EEG-Based Disease Classification

• EEG-based disease classification is a field that uses brain electrical recordings and machine learning to detect and differentiate neurological and psychiatric disorders.
• It integrates rigorous data preprocessing, artifact removal, and feature extraction methods such as spectral, nonlinear, and connectivity analyses.
• Spatial mapping and deep learning models are applied to improve diagnostic accuracy and interpretability for conditions like epilepsy, Alzheimer’s, and Parkinson’s.

Electroencephalography-based disease classification leverages multivariate brain electrical recordings to detect, discriminate, and subtype neurological and psychiatric conditions. Modern EEG-based disease classifiers utilize diverse feature construction, denoising, and machine learning strategies, ranging from signal decomposition and network analytics to deep learning on spatially organized representations. Classification algorithms are validated on a wide spectrum of clinical tasks, including the detection of epilepsy, Alzheimer's disease, Parkinson's disease, schizophrenia, and multi-disease screening. Below, key principles, methodologies, and comparative results are systematically delineated.

1. Data Acquisition, Denoising, and Preprocessing

Technical rigor in data acquisition and preprocessing is foundational. Routine disease-classification pipelines acquire EEG using standard montages (e.g., 10–20, 10–10, or high-density 64/128-channel arrays) at sampling rates between 250–1000 Hz for resting-state, task-based, or TMS-evoked paradigms. Preprocessing strategies typically include:

• Referencing and Artifact Removal: Use of bipolar, common-average, or REST referencing. Ocular, muscular, and cardiac artifacts are removed using Independent Component Analysis (FastICA) and visual/automatic IC rejection (Redwan et al., 2022, Sahota et al., 2023).
• Channel and Epoch Selection: Bad channels are interpolated or excluded, and only artifact-free epochs are retained, as in large-scale TUAB or multi-center clinical datasets (Sahota et al., 1 May 2025, Samanta et al., 27 Dec 2025).
• Filtering: The chains employ high-pass (0.5–1 Hz), low-pass (30–70 Hz), and notch filters to minimize physiological and instrumentation noise.
• Denoising: Wavelet-based approaches such as Tunable Q-factor Wavelet Transform (TQWT, with $Q=6$, $r=5$, $J=10$ levels) are applied to preserve oscillatory content while reducing noise prior to feature derivation (SV, 2023). Empirical Wavelet Transform (EWT), Variational Mode Decomposition (VMD), and band-specific windowing are used for adaptive spectral isolation (Akbari et al., 2019, Ullal et al., 2020).

2. Feature Extraction: Spectral, Temporal, Nonlinear, and Network Dimensions

2.1 Spectral Features

Broadly adopted pipelines compute per-channel power spectral density (PSD), usually via Welch or multitaper methods, integrating absolute or relative power in canonical frequency bands ($\delta$, $\theta$, $\alpha$, $\sigma$, $\beta$, $\gamma$) (Redwan et al., 2022, Sahota et al., 1 May 2025, Samanta et al., 27 Dec 2025). Peak frequency, median frequency, spectral entropy, and spectral edge frequency (SEF95) are included for finer resolution.

2.2 Nonlinear and Complexity Features

Nonlinear dynamics are increasingly used to capture the long-range dependence and chaoticity of EEG:

• Maximal Lyapunov Exponent (MLE): Quantifies phase-space divergence using Rosenstein’s algorithm, revealing dynamical instability alterations (especially relevant for schizophrenia classification) (SV, 2023).
• Hurst Exponent (HE): Characterizes long-term autocorrelation structure, relevant for disorders altering temporal dependencies (e.g., schizophrenia, dementia) (SV, 2023).
• Entropy Measures: Shannon, permutation, and sample entropy are employed to quantify signal complexity and differentiate disease-relevant states (Sahota et al., 1 May 2025, Samanta et al., 27 Dec 2025).
• 95% Confidence Ellipse Area: Derived from phase space reconstruction (PSR) of decomposed modes/rhythms, this geometric metric robustly separates ictal from non-ictal states in epilepsy (Akbari et al., 2019, Ullal et al., 2020).

2.3 Time-Domain and Statistical Features

Moments (mean, variance, skewness, kurtosis), line length, RMS energy, Hjorth parameters (activity, mobility, complexity), and autocorrelation coefficients are systematically computed for all channels/waves (Tautan et al., 2022, Samanta et al., 27 Dec 2025). TMS-evoked features (peak amplitudes/latencies, root-mean-square) and event-specific measures are applied in paradigms with evoked/induced responses (Tautan et al., 2022, Tautan et al., 2022).

2.4 Network and Connectivity Features

Disease-induced reorganization of functional connectivity is probed using:

• Coherence, Phase-Locking Value (PLV), Phase-Lag Index (PLI): Channel-wise connectivity metrics are calculated per band and input to classifiers, especially for disorders with altered synchrony (Alzheimer's, epilepsy) (Sahota et al., 1 May 2025, 2404.1329, Detti et al., 2018).
• Graph Construction: Vertices represent electrodes; edges/weights represent similarity (Fréchet, correlation) or connectivity (coherence, PLV). Maximum-weight clique selection (mwcEEGs) is used to select homogeneous, artifact-free trial subsets (Dai et al., 2018).
• Quaternion PCA: Captures high-dimensional inter-channel relationships for connectivity analysis in small-channel paradigms (Hung et al., 21 May 2025).

3. Spatial Representation and the Importance of Electrode Configuration

Encoding spatial information is shown to substantially improve classifier performance:

• Topographic Mapping: Channels are arranged on regular grids reflective of physical scalp topology, enabling the construction of 2D (or interpolated) images processed by convolutional networks (SV, 2023, Wang et al., 2021, Mokatren et al., 2019).
• Spatial Interpolation: Band energies/power are interpolated across grids (e.g., via azimuthal equidistant projection, inverse-distance weighting) to generate dense maps for spatial feature extraction (Wang et al., 2021).
• Spatial vs. Temporal vs. Spectral Resolution: Empirical studies demonstrate spatial information is at least as informative as spectral, and more valuable than temporal resolution for classification of Alzheimer's and similar disorders, when encoded using graph signal processing or spatial pooling (Goerttler et al., 2024).
• Deep Models with Sensor Configuration: Factoring in electrode layout for image-like inputs yields 5–8% absolute accuracy improvement over feature-concatenation methods when tested with CNNs, SVMs, or kNNs (Mokatren et al., 2019).

4. Machine Learning and Deep Neural Architectures

4.1 Classical ML Approaches

Random Forests, SVMs (typically with RBF kernels), AdaBoost, and Gaussian Process Classifiers are widely implemented. For example, GPCs achieved 95.5% accuracy, 95.8% specificity, and 95.3% sensitivity for first-episode psychosis vs. controls using bandpower vectors (Redwan et al., 2022). Random forests remain a preferred solution for explainability, as seen in Knowledge-driven Fusion Forests (Sahota et al., 1 May 2025) and TMS-EEG feature pipelines (Tautan et al., 2022, Tautan et al., 2022).

4.2 Deep Learning and End-to-End Methods

• 1D CNNs: Raw time-domain vectors as input to deep residual 1D CNNs have yielded as high as 99.9% specificity and 99.5% sensitivity in seizure-vs-healthy binary tasks, and 81% sensitivity in five-way epilepsy (Gupta et al., 2021).
• 2D CNNs with Spatial Encodings: MobileNetV2 applied to nonlinear heatmaps, EfficientNet backbones following 1D-to-pseudoimage transformation (VIPEEGNet), and classical CNN stacks on interpolated bandpower images have demonstrated high efficacy (AUROC up to 0.99) (SV, 2023, Sun et al., 10 Jul 2025, Mokatren et al., 2019).
• Minimalist CNNs/Interpretable Models: Single-layer architectures (LightCNN) with direct multi-channel convolutions surpass deeper ensembles in Parkinson’s disease classification, achieving 98.9% accuracy and 100% precision (Anjum, 2024).
• Capsule Networks, Spatiotemporal Graph Neural Networks: Pixelwise spatial mapping and capsule routing for band energy distributions achieves 89.3% (±4.1%) accuracy for PD (Wang et al., 2021). Personalized graph neural networks and Gated GCNs using node (feature) and edge (connectivity) encodings deliver interpretable state-of-the-art AUC and accuracy (AUC = 0.991, accuracy = 0.984) for moderate-to-severe Alzheimer’s (Wang et al., 2 Apr 2025).

4.3 Explainable Approaches

Feature importance analyses elucidate physiological drivers (e.g., increased θ-coherence, variance, entropy reduction) underlying classifier output (Sahota et al., 1 May 2025, Samanta et al., 27 Dec 2025). Explainable GNNs and Fusion Forests report clinician-readable feature importances and offer visualizations of discriminative neural networks (e.g., frontal-parietal synchrony in AD) (Wang et al., 2 Apr 2025).

5. Performance Metrics and Validation Protocols

Standard accuracy, sensitivity, specificity, precision, F1-score, and AUROC are reported across studies, with leave-one-subject-out CV, repeated K-fold, and stratified splits being the norm. In multi-disorder clinical screening, disorder-specific threshold calibration is explicitly used to prioritize sensitivity (recall) in low-prevalence classes, yielding gains of 15–50% in recall compared to default thresholds (Samanta et al., 27 Dec 2025).

Classifier/Task	Accuracy (%)	Sensitivity (%)	Specificity (%)	F1 / AUROC	Notable Design
RandomForest (TMS-EEG, AD)	92.7	96.6	88.8	92.7	High-density montage, 14 time-domain features (Tautan et al., 2022)
GPC (PSD, Psychosis)	95.5	95.3	95.8	—	Resting PSD, 240 features (Redwan et al., 2022)
1D-ResNet (EEG, Epilepsy)	99.5 (bin.)/81.0 (5-way)	99.5/81.0	99.9/81.4	—	Raw waveform input (Gupta et al., 2021)
LightCNN (PD, EEG)	98.9	97.7	100.0	0.998 (AUC); F1=0.99	Single convolutional layer, raw multi-chan. (Anjum, 2024)
KnowEEG Fusion Forest	80.2	78.0	82.0	87.7 (AUROC)	Per-electrode + connectivity features, explainable (Sahota et al., 1 May 2025)
VIPEEGNet (Multiclass)	—	36.8–88.2	55.6–80.4 (prec.)	0.93–0.97 (AUROC)	Vision-inspired transfer learning (Sun et al., 10 Jul 2025)
GGCN (β-PLV, AD)	98.4	97.2	100.0	0.991 (AUC)	Graph-based, explainable, β-band PLV (Wang et al., 2 Apr 2025)

6. Task-Specific Pipelines and Multi-Disorder Benchmarking

• Alzheimer’s Disease: Multichannel graph representations (spectral clustering, eigenvector pooling), PLV/PLI adjacency, and network decomposition (QPCA) reveal alpha-band connectivity drop and frontal-parietal disruption as robust AD markers (Wang et al., 2 Apr 2025, Goerttler et al., 2024, Hung et al., 21 May 2025).
• Epilepsy/Seizure: Phase synchrony indices (PLV, PLI, WPLI), graph-theoretic strength, and hybrid trend features (MAACD) allow patient-specific preictal warning (TH classifier, 50 s warning, zero FPs) (Detti et al., 2018). 1D CNNs and VMD-based MLPs achieve >98% accuracy for binary/ternary detection (Ullal et al., 2020, Gupta et al., 2021).
• Parkinson’s Disease: Multi-modal approaches (15-variate bandpower/peak-freq representations with AdaBoost, CapsNet on spatial-spectral images, 1D/2D CNNs) harness NREM sleep, gamma-band topography, and delta-beta rhythms as sensitive features, reaching 85–98.9% accuracy (Sahota et al., 2023, Wang et al., 2021, Anjum, 2024).
• Schizophrenia/Psychosis: Nonlinear feature mapping (MLE/HE) organized into spatial heatmaps, processed by CNNs, achieves up to 75% accuracy for schizophrenia (SV, 2023), and ∼95% accuracy for first-episode psychosis via PSD-GPC (Redwan et al., 2022).
• Multi-Disorder Screening: Clinical EEG pipelines with disorder-calibrated thresholds and multi-domain features achieve sensitivity-oriented recall ≥80% across 11 neurological disorders, demonstrating plausible feature importances for each category (Samanta et al., 27 Dec 2025).

7. Limitations, Challenges, and Future Directions

Current pipelines face constraints from small/biased datasets, lack of rigorous cross-validation, limited generalizability across acquisition hardware, and reliance on shallow feature engineering in many clinical environments. Critical directions include:

• Dataset Diversification: Expansion to larger, multi-institution, multi-modal databases to improve generalizability and robustness (SV, 2023, Anjum, 2024).
• Integration of Advanced Features: Incorporation of fractal dimension, dynamic network topologies, cross-frequency coupling, and novel statistical features is recommended for multidimensional heatmaps and explainable frameworks (SV, 2023, Sahota et al., 1 May 2025).
• Hyperparameter and Architecture Optimization: Systematic tuning (learning rates, pooling ratios, number of GNN layers) via multi-objective Bayesian methods (MOTPE) is demonstrated to yield Pareto-optimal trade-offs (Wang et al., 2 Apr 2025).
• Explainability and Clinical Integration: Increasing emphasis on feature/pathway interpretability, saliency mapping, and convergence with clinician-explainable output is evident in the design of Fusion Forests, GGN architectures, and visually driven topographic analyses (Sahota et al., 1 May 2025, Wang et al., 2 Apr 2025).
• Temporal and Event-Related Features: While resting-state remains the norm, event-related paradigms require enhanced temporal resolution and may benefit from time–frequency balancing frameworks (Goerttler et al., 2024).

EEG-based disease classification has thus evolved into a confluence of advanced feature engineering, spatial-spectral analytics, and interpretable deep learning, establishing robust baselines and innovative pathways for non-invasive neurological diagnostics across the lifespan.