limits overfitting (Suárez et al., Mar. 2025)	Ensemble (Bagging)	Number of trees: 100; bootstrapped sampling; random feature selection.
AdaBoost	AdaBoost is a sequential ensemble learning algorithm that combines multiple weak learners by iteratively adjusting sample weights, emphasizing misclassified instances in subsequent iterations. The final model aggregates the weighted predictions of all learners to form a strong classifier (Al-Hadeethi et al., Dec. 2021).	Ensemble (Boosting)	Estimators: 50; learning rate: 1.0; random state: 42.
XGBoost	A highly effective gradient boosting model based on constructing additive models with regularized trees (Panigrahi et al., Feb. 2022). It is very accurate with the ability to resist overfitting.	Gradient Boosting	Learning rate: 0.1; max depth: 3; estimators: 100.
SVM + RF	A mixed classifier that uses the margin maximum characteristics of SVM and then fusing the decision level using Random Forest to enhance reliability (Polat & Nour, Dec. 2020).	Hybrid (Margin + Bagging)	Soft voting; SVM (RBF kernel); RF (100 trees).
XGBoost + RF	Combines the strengths of the regularized boosting of the XGBoost and the diversity of the ensemble to acquire increased generalization (Kode et al., 2024).	Hybrid Ensemble	XGBoost (depth = 3); RF (100 trees).
XGBoost + AdaBoost	A finer integrated composite model which combines gradient boosting and adaptive weighting to focus on hard samples.	Boosting Hybrid	50 adaptive iterations; shallow boosted trees.
Stacked Classifier	A meta-ensemble framework that combines the output of multiple base learner's using a meta-classifier, while capturing diverse learning patterns across classifiers (Anuragi et al., Jan. 2022).	Meta-Learning Ensemble	Base: SVM, RF, AdaBoost; Meta: Logistic Regression.

3.5 Two-Stage Dimensionality Reduction

High dimensional feature space often contains redundant features that can impair classifier performance and increase computational cost. To address this, dimensionality reduction techniques were applied to transform the feature space into a lower-dimensional representation. A sequential dimensionality reduction framework combining Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA) were adopted to enhance model performance. PCA was initially applied to reduce noise and redundancy by projecting the data onto a lower-dimensional space that captures the most significant variance. The number of principal components retained after PCA was set to 100 to preserve the most significant variance while reducing dimensionality. Since PCA does not utilize class label information, it may overlook features that are critical for class discrimination. To address the above, LDA was subsequently applied to the PCA-transformed data to maximize class separability by optimizing the ratio of inter-class to intra-class variance. As this work involves binary classification, LDA produces a single discriminant component, resulting in a one-dimensional discriminative subspace. This hybrid approach leverages the strength of both unsupervised and supervised techniques. Thus PCA results in compression and noise filtering, while LDA enhances class distinction. A comparative evaluation of feature sets with and without dimensionality reduction indicates a consistent improvement in classification performance when the PCA–LDA framework is applied. This improvement is evident across multiple classifiers, where DR-based configurations demonstrate enhanced class separability compared to their non-reduced counterparts. The reduction of feature redundancy through PCA, followed by discriminative projection via LDA, results in a more compact and informative feature representation, which facilitates improved decision boundaries during classification. Furthermore, the reduced dimensionality contributes to improved computational efficiency and model stability, particularly in high-dimensional EEG feature spaces where redundancy and noise can adversely affect performance.

To prevent data leakage, PCA was fitted using only the training data and the learned transformation was applied to the test data; similarly, LDA was fitted using the PCA-transformed training data and corresponding labels, and the learned projection was applied to the test data. The combined PCA–LDA pipeline not only improves computational efficiency and numerical stability but also consistently yields superior results in EEG-based classification tasks by preserving discriminative structure in a reduced feature space. The resulting PCA–LDA subspace represents a noise-reduced and class-discriminative embedding, which was fused with the extracted feature sets to improve class separability and overall classification performance.

3.6 Data Fusion

To enhance class separability, a Dimensionality Reduced–Fused Feature strategy integrates multi-domain descriptors with their low-dimensional embeddings, yielding a compact representation that preserves essential EEG dynamics while improving generalization across classifiers.

Unlike conventional fusion schemes that rely on either raw EEG signals or isolated feature domains, the present framework integrates dimensionality reduction with multi-domain feature representations to achieve enhanced discriminative capability. EEG signals were first denoised using a hybrid Butterworth–WPD filter, after which three independent feature sets—linear, nonlinear, and hybrid were extracted. In parallel, a combined PCA–LDA dimensionality reduction process was applied to the filtered signals to generate a compact subspace optimized for class separability. This reduced representation was subsequently fused with each feature domain, forming three distinctive Dimensionality Reduced–Fused (DR–Fused) feature sets: Linear–DR, Nonlinear–DR, and Hybrid–DR. The fusion process retained domain-specific characteristics while embedding low-dimensional discriminative structure, thereby enriching the feature space with complementary, non-redundant information. Both the original and fused features were evaluated using multiple classifiers to ensure generalizable differentiation between focal and non-focal EEG patterns.

The last phase uses the same classifiers as in Section 3.4, to test the effect of dimensionality reduction and fusion on classification accuracy. Since there is uniformity in the classification models, it is possible to directly compare the results of both the pre-fusion and post-fusion feature sets.

3.7 Experimental Setup and Implementation Details

All experiments were implemented in Python using scientific computing and machine learning libraries, including NumPy, SciPy, Pandas, PyWavelets, Scikit-learn, etc. The experiments were conducted on a system with an Intel Core i7 processor and 16 GB RAM. The proposed framework is based on feature engineering and classical machine learning models; therefore, GPU acceleration was not required.

Signal preprocessing was performed using a Butterworth bandpass filter (0.5–50 Hz) followed by Wavelet Packet Decomposition (WPD) using the Daubechies wavelet (db6) up to level 7 for denoising. Linear, nonlinear, and hybrid features were extracted from the filtered EEG signals. Dimensionality reduction was performed using Principal Component Analysis (PCA) followed by Linear Discriminant Analysis (LDA), and the reduced features were fused with the original feature domains to construct fused feature sets.

The dataset was divided into training and testing sets using stratified random sampling to maintain class balance, with 80% of the data used for training and 20% used for testing. To ensure stability and reduce bias due to random partitioning, Stratified K-Fold cross-validation (k = 5) was performed on the training set during model development. Feature scaling was performed using standardization prior to classification. Multiple classifiers, including individual, ensemble, hybrid, and stacked models, were used to evaluate the effectiveness of the proposed feature fusion framework. The hyperparameters used for each classifier are summarized in Table 5. Model performance was evaluated using accuracy, sensitivity, specificity, F1-score, AUC, and Brier score.

Table 5.
Experimental Setup and Parameter Configuration.

Stage Method Parameters

Signal Filtering Butterworth Bandpass Filter Order = 5, Frequency range = 0.5–50 Hz

Denoising Wavelet Packet Decomposition Wavelet = db6, Level = 7

Manifold Learning Isomap n_neighbors = 10, n_components = 5

Dimensionality Reduction PCA n_components = 100

Dimensionality Reduction LDA n_components = 1

Data Splitting Stratified Train–Test Split Train = 80%, Test = 20%, random_state = 42

SVM RBF Kernel C = 10, Gamma = 0.01

Random Forest Ensemble n_estimators = 300, max_depth = 25

AdaBoost Boosting n_estimators = 200, learning_rate = 0.4

XGBoost Gradient Boosting n_estimators = 300, max_depth = 10, learning_rate = 0.05

SVM + RF Soft Voting Ensemble Base learners: SVM and Random Forest

XGB + RF Soft Voting Ensemble Base learners: XGBoost and Random Forest

XGB + AdaBoost Soft Voting Ensemble Base learners: XGBoost and AdaBoost

Stacked Classifier Stacking Ensemble Base learners: SVM, AdaBoost, RF, XGBoost, LightGBM, CatBoost; Meta-learner: Logistic Regression

Stage	Method	Parameters
Signal Filtering	Butterworth Bandpass Filter	Order = 5, Frequency range = 0.5–50 Hz
Denoising	Wavelet Packet Decomposition	Wavelet = db6, Level = 7
Manifold Learning	Isomap	n_neighbors = 10, n_components = 5
Dimensionality Reduction	PCA	n_components = 100
Dimensionality Reduction	LDA	n_components = 1
Data Splitting	Stratified Train–Test Split	Train = 80%, Test = 20%, random_state = 42
SVM	RBF Kernel	C = 10, Gamma = 0.01
Random Forest	Ensemble	n_estimators = 300, max_depth = 25
AdaBoost	Boosting	n_estimators = 200, learning_rate = 0.4
XGBoost	Gradient Boosting	n_estimators = 300, max_depth = 10, learning_rate = 0.05
SVM + RF	Soft Voting Ensemble	Base learners: SVM and Random Forest
XGB + RF	Soft Voting Ensemble	Base learners: XGBoost and Random Forest
XGB + AdaBoost	Soft Voting Ensemble	Base learners: XGBoost and AdaBoost
Stacked Classifier	Stacking Ensemble	Base learners: SVM, AdaBoost, RF, XGBoost, LightGBM, CatBoost; Meta-learner: Logistic Regression

The computational cost of the proposed framework arises from multiple stages, including preprocessing, feature extraction, dimensionality reduction, and classification. The preprocessing stage involves Butterworth filtering and Wavelet Packet Decomposition (WPD), which are computationally optimised signal processing operations and are performed once per signal. The feature extraction stage includes linear, nonlinear, and graph-theoretic features, which introduce moderate computational complexity due to entropy, fractal, and connectivity calculations. However, these features provide highly discriminative information, reducing the need for very complex classifiers. The classification stage includes single, hybrid, and stacked models. Among these, stacked and hybrid models require higher computational time compared to single classifiers due to the combination of multiple learning algorithms. However, the use of dimensionality reduction significantly reduces the feature space, which helps in reducing the overall training time and computational complexity.

Overall, the proposed framework does not require GPU acceleration and can be implemented on a standard computing system. Compared to deep learning-based approaches that require large training data and high computational resources, the proposed method is computationally efficient while still achieving high classification performance, making it suitable for practical EEG analysis applications.

4. Results and Discussion

This section presents a critical evaluation of the classification framework developed for distinguishing focal from non-focal EEG signals. The analysis is structured according to the sequential stages of the pipeline, highlighting the influence of hybrid filtering approaches, the benefits conferred by dimensionality reduction, the role of feature extraction, and the impact of multi-domain data fusion on classifier performance. Observations emphasize how the integration of these components contributes to enhanced signal discrimination and classification process.

4.1 Feature Visualization of Filtered EEG Signals

To gain a comprehensive understanding of the nonlinear dynamics inherent in the filtered EEG signals, visual feature representations were utilized. These representations offer complementary insights into the temporal and dynamical characteristics of the system. In particular, the Poincaré Section and Recurrence Plot analyses were employed to elucidate distinctive spatiotemporal patterns corresponding to focal and non-focal EEG activities, thereby facilitating a more nuanced characterization of signal dynamics.

4.1.1 Poincaré Section Representation (PSR)

To characterize the nonlinear dynamics of EEG signals beyond conventional time- and frequency-domain analyses, Phase Space Reconstruction (PSR) was applied. PSR provides a geometrical depiction of signal evolution by mapping trajectories in a two-dimensional phase plane, capturing stability, periodicity, and potential chaotic behavior.

As illustrated in Figure 5, focal EEG signals exhibit widely dispersed trajectories, reflecting higher dynamical variability and the presence of irregular events such as bursts or spikes, which are typically associated with epileptogenic regions. In contrast, non-focal signals demonstrate compact, clustered trajectories indicative of stable and periodic dynamics. A quantitative comparison of trajectory dispersion reveals that focal EEG signals exhibit significantly higher spread, with visibly larger spatial occupation in the phase space compared to non-focal signals, whose trajectories remain confined within a narrow region. This increased dispersion indicates higher variance and complexity in focal signals, whereas the tightly clustered structure of non-focal signals reflects lower variability and more regular dynamics. Furthermore, the presence of irregular loops and extended trajectory deviations in focal signals, as opposed to the smoother and more consistent patterns observed in non-focal signals, highlights the difference in underlying dynamical behavior. These distinct visual features—trajectory divergence, loop formations, and clustering—serve as markers for discriminating between focal and non-focal EEG activity, highlighting the utility of PSR for enhancing classification performance.

Figure 5.

Poincaré section representation of filtered EEG signals focal (left) and non-focal(right).

Figure 6.

3D recurrence plot visualization of filtered EEG signals (a) focal (b) non-focal.

4.1.2 Recurrence Plot Visualization

To further investigate the temporal recurrence characteristics of EEG dynamics, Recurrence Plot (RP) analysis was employed. RP identifies points at which the system revisits similar states, providing insight into periodicity and chaotic behavior. Figure 6 presents three-dimensional recurrence surfaces for focal and non-focal EEG segments. Focal signals display irregular, fragmented recurrence patterns, indicative of highly unstable and chaotic transitions commonly associated with epileptogenic activity. In contrast, non-focal signals exhibit structured, repetitive patterns, reflecting stable and periodic dynamics. A quantitative interpretation of the recurrence structure reveals that focal EEG signals exhibit lower recurrence uniformity and higher variability, as evidenced by the uneven distribution and discontinuity of recurrence intensities across the surface. In comparison, non-focal signals demonstrate higher recurrence density and consistency, with more uniformly distributed recurrence values indicating stable temporal evolution. The increased fragmentation and irregularity observed in focal signals correspond to higher dynamical complexity, whereas the smoother and more continuous recurrence patterns in non-focal signals reflect lower complexity and stronger periodic behavior. These visual distinctions underscore the greater dynamical complexity of focal EEG signals and highlight RP analysis as a valuable tool for discriminating between focal and non-focal activity, thereby supporting improved classification outcomes.

4.2 Classifier Output

This section presents a detailed analysis of the experimental outcomes derived from the proposed EEG classification framework. Classification was performed at two critical stages: (i) immediately after feature extraction and (ii) following the application of dimensionality reduction and feature fusion. To ensure consistency and comparability, identical classifiers were employed across both stages. A total of eight classifiers—comprising traditional (SVM), ensemble (AdaBoost, Random Forest, XGBoost), and hybrid models (SVM + RF, XGB + RF, XGB + AdaBoost, and Stacked)—were evaluated on six feature configurations, namely Linear, Nonlinear, Hybrid, and their respective dimensionality-reduced (DR) fused counterparts. To enhance model robustness and eliminate redundant features, RFECV was applied prior to classification. Model evaluation was performed using stratified 5-fold cross-validation to ensure balanced class representation across folds. The variation in performance across folds was minimal, with deviations observed in the range of approximately 0.001–0.008, indicating stable and consistent model performance. Performance evaluation was conducted using accuracy, sensitivity, specificity, F1-score, AUC, and Brier score as key metrics.

For clarity and brevity, Table 6 reports only the results of the best-performing classifiers, including XGBoost (individual), XGB + AdaBoost (hybrid), and the Stacked Classifier (meta-ensemble), to illustrate the impact of feature complexity, dimensionality reduction, and model architecture on classification performance.

Table 6.
Performance Comparison of Classification Models Using Various Feature Sets (Pre- and Post- Fusion).

Classifier Type Classifier Feature Set ACC (%) SEN SPE F1-Score AUC Brier Score

Base/ Individual Classifier XGBoost Linear 82.500 0.8040 0.8460 0.8212 0.8962 0.1250

Non-Linear 90.23 0.9000 0.9046 0.9021 0.9692 0.0710

Hybrid 93.36 0.9366 0.9306 0.9338 0.9525 0.0503

Linear + DR 99.03 0.9986 0.9940 0.9963 0.9136 0.0051

Non-Linear + DR 99.50 0.9960 0.9940 0.9950 0.9777 0.0051

Hybrid + DR 97.73 0.9786 0.9760 0.9773 0.9623 0.0182

Hybrid Classifier XGB + AdaBoost Linear 82.467 0.8040 0.8453 0.8209 0.9037 0.1437

Non-Linear 90.23 0.9000 0.9046 0.9021 0.9589 0.1109

Hybrid 93.33 0.9360 0.9306 0.9335 0.9421 0.0969

Linear + DR 99.03 0.9986 0.9940 0.9963 0.9265 0.0700

Non-Linear + DR 99.50 0.9960 0.9940 0.9950 0.9877 0.0692

Hybrid + DR 97.70 0.9786 0.9753 0.9770 0.9655 0.0772

Meta-ensemble Stacked classifier Linear 83.20 0.8206 0.8453 0.8300 0.9143 0.1200

Non-Linear 92.40 0.9180 0.9300 0.9235 0.9856 0.0560

Hybrid 94.53 0.9420 0.9486 0.9451 0.9673 0.0426

Linear + DR 98.99 0.9993 0.9980 0.9986 0.9382 0.0011

Non-Linear + DR 99.80 0.9986 0.9973 0.9980 0.9982 0.0015

Hybrid + DR 98.16 0.9806 0.9826 0.9816 0.9788 0.0139

Classifier Type	Classifier	Feature Set	ACC (%)	SEN	SPE	F1-Score	AUC	Brier Score
Base/ Individual Classifier	XGBoost	Linear	82.500	0.8040	0.8460	0.8212	0.8962	0.1250
Non-Linear	90.23	0.9000	0.9046	0.9021	0.9692	0.0710
Hybrid	93.36	0.9366	0.9306	0.9338	0.9525	0.0503
Linear + DR	99.03	0.9986	0.9940	0.9963	0.9136	0.0051
Non-Linear + DR	99.50	0.9960	0.9940	0.9950	0.9777	0.0051
Hybrid + DR	97.73	0.9786	0.9760	0.9773	0.9623	0.0182
Hybrid Classifier	XGB + AdaBoost	Linear	82.467	0.8040	0.8453	0.8209	0.9037	0.1437
Non-Linear	90.23	0.9000	0.9046	0.9021	0.9589	0.1109
Hybrid	93.33	0.9360	0.9306	0.9335	0.9421	0.0969
Linear + DR	99.03	0.9986	0.9940	0.9963	0.9265	0.0700
Non-Linear + DR	99.50	0.9960	0.9940	0.9950	0.9877	0.0692
Hybrid + DR	97.70	0.9786	0.9753	0.9770	0.9655	0.0772
Meta-ensemble	Stacked classifier	Linear	83.20	0.8206	0.8453	0.8300	0.9143	0.1200
		Non-Linear	92.40	0.9180	0.9300	0.9235	0.9856	0.0560
		Hybrid	94.53	0.9420	0.9486	0.9451	0.9673	0.0426
		Linear + DR	98.99	0.9993	0.9980	0.9986	0.9382	0.0011
		Non-Linear + DR	99.80	0.9986	0.9973	0.9980	0.9982	0.0015
		Hybrid + DR	98.16	0.9806	0.9826	0.9816	0.9788	0.0139

Table 6 summarizes the classification performance of the three best-performing models-XGBoost, XGB + AdaBoost, and the Stacked Classifier—evaluated across six distinct feature configurations. The results reveal a consistent performance improvement when dimensionality reduction (DR) is applied to fused feature sets, highlighting its role in enhancing discriminative capability by minimizing feature redundancy while retaining essential information.

The Non-linear + DR and Hybrid + DR feature configurations achieve near-optimal outcomes across all evaluation metrics. The Non-linear + DR feature set in particular attains the highest overall performance, with the stacked ensemble achieving an accuracy of 99.80%, AUC of 0.99, and the lowest Brier score (0.015). This superior result indicates excellent calibration, reliability, and generalization, affirming that the combination of nonlinear features with dimensionality reduction best captures the intrinsic chaotic and complex dynamics of EEG signals.

The XGB + AdaBoost hybrid model also demonstrates strong discriminative performance (ACC = 99.50%, SEN = 0.996, SPE = 0.994, Brier = 0.0692), but the Stacked Classifier marginally outperforms it across all metrics. Although the numerical difference is small, the stacked model exhibits enhanced stability benefiting from the integration of complementary decision boundaries from multiple base learners.

Linear feature sets, in contrast, produce relatively modest results (<84% accuracy) and higher Brier scores, reflecting their limited ability to capture the nonlinear variability of EEG dynamics. Hybrid and nonlinear feature domains, particularly when fused with DR, exhibit substantial gains in sensitivity, specificity, and F1-scores, confirming their superior discriminative strength.

Overall, the findings conclusively demonstrate that feature fusion integrated with dimensionality reduction yields the most reliable and high-performing EEG classification outcomes, with the Non-linear + DR configuration emerging as the most effective representation and the Stacked Classifier providing the most consistent and generalizable performance across all feature types.

Figure 7 illustrates the comparative classification accuracies achieved across eight classifiers for all six feature configurations. A clear trend is observed wherein feature sets fused with dimensionality reduction (DR) consistently outperform their non-fused counterparts, reaffirming the effectiveness of DR in enhancing discriminative representation. Specifically, DR-based feature configurations demonstrate an improvement of approximately 3–7% in accuracy across most classifiers when compared to their corresponding non-DR variants, indicating a consistent enhancement in model performance. For instance, in the case of ensemble classifiers such as XGBoost and Stacked models, hybrid feature sets combined with DR achieve peak accuracies approaching 99%, whereas the corresponding non-DR configurations remain below this range. Notably, configurations integrating nonlinear and hybrid features with DR achieve the highest accuracies across nearly all classifiers, demonstrating superior adaptability to the complex, nonstationary nature of EEG dynamics.

Figure 7.

Comparative accuracy of different classifiers across feature configurations.

Among the evaluated classifiers, ensemble-based methods (XGBoost, XGB + RF, and Stacked) consistently outperform individual classifiers, achieving accuracy gains of approximately 4–8% over baseline models such as AdaBoost and Random Forest. This trend indicates that the combination of feature fusion and ensemble learning effectively captures both linear and nonlinear EEG characteristics, leading to improved classification robustness. While classification accuracy provides an initial measure of model capability, it does not fully capture aspects such as class-specific sensitivity or calibration reliability—factors critical to EEG-based clinical applications. Therefore, to obtain a more comprehensive assessment, additional evaluation metrics including sensitivity, specificity, F1-score, AUC, and Brier score were incorporated.

To enable fair cross-metric comparison, all metric values were normalized using min–max scaling between 0 and 1, ensuring that differences in scale did not bias interpretation. Subsequently, the normalized metrics were aggregated to compute an overall normalized score for each classifier–feature combination, forming the basis for ranking and further visualization through a composite heatmap (Figure 8). This approach provides a balanced and integrative view of classifier behavior, highlighting those configurations that maintain consistent performance across diverse evaluation criteria.

Figure 8.

Heatmap of normalized performance scores of feature sets across classifiers.

Figure 8 presents a heatmap illustrating the normalized, multi-metric performance across all classifiers and feature configurations. The visualization provides an integrative perspective of classifier behavior, highlighting subtle yet meaningful variations that may not be evident from individual metrics alone. A quantitative inspection of the heatmap reveals that feature configurations incorporating dimensionality reduction (DR) consistently achieve higher normalized scores, typically in the range of 0.93 to 0.99, compared to non-DR configurations which generally remain below 0.90. Among the fused feature configurations, the Non-linear + DR set demonstrates the most consistent and superior performance, maintaining high normalized scores across nearly all classifiers. In particular, the Stacked classifier achieves the highest performance with a peak normalized score of approximately 0.998 under the Non-linear + DR configuration, followed closely by XGBoost and XGB-based ensemble models with scores exceeding 0.98. The Hybrid + DR configuration follows closely, displaying stable performance—particularly when coupled with ensemble and hybrid learning models.

Although the Linear + DR feature set achieves competitive results with classifiers such as XGBoost and the stacked ensemble, its performance exhibits greater variability across other models, suggesting classifier-dependent sensitivity. The heatmap thus reinforces earlier observations from the quantitative analysis, confirming that Non-linear + DR and Hybrid + DR configurations provide the most reliable and generalizable feature representations. The normalization process ensures an equitable evaluation across all metrics, allowing the heatmap to clearly emphasize configurations that maintain balanced excellence in accuracy, sensitivity, specificity, and calibration quality.

Table 7 summarizes the averaged normalized performance scores and corresponding ranks obtained for all evaluated feature set combinations. The Non-linear + DR configuration achieved the highest normalized score (0.95), followed closely by the Hybrid + DR configuration (0.94). Although these two feature sets appear visually comparable in the heatmap (Figure 8), the quantitative ranking confirms the superior discriminative capacity of nonlinear descriptors when integrated with dimensionality reduction. The Hybrid + DR feature set also demonstrated highly competitive performance, suggesting that combining linear and nonlinear feature domains enhances the overall representational richness of EEG dynamics.

Table 7.

Averaged Normalized Performance Scores and Ranks of Feature Sets Across Classifiers.

Feature Set	Normalized Score	Rank
Non-Linear + DR	0.95	1
Hybrid + DR	0.94	2
Hybrid	0.92	3
Linear + DR	0.91	4
Non-Linear	0.90	5
Linear	0.83	6

In contrast, the Linear feature set yielded the lowest normalized score (0.83), reinforcing its limited ability to capture the complex, nonstationary characteristics of EEG signals. These findings highlight the critical role of selecting appropriate feature representations and fusion strategies to achieve reliable discrimination between focal and non-focal EEG patterns.

Among all evaluated combinations, the Non-linear + DR configuration consistently emerged as the most stable and robust performer across multiple evaluation metrics, closely followed by Hybrid + DR. The integration of dimensionality reduction not only improved accuracy but also enhanced calibration and generalization, leading to more dependable classification outcomes. Furthermore, ensemble-based architectures—particularly the Stacked, XGB + RF, and XGBoost classifiers—consistently delivered superior results, demonstrating their strength in exploiting the structured variability of the reduced nonlinear and hybrid feature spaces.

4.3 Statistical Validation of Classification Performance

To evaluate the statistical reliability of the proposed EEG classification framework, stratified cross-validation was performed and the performance metrics were analyzed using statistical measures. Table 8 presents the classification performance in terms of mean, standard deviation, and 95% confidence intervals for the best performing featureset- Nonlinear + DR. The results show very small standard deviation values (approximately 0.0005–0.008), indicating that the model performance is stable and not dependent on a specific data split. The narrow confidence intervals further confirm the reliability and consistency of the proposed method.

Table 8.
Cross-Validation Performance with Confidence Interval.

Model Accuracy (Mean ± Std) 95% Confidence Interval

SVM 0.8113 ± 0.0051 ±0.0045

AdaBoost 0.7977 ± 0.0077 ±0.0067

Random Forest 0.8883 ± 0.0064 ±0.0056

XGBoost 0.9985 ± 0.0005 ±0.0004

SVM + RF 0.8541 ± 0.0041 ±0.0036

XGB + RF 0.9894 ± 0.0018 ±0.0015

Stacked 0.8821 ± 0.0042 ±0.0037

Model	Accuracy (Mean ± Std)	95% Confidence Interval
SVM	0.8113 ± 0.0051	±0.0045
AdaBoost	0.7977 ± 0.0077	±0.0067
Random Forest	0.8883 ± 0.0064	±0.0056
XGBoost	0.9985 ± 0.0005	±0.0004
SVM + RF	0.8541 ± 0.0041	±0.0036
XGB + RF	0.9894 ± 0.0018	±0.0015
Stacked	0.8821 ± 0.0042	±0.0037

Furthermore, paired statistical tests were conducted across cross-validation folds to compare classifier performance. The statistical analysis indicated that the performance differences between classifiers are statistically significant (p < 0.05) in most cases, demonstrating that the observed performance improvements are not due to random variation but are statistically meaningful. Overall, the statistical results confirm that the proposed feature extraction and feature fusion framework provides stable and statistically reliable classification performance.

4.4 Impact of PCA + LDA Fusion on Classification Performance

The key novelty of this work lies in the fusion of dimensionally reduced features (PCA + LDA) with the original Linear, Nonlinear, and Hybrid feature sets. Rather than relying solely on the original or reduced representations, the proposed framework reintegrates the reduced components into their corresponding feature spaces, forming enriched composite sets: Linear + DR, Nonlinear + DR, and Hybrid + DR. This design synergistically combines the variance-preserving nature of PCA and the class-discriminative strength of LDA with the intrinsic statistical and dynamical characteristics of each feature domain. Consequently, the fused representations provide a compact yet information-rich description of EEG signal behavior, improving both separability and stability. A comparative analysis between pre-fusion and post-fusion feature sets reveals a consistent improvement in classification performance across all classifiers, indicating the effectiveness of the proposed fusion strategy.

Figure 9 illustrates the normalized classification performance of all feature sets before and after fusion. As evident from the figure, the post-fusion feature sets consistently outperform their pre-fusion counterparts across all classifiers. Quantitatively, the fused feature sets demonstrate an improvement of approximately 3–6% in accuracy compared to their corresponding pre-fusion representations, with the most significant gains observed in nonlinear and hybrid configurations. The Nonlinear + DR and Hybrid + DR configurations exhibit the most pronounced improvements, demonstrating that the integration of reduced representations significantly enhances discriminative capacity.

Figure 9.

Comparison of pre-fusion and post-fusion feature set performance.

Furthermore, ensemble-based classifiers such as XGBoost, XGBoost + RF, XGBoost + AdaBoost, and Stacked show a substantial increase in performance following fusion. In particular, ensemble models consistently achieve higher accuracy after fusion, indicating that the enriched feature representation enables better learning of complex decision boundaries. This improvement highlights the complementary nature of original and reduced features, where fusion effectively reduces redundancy while preserving discriminative information. This observation confirms the efficacy of the proposed fusion strategy in stabilizing classification outcomes, reducing redundancy, and capturing the complex, nonstationary dynamics inherent in EEG signals. Overall, the results validate that the fusion of dimensionality reduction with original feature domains yields superior, balanced, and generalizable performance, setting a promising direction for EEG-based focal versus non-focal signal classification.

4.5 Mitigation of Overfitting

To minimize the risk of overfitting, the performance of the optimal stacked model was further validated through Receiver Operating Characteristic (ROC) curve analysis and confusion matrix evaluation. The ROC curve quantifies the model's discriminative capability across varying thresholds, whereas the confusion matrix provides a detailed representation of correct and misclassified instances, offering complementary insight into classification reliability.

4.5.1 Receiver Operating Characteristic (ROC) Curve

The Receiver Operating Characteristic (ROC) curve serves as a graphical tool to assess a classifier's capability to differentiate between classes across various decision thresholds. It plots the True Positive Rate (sensitivity) against the False Positive Rate (1 – specificity), thereby offering an aggregate view of classification performance. Figure 10 illustrates the ROC curves obtained for all feature sets using the best-performing stacked model. Among these, the dimensionality-reduced fused features exhibited the highest Area Under the Curve (AUC) value of 0.9982 for the Nonlinear + DR configuration, demonstrating superior discriminative strength. The remaining feature sets also achieved consistently high AUCs, ranging from 0.912 to 0.952, underscoring the model's reliability in accurately distinguishing focal from non-focal EEG signals across diverse feature representations.

Figure 10.

ROC curves for the feature sets using the stacked model.

Figure 11.

Confusion matrices for the feature sets using the stacked model.

These findings indicate that the model's superior performance results from the combined effect of feature fusion and dimensionality reduction rather than overfitting. By integrating complementary neurodynamic characteristics and minimizing redundant information, the model generalizes effectively to unseen EEG data. The corresponding confusion matrix analysis further supports this observation by highlighting consistently high true positive rates and minimal misclassifications across feature sets.

4.5.2 Confusion Matrix

The confusion matrix provides a comprehensive assessment of the classifier's performance by displaying the counts of true positives, true negatives, false positives, and false negatives. It enables evaluation not only of the overall accuracy but also of the error distribution across classes.

Figure 11 presents the confusion matrices for all feature sets using the best-performing stacked model. The dimensionality-reduced fused features exhibit the highest discriminative capability, with the Nonlinear + DR configuration correctly identifying 1,493 focal and 1,488 non-focal samples, resulting in only a few misclassifications. Similarly, the Linear + DR and Hybrid + DR feature sets demonstrate strong predictive stability, underscoring the model's robustness in capturing subtle neurodynamic differences between focal and non-focal EEG patterns.

4.6 Comparison with Existing Methods

To contextualize the performance of the proposed framework, it was compared with recent studies on focal and non-focal EEG classification that used the same Bern–Barcelona EEG dataset. Since classification performance can vary depending on preprocessing methods, feature extraction techniques, and evaluation protocols, comparison across different datasets or experimental setups may not be directly comparable. Therefore, the comparison presented in Table 9 focuses on studies conducted on the same dataset to ensure a fair and consistent evaluation. As summarized in the table, the proposed stacked model demonstrates superior discriminative performance, highlighting the advantage of integrating feature fusion with dimensionality reduction for reliable EEG signal classification.

Table 9.
Comparative Analysis of the Proposed Method with Existing State-of-the-Art Approaches.

Methodology Accuracy (%)

MPLZC + Least-Squares Support Vector Machine (LS-SVM) (Borowska, Jul. 2021) 86

Multi-scale PCA + AR Burg Algorithm + Rotation Forest (Jukic et al., Sep. 2020) 98.9

Differential Operator (X–Y) + Tunable Q-Factor Wavelet Transform (TQWT) + Entropy Features + Cascade-Forward Neural Network (Sadiq et al., 2021) 97.68

Time–Frequency Hybrid Network + CNN (Sui et al., 2021) 94.3

1.GASF + Scale-Invariant Feature Transform + Random Forest

2.GASF + Oriented FAST and Rotated BRIEF + Random Forest (Krishnan et al., 2024)

9694

EMD–DWT + Log-Energy Entropy + KNN (Das & Bhuiyan, Aug. 2016) 89.4

DWT-Based Time, Frequency, and Statistical Features + Neighborhood Component Analysis + Support Vector Machine (Raghu & Sriraam, Dec. 2018) 96.1

Wavelet Decomposition + TQWT + Hjorth & Statistical Features + Human Learning Optimization –Trained Hidden Markov Model (HMM) (Chavan & Desai, May 2023) 99.15

Adaptive DCT Filter Bank + Time-Domain Features + Non-Linear Features + SVM (Mehla et al., Aug. 2023) 99.43

Band-Pass Filtering + Time & Frequency Domain Features + SVM (S et al., Oct. 2024) 99.6

Fourier–Bessel Series Expansion-Based Empirical Wavelet Transform (FBSE–EWT) + 132 Non-Linear Geometrical Features from Second-Order Difference Plot (SODP) + Kruskal–Wallis Test with VIKOR-Based Ranking (KWS–VIKOR) + LS-SVM (Anuragi et al., May 2023) 98.7

1.Ensemble Empirical Mode Decomposition (EEMD)) + Statistical and Entropy-Based Features + KNN

2.Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) + Statistical Features + KNN (Murariu et al., Jun. 2024)

93.3294.54

BW -WPD + Non-linear features fused with PCA-LDA + Stacked Classifier (This work) 99.80

Methodology	Accuracy (%)
MPLZC + Least-Squares Support Vector Machine (LS-SVM) (Borowska, Jul. 2021)	86
Multi-scale PCA + AR Burg Algorithm + Rotation Forest (Jukic et al., Sep. 2020)	98.9
Differential Operator (X–Y) + Tunable Q-Factor Wavelet Transform (TQWT) + Entropy Features + Cascade-Forward Neural Network (Sadiq et al., 2021)	97.68
Time–Frequency Hybrid Network + CNN (Sui et al., 2021)	94.3
1.GASF + Scale-Invariant Feature Transform + Random Forest 2.GASF + Oriented FAST and Rotated BRIEF + Random Forest (Krishnan et al., 2024)	9694
EMD–DWT + Log-Energy Entropy + KNN (Das & Bhuiyan, Aug. 2016)	89.4
DWT-Based Time, Frequency, and Statistical Features + Neighborhood Component Analysis + Support Vector Machine (Raghu & Sriraam, Dec. 2018)	96.1
Wavelet Decomposition + TQWT + Hjorth & Statistical Features + Human Learning Optimization –Trained Hidden Markov Model (HMM) (Chavan & Desai, May 2023)	99.15
Adaptive DCT Filter Bank + Time-Domain Features + Non-Linear Features + SVM (Mehla et al., Aug. 2023)	99.43
Band-Pass Filtering + Time & Frequency Domain Features + SVM (S et al., Oct. 2024)	99.6
Fourier–Bessel Series Expansion-Based Empirical Wavelet Transform (FBSE–EWT) + 132 Non-Linear Geometrical Features from Second-Order Difference Plot (SODP) + Kruskal–Wallis Test with VIKOR-Based Ranking (KWS–VIKOR) + LS-SVM (Anuragi et al., May 2023)	98.7
1.Ensemble Empirical Mode Decomposition (EEMD)) + Statistical and Entropy-Based Features + KNN 2.Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) + Statistical Features + KNN (Murariu et al., Jun. 2024)	93.3294.54
BW -WPD + Non-linear features fused with PCA-LDA + Stacked Classifier (This work)	99.80

Although previous studies, summarized in Table 9, report relatively high accuracy in distinguishing focal from non-focal EEG signals, they predominantly emphasize accuracy alone, which may overlook important facets of classifier performance. In this work, evaluation extends beyond accuracy to include sensitivity, specificity, F1-score, area under the ROC curve (AUC), and Brier score, with all metrics normalized to facilitate unbiased comparison. Such a comprehensive assessment is essential because high accuracy by itself does not necessarily indicate clinical reliability and can obscure the presence of false negatives, which are critical in epileptic EEG diagnosis. By integrating both performance-based and loss-based measures, the proposed framework provides a more holistic and clinically meaningful appraisal of classifier efficacy.

5. Conclusion

This work presents a systematic evaluation of EEG signal classification schemes through an integrated framework that combines linear, nonlinear, and hybrid feature extraction techniques with dimensionality reduction and data fusion. The proposed framework adopts a normalized multi-metric evaluation strategy, ensuring unbiased and comprehensive performance assessment across multiple classifiers. Unlike conventional approaches that rely solely on accuracy, this multi-metric method provides deeper insights into model behavior and generalization capability.

Among the various feature combinations analyzed, nonlinear features fused with dimensionally reduced data consistently demonstrated superior performance across several evaluation metrics, including accuracy, sensitivity, specificity, F1-score, AUC, and Brier score. The Hybrid-DR fused feature set also achieved strong results, indicating that integrating diverse descriptors can be beneficial. However, its relatively lower performance compared to the Nonlinear-DR combination suggests that certain linear features may introduce redundancy. Overall, all feature sets exhibited notable improvements when fused with dimensionality-reduced representations.

The novelty of this work lies in two major contributions. First, it introduces a multi-metric normalization and ranking framework that enables a fair and unbiased comparison across experimental configurations. Second, it demonstrates that the strategic fusion of nonlinear features with dimensionally reduced data yields a compact and reliable representation of EEG signals, thereby enhancing the classification of focal and non-focal patterns.

Experimental results confirm that the Nonlinear-DR fused feature set, when integrated with ensemble classifiers, provides a reliable and computationally efficient solution for focal and non-focal EEG classification. The proposed framework achieves a balanced trade-off among accuracy, interpretability, and computational cost, making it suitable for both research and clinical applications in automated EEG-based epilepsy detection. While earlier studies have examined individual aspects of this pipeline, the present work offers a unified methodological and analytical framework that integrates these processes cohesively.

Future research should aim to validate the framework using large-scale, clinically recorded EEG datasets to ensure robustness across diverse populations. Additionally, integrating complementary neuroimaging modalities such as Magnetoencephalography (MEG) could enable multimodal analyses, offering enhanced insight into the spatiotemporal dynamics of epileptic brain activity.

Footnotes

ORCID iD

Rajani Rai Balakrishna

Karunakara Rai Balkady

Aruvanalli Shivaraj Mamatha

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of Conflicting Interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

References

Abirami

, et al. (2024). Automated Multi-class Seizure-type Classification System using EEG Signals and Machine Learning Algorithms. IEEE Access, 12(2024): 136524–136541. https://doi.org/10.1109/ACCESS.2024.3462772

Alalayah

K. M.

Senan

E. M.

Atlam

H. F.

Ahmed

I. A.

Shatnawi

H. S. A.

(Jun. 2023). Effective Early Detection of Epileptic Seizures through EEG Signals Using Classification Algorithms Based on t-Distributed Stochastic Neighbor Embedding and K-Means. Diagnostics, 13(11). https://doi.org/10.3390/diagnostics13111957

Alazrai

Homoud

Alwanni

Daoud

M. I.

(Aug. 2018). EEG-based emotion recognition using quadratic time-frequency distribution. Sensors (Switzerland), 18(8). https://doi.org/10.3390/s18082739

Al-Hadeethi

Abdulla

Diykh

Green

J. H.

(Dec. 2021). Determinant of covariance matrix model coupled with AdaBoost classification algorithm for EEG seizure detection. Diagnostics, 12(1), 74. https://doi.org/10.3390/diagnostics12010074

Al-Nafjan

Hosny

Al-Ohali

Al-Wabil

(Dec. 01, 2017). “Review and classification of emotion recognition based on EEG brain-computer interface system research: A systematic review,” MDPI AG. https://doi.org/10.3390/app7121239

al-Qerem

Kharbat

Nashwan

Ashraf

blaou

(Mar. 2020). General model for best feature extraction of EEG using discrete wavelet transform wavelet family and differential evolution. Int. J. Distrib. Sens. Netw., 16(3). https://doi.org/10.1177/1550147720911009

Al-Salman

Wen

Miften

F. S.

Oudah

A. Y.

Al Ghayab

H. R.

(Mar. 2022). Extracting epileptic features in EEGs using a dual-tree complex wavelet transform coupled with a classification algorithm. Brain Res., 1779, https://doi.org/10.1016/j.brainres.2022.147777

Andrzejak

R. G.

Schindler

Rummel

(Oct. 2012). Nonrandomness, nonlinear dependence, and nonstationarity of electroencephalographic recordings from epilepsy patients. Phys. Rev. E Stat. Nonlin. Soft Matter Phys., 86(4). https://doi.org/10.1103/PhysRevE.86.046206

Anuragi

Singh Sisodia

Pachori

R. B.

(Jan. 2022). Epileptic-seizure classification using phase-space representation of FBSE-EWT based EEG sub-band signals and ensemble learners. Biomed. Signal Process. Control, 71, 103138. https://doi.org/10.1016/j.bspc.2021.103138

10.

Anuragi

Sisodia

D. S.

Pachori

R. B.

(May 2023). Classification of focal and non-focal EEG signals using optimal geometrical features derived from a second-order difference plot of FBSE-EWT rhythms. Artif. Intell. Med., 139, 102542. https://doi.org/10.1016/j.artmed.2023.102542

11.

Bashashati

Fatourechi

Ward

R. K.

Birch

G. E.

(Jun. 01, 2007). A survey of signal processing algorithms in brain-computer interfaces based on electrical brain signals,” https://doi.org/10.1088/1741-2560/4/2/R03

12.

Borowska

(Jul. 2021). Multiscale permutation Lempel–Ziv complexity measure for biomedical signal analysis: Interpretation and application to focal EEG signals. Entropy, 23(7). https://doi.org/10.3390/e23070832

13.

Brari

Belghith

(2021). A Novel Machine Learning Model for the Detection of Epilepsy and Epileptic Seizures Using Electroencephalographic Signals Based on Chaos and Fractal Theories. Math. Probl. Eng., 2021, https://doi.org/10.1155/2021/2107113

14.

Buchanna

Premchand

Govardhan

(Feb. 2022). Classification of epileptic and non-epileptic electroencephalogram (EEG) signals using fractal analysis and support vector regression. Emerging Science Journal, 6(1), 138–150. https://doi.org/10.28991/ESJ-2022-06-01-011

15.

Chaddad

Kateb

Bouridane

(Jul. 01, 2023). Electroencephalography Signal Processing: A Comprehensive Review and Analysis of Methods and Techniques,” Multidisciplinary Digital Publishing Institute (MDPI). https://doi.org/10.3390/s23146434

16.

Chavan

P. A.

Desai

(May 2023). Effective epileptic seizure detection by classifying focal and non-focal EEG signals using human learning optimization-based hidden markov model. Biomed. Signal Process. Control, 83, 104682. https://doi.org/10.1016/j.bspc.2023.104682

17.

Chen

Huang

Bao

Pan

(2023). An EEG-based attention recognition method: fusion of time domain, frequency domain, and non-linear dynamics features. Front. Neurosci., 17, https://doi.org/10.3389/fnins.2023.1194554

18.

Das

A. B.

Bhuiyan

M. I. H.

(Aug. 2016). Discrimination and classification of focal and non-focal EEG signals using entropy-based features in the EMD-DWT domain. Biomed. Signal Process. Control, 29, 11–21. https://doi.org/10.1016/j.bspc.2016.05.004

19.

Deepa

Anand

Pandey

B. K.

Karki

(2022). Comprehensive Performance Analysis of Classifiers in Diagnosis of Epilepsy. Math. Probl. Eng., 2022, https://doi.org/10.1155/2022/1559312

20.

Ferdous

Ali

Hamid

Molla

K. I.

(2021). Sub-band selection approach to artifact suppression from electroencephalography signal using hybrid wavelet transform. Int. J. Adv. Robot. Syst., 18(1). https://doi.org/10.1177/1729881421992269

21.

Hao

Cheng

(Dec. 2025). Motor imagery EEG classification method using 3D CNN and LSTM for rehabilitation application. Cogn. Neurodyn., 19(1), 131. https://doi.org/10.1007/s11571-025-10317-y

22.

Hassan

Hussain

S. F.

Qaisar

S. M.

(2022a). Epileptic Seizure Detection Using a Hybrid 1D CNN-Machine Learning Approach from EEG Data. J. Healthc. Eng., 2022, https://doi.org/10.1155/2022/9579422

23.

Hassan

Hussain

S. F.

Qaisar

S. M.

(2022b). Epileptic Seizure Detection Using a Hybrid 1D CNN-Machine Learning Approach from EEG Data. J. Healthc. Eng., 2022, https://doi.org/10.1155/2022/9579422

24.

Liu

Zhu

(Dec. 17, 2021). Data Augmentation for Deep Neural Networks Model in EEG Classification Task: A Review. Frontiers Media S.A. https://doi.org/10.3389/fnhum.2021.765525

25.

Hemachandira

V. S.

Viswanathan

(2022). A Framework on Performance Analysis of Mathematical Model-Based Classifiers in Detection of Epileptic Seizure from EEG Signals with Efficient Feature Selection. J. Healthc. Eng., 2022, https://doi.org/10.1155/2022/7654666

26.

Hussain

(Jun. 2018). Detecting epileptic seizure with different feature extracting strategies using robust machine learning classification techniques by applying advance parameter optimization approach. Cogn. Neurodyn., 12(3), 271–294. https://doi.org/10.1007/s11571-018-9477-1

27.

Jang

S. W.

Lee

S. H.

(Aug. 2020). Detection of epileptic seizures using wavelet transform, peak extraction and PSR from EEG signals. Symmetry (Basel)., 12(8). https://doi.org/10.3390/SYM12081239

28.

Jukic

Saracevic

Subasi

Kevric

(Sep. 2020). Comparison of ensemble machine learning methods for automated classification of focal and non-focal epileptic EEG signals. Mathematics, 8(9). https://doi.org/10.3390/math8091481

29.

Jung

T.-P.

, et al. (2000). REMOVING ELECTROENCEPHALOGRAPHIC ARTIFACTS BY BLIND SOURCE SEPARATION.

30.

Kode

Elleithy

Almazaydeh

(2024). Epileptic seizure detection in EEG signals using machine learning and deep learning techniques. IEEE Access, 12, 80657–80668. https://doi.org/10.1109/ACCESS.2024.3409581

31.

Krishnan

P. T.

Erramchetty

S. K.

Balusa

B. C.

(2024). Advanced framework for epilepsy detection through image-based EEG signal analysis. Front. Hum. Neurosci., 18, https://doi.org/10.3389/fnhum.2024.1336157

32.

Liu

C. C.

Hajra

S. G.

Fickling

S. D.

Pawlowski

Song

D’Arcy

R. C. N.

(Feb. 2020). Novel signal processing technique for capture and isolation of blink-related oscillations using a low-density electrode array for bedside evaluation of consciousness. IEEE Trans. Biomed. Eng., 67(2), 453–463. https://doi.org/10.1109/TBME.2019.2915185

33.

Luján

MÁ

Jimeno

M. V.

Sotos

J. M.

Ricarte

J. J.

Borja

A. L.

(Dec. 01, 2021). “A survey on eeg signal processing techniques and machine learning: Applications to the neurofeedback of autobiographical memory deficits in schizophrenia,” MDPI. https://doi.org/10.3390/electronics10233037

34.

Mardini

Bani Yassein

M. M.

Al-Rawashdeh

Aljawarneh

Khamayseh

Meqdadi

(2020). Enhanced detection of epileptic seizure using EEG signals in combination with machine learning classifiers. IEEE Access, 8, 24046–24055. https://doi.org/10.1109/ACCESS.2020.2970012

35.

Mehla

V. K.

Singhal

Singh

(Aug. 2023). An efficient classification of focal and non-focal EEG signals using adaptive DCT filter bank. Circuits Syst. Signal Process., 42(8), 4691–4712. https://doi.org/10.1007/s00034-023-02328-z

36.

Murariu

M.-G.

Dorobanțu

F.-R.

Tărniceriu

(Jun. 2024). Enhanced classification of focal and generalized epilepsy using EEMD and CEEMDAN methods. Traitement du Signal, 41(3), 1315–1322. https://doi.org/10.18280/ts.410320

37.

Panigrahi

Behera

D. K.

Patra

K. C.

(Feb. 2022). Epileptic seizure classification of electroencephalogram signals using extreme gradient boosting classifier. Indonesian Journal of Electrical Engineering and Computer Science, 25(2), 884. https://doi.org/10.11591/ijeecs.v25.i2.pp884-891

38.

Polat

Nour

(Dec. 2020). Epileptic seizure detection based on new hybrid models with electroencephalogram signals. IRBM, 41(6), 331–353. https://doi.org/10.1016/j.irbm.2020.06.008

39.

Prabhakar

S. K.

Won

D. O.

(2023). Performance comparison of bio-inspired and learning-based clustering analysis with machine learning techniques for classification of EEG signals. Front. Artif. Intell., 6, https://doi.org/10.3389/frai.2023.1156269

40.

Rafiuddin

Khan

Y. U.

Farooq

(2022). A novel wavelet approach for multiclass iEEG signal classification in automated diagnosis of epilepsy. IEEE Trans. Instrum. Meas., 71, 1–10. https://doi.org/10.1109/TIM.2022.3207799

41.

Raghu

Sriraam

(Dec. 2018). Classification of focal and non-focal EEG signals using neighborhood component analysis and machine learning algorithms. Expert Syst. Appl., 113, 18–32. https://doi.org/10.1016/j.eswa.2018.06.031

42.

Rai B

A S

Nikshitha , (Dec. 2025b). An analytical framework for enhancing brain signal classification through hybrid filtering and dimensionality reduction. Healthcare Analytics, 8, https://doi.org/10.1016/j.health.2025.100435

43.

Rai B

A S

Sooda

(Jun. 01, 2025c). Advancements in epilepsy classification: Current trends and future directions. Elsevier B.V. https://doi.org/10.1016/j.mex.2025.103257

44.

Rai B

Mamatha

Kumar N

(Nov. 2025a). A VIKOR-Driven Deep Learning Framework for Classification of Focal and Non-Focal EEG Signals. Journal of Intelligent & Fuzzy Systems: Applications in Engineering and Technology. https://doi.org/10.1177/18758967251395466

45.

Rajani Rai

Martis

R. J.

(2021). Automated Decision Support System for Focal Epilepsy Detection using Electroencephalogram. 2021 IEEE Bombay Section Signature Conference, IBSSC 2021, Institute of Electrical and Electronics Engineers Inc., https://doi.org/10.1109/IBSSC53889.2021.9673180

46.

Ranjan

Sahana

B. C.

(Oct. 2024). Multiresolution feature fusion for smart diagnosis of schizophrenia in adolescents using EEG signals. Cogn. Neurodyn., 18(5), 2779–2807. https://doi.org/10.1007/s11571-024-10120-1

47.

K. S.

Mohan

K.P.

(Oct. 2024). EEG Based automated detection of seizure using machine learning approach and traditional features. Expert Syst. Appl., 251, 123991. https://doi.org/10.1016/j.eswa.2024.123991

48.

Sadiq

M. T.

, et al. (2021). Exploiting Feature Selection and Neural Network Techniques for Identification of Focal and Nonfocal EEG Signals in TQWT Domain. J. Healthc. Eng., 2021, https://doi.org/10.1155/2021/6283900

49.

Saemaldahr

Ilyas

(Jul. 2023). Patient-Specific Preictal Pattern-Aware Epileptic Seizure Prediction with Federated Learning. Sensors, 23(14). https://doi.org/10.3390/s23146578

50.

Sahoo

S. K.

Mohapatra

S. K.

(2022). Recognition of Ocular Artifacts in EEG Signal through a Hybrid Optimized Scheme. Biomed Res. Int., 2022, https://doi.org/10.1155/2022/4875399

51.

Sairamya

N. J.

Subathra

M. S. P.

Suviseshamuthu

E. S.

Thomas George

(Jan. 2021). A new approach for automatic detection of focal EEG signals using wavelet packet decomposition and quad binary pattern method. Biomed. Signal Process. Control, 63, https://doi.org/10.1016/j.bspc.2020.102096

52.

Savadkoohi

Oladunni

Thompson

(Jul. 2020). A machine learning approach to epileptic seizure prediction using electroencephalogram (EEG) signal. Biocybern. Biomed. Eng., 40(3), 1328–1341. https://doi.org/10.1016/j.bbe.2020.07.004

53.

Sharmila

Geethanjali

(May 2020). Evaluation of time domain features on detection of epileptic seizure from EEG signals. Health Technol. (Berl), 10(3), 711–722. https://doi.org/10.1007/s12553-019-00363-y

54.

Sohaib

Ghaffar

Shin

Hasan

M. J.

Suleman

M. T.

(Oct. 2022). Automated Analysis of Sleep Study Parameters Using Signal Processing and Artificial Intelligence. Int. J. Environ. Res. Public Health, 19(20). https://doi.org/10.3390/ijerph192013256

55.

Suárez

Torres

A. M.

Blasco-Segura

Mateo

(2025). Application of the random forest algorithm for accurate bipolar disorder classification. Life, 15(3), 394. https://doi.org/10.3390/life15030394

56.

Sui

Zhao

Tanaka

Cao

(2021). Hybrid Convolutional Neural Network for Localization of Epileptic Focus Based on iEEG. Neural Plast., 2021, https://doi.org/10.1155/2021/6644365

57.

Sun

Mou

(2023). “Survey on the research direction of EEG-based signal processing,” Frontiers Media SA. https://doi.org/10.3389/fnins.2023.1203059

58.

Tahernezhad-Javazm

Azimirad

Shoaran

(Feb. 19, 2018). “A review and experimental study on the application of classifiers and evolutionary algorithms in EEG-based brain-machine interface systems,” Institute of Physics Publishing. https://doi.org/10.1088/1741-2552/aa8063

59.

Tran

L. V.

Tran

H. M.

T. M.

Huynh

T. T. M.

Tran

H. T.

Dao

S. V. T.

(Nov. 2022). Application of Machine Learning in Epileptic Seizure Detection. Diagnostics, 12(11). https://doi.org/10.3390/diagnostics12112879

60.

Lin

(Sep. 2022). Epilepsy Seizures Prediction Based on Nonlinear Features of EEG Signal and Gradient Boosting Decision Tree. Int. J. Environ. Res. Public Health, 19(18). https://doi.org/10.3390/ijerph191811326

61.

Yan

(Nov. 2022). A time-frequency denoising method for single-channel event-related EEG. Front. Neurosci., 16, https://doi.org/10.3389/fnins.2022.991136

62.

Yuan

, et al. (Aug. 2017). Epileptic seizure detection based on imbalanced classification and wavelet packet transform. Seizure, 50, 99–108. https://doi.org/10.1016/j.seizure.2017.05.018

63.

Zhou

(Jun. 2020). Epilepsy EEG Signal Classification Algorithm Based on Improved RBF. Front. Neurosci., 14, https://doi.org/10.3389/fnins.2020.00606