Ilmu Komputer & AI editorial
Open AccessOA2026
From Neurons to Networks: A Holistic Review of Electroencephalography (EEG) from Neurophysiological Foundations to AI Techniques
A comprehensive synthesis of EEG analysis, from neurophysiological principles to foundation models and neuro-symbolic AI
Christos Kalogeropoulos; Konstantinos A. Theofilatos; Seferina Mavroudi· Signals· 2026· DOI 10.3390/signals7010017
The core problem
Electroencephalography (EEG) has undergone a profound transformation over the past century, evolving from a subjective observational tool into a data-intensive analytical discipline that leverages sophisticated algorithms and mathematical models. This review provides a holistic foundation by first detailing the neurophysiological basis of EEG, the recording techniques employed, and the broad spectrum of applications. It then offers a rigorous examination of both traditional and modern analytical pillars, including statistical and time-series analysis, spectral and time-frequency analysis, spatial analysis and source modelling, connectivity and network analysis, and nonlinear and chaotic analysis. While acknowledging the historical role of machine learning (ML) and deep learning (DL) architectures such as support vector machines (SVMs) and convolutional neural networks (CNNs), the review shifts its primary focus toward current state-of-the-art artificial intelligence (AI) trends. Emphasis is placed on the emergence of foundation models, including large language models (LLMs) and large vision models (LVMs), adapted for high-dimensional neural sequences. Finally, the integration of ge
Innovation
The review synthesises findings across multiple domains. Traditional analytical methods continue to provide robust, interpretable features: for instance, spectral analysis reveals power in frequency bands (delta: 0.5–4 Hz, theta: 4–8 Hz, alpha: 8–13 Hz, beta: 13–30 Hz, gamma: >30 Hz), while connectivity analysis uncovers network topologies underlying cognitive processes. Machine learning models, particularly SVMs and CNNs, have achieved high accuracy in tasks such as seizure detection and sleep staging. Deep learning architectures, including CNNs and RNNs, have automated feature extraction, but often lack interpretability. Foundation models, such as LLMs and LVMs, have demonstrated remarkable versatility in handling high-dimensional EEG sequences, enabling zero-shot and few-shot learning. Generative AI, including variational autoencoders (VAEs) and generative adversarial networks (GANs), has been used for data augmentation, addressing the scarcity of labelled EEG data. XAI frameworks, such as SHAP and LIME, have been applied to decode black-box models, providing insights into feature importance and model decisions. The review highlights that neuro-symbolic architectures, which inte
Electroencephalography (EEG) has undergone a profound transformation over the past century, evolving from a subjective observational tool into a data-intensive analytical discipline that leverages sophisticated algorithms and mathematical models. This review provides a holistic foundation by first detailing the neurophysiological basis of EEG, the recording techniques employed, and the broad spectrum of applications. It then offers a rigorous examination of both traditional and modern analytical pillars, including statistical and time-series analysis, spectral and time-frequency analysis, spatial analysis and source modelling, connectivity and network analysis, and nonlinear and chaotic analysis. While acknowledging the historical role of machine learning (ML) and deep learning (DL) architectures such as support vector machines (SVMs) and convolutional neural networks (CNNs), the review shifts its primary focus toward current state-of-the-art artificial intelligence (AI) trends. Emphasis is placed on the emergence of foundation models, including large language models (LLMs) and large vision models (LVMs), adapted for high-dimensional neural sequences. Finally, the integration of generative AI for data augmentation and explainable AI (XAI) frameworks designed to bridge the gap between “black-box” decoding and clinical interpretability is explored. The review concludes that the next generation of EEG analysis will likely converge into neuro-symbolic architectures, synergising the massive generative power of foundation models with the rigorous, rule-based interpretability of classical signal theory.
The review adopts a holistic approach, systematically covering the neurophysiological foundations of EEG, recording techniques, and applications before delving into analytical methodologies. The analytical pillars are categorised into five main areas: (1) Statistical and Time-Series Analysis, which includes measures such as mean, variance, autocorrelation, and autoregressive models; (2) Spectral and Time-Frequency Analysis, encompassing Fourier transform, wavelet transform, and spectrograms; (3) Spatial Analysis and Source Modelling, involving electrode placement, spatial filtering, and inverse solutions; (4) Connectivity and Network Analysis, utilising coherence, phase synchronisation, and graph-theoretical metrics; and (5) Nonlinear and Chaotic Analysis, employing measures like Lyapunov exponents, entropy, and fractal dimensions. The review then examines machine learning and deep learning architectures, including SVMs, CNNs, recurrent neural networks (RNNs), and transformers, before focusing on foundation models (LLMs, LVMs) adapted for EEG. Generative AI techniques for data augmentation and XAI frameworks are also reviewed. The methodology emphasises a transition from traditional signal processing to modern AI-driven approaches, culminating in the proposal of neuro-symbolic architectures that combine generative power with interpretability.
Why it matters
The discussion underscores the evolution of EEG analysis from subjective observation to data-intensive computation. Traditional signal processing techniques remain valuable for their interpretability and theoretical grounding, but they often require manual feature engineering and may not capture complex nonlinear dynamics. Machine learning and deep learning have revolutionised EEG analysis by automating feature extraction and achieving state-of-the-art performance, yet their “black-box” nature poses challenges for clinical adoption. Foundation models represent a paradigm shift, leveraging large-scale pre-training on diverse EEG datasets to enable transfer learning and few-shot adaptation. However, their computational demands and potential biases necessitate careful validation. Generative AI addresses data scarcity but raises ethical concerns regarding synthetic data. XAI bridges the interpretability gap, but current methods are often post-hoc and may not fully capture causal relationships. The review argues that neuro-symbolic AI, which combines the generative power of foundation models with the rigorous, rule-based interpretability of classical signal theory, is the most promising direction. Such architectures could enable both high accuracy and transparent decision-making, facilitating clinical translation. Future work should focus on developing standardised benchmarks, addressing ethical and privacy concerns, and fostering interdisciplinary collaboration between neuroscientists, clinicians, and AI researchers.
Who should read this
CS practitioners and researchers
Opening member content…