Effective Denoising of Epileptic High-Frequency Oscillations (HFOs) in Scalp EEG Using a Temporal Generative Adversarial Network (TimeGAN)

Abstract

Over the past two decades, High-Frequency Oscillations (HFOs) have been widely recognized as promising biomarkers for delineating the specific regions of the brain responsible for seizure onset, and for monitoring the spatiotemporal dynamics of epileptogenic activity. While deep learning (DL) techniques have shown promise in processing HFOs in intracranial EEG (iEEG) data, they often struggle to perform effectively on scalp electroencephalography (EEG). This limitation is particularly owing to the inherently lower signal-to-noise ratio of scalp HFOs, their smaller amplitude, and their increased susceptibility to artifact contamination, which together obscure subtle high-frequency phenomena. Despite these restrictions or barriers, clinicians predominantly rely on scalp EEG recordings due to their noninvasive nature, low cost, and recognized safety. To extend the benefits or applicability of DL-based HFO detection to a broader patient population, automated scalp EEG systems incorporating HFO denoising algorithms are highly warranted and necessary. Hence, the present study aimed at proposing, developing, and comparing the effectiveness of five optimized DL-based paradigms for denoising unwanted artifacts from realistic scalp EEG. The target data was created by combining intracranial HFO samples with injected realistic contaminants such as electrooculogram (EOG) and electromyography (EMG), accurately replicating scalp EEG profiles. In this regard, we considered evaluating and comparing the following time-series DL architectures, referred to as: one-dimensional Convolutional Neural Networks (CNN-1D), Long Short-Term Memory (LSTM) networks, Stacked Autoencoders (SAE), one-dimensional Transformer (Transformer-1D), and Generative Adversarial Networks (GAN). This analysis aims to elucidate each model’s denoising performance strengths and weaknesses. The reached quantitative and qualitative experimental results revealed that the proposed current GAN model significantly achieved superior performance compared to other DL approaches, suggesting its effectiveness as a robust solution for denoising scalp HFOs, thereby enhancing noninvasive HFOs detection.

Received: 7 April 2025 | Revised: 15 August 2025 | Accepted: 30 September 2025

Conflicts of Interest

The authors declare that they have no conflicts of interest to this work.

Data Availability Statement

Data were available from the corresponding author upon reasonable request.

Author Contribution Statement

Sahbi Chaibi: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Resources, Data curation, Writing – original draft, Writing – review & editing, Visualization, Project administration. Wadhah Ayadi: Conceptualization, Methodology, Validation. Abdennaceur Kachouri: Validation, Supervision.