Robust Deep Learning for EEG Decoding: Scaling from Seizure Detection to Multi-Task Cross-Species Foundation Models

Mots clé :

• Deep Learning
• EEG
• Brain Activity Decoding

Decoding brain activity from electroencephalography (EEG) signals offers significant potential for clinical diagnosis, pharmacological research, and brain–computer interfaces. By providing a real-time measure of neural dynamics, EEG enables the study of physiological and pathological brain states and supports predictive analysis of the effects of pharmacological agents on neural oscillations (pharmaco-EEG). However, EEG analysis remains challenging due to its low signal-to-noise ratio, non-stationarity, inter- and intra-subject variability, and limited availability of annotated data. These constraints hinder the development of reliable and generalizable automated systems, particularly in real-world settings.

This thesis aims to address these challenges through a progressive framework, moving from task-specific applications toward general EEG representation learning. First, we investigate automated epileptic seizure detection and highlight a key gap between classification-based benchmarks and real-world detection scenarios. While many approaches rely on pre-segmented data, clinical practice requires continuous monitoring and precise temporal localization of seizure events. To bridge this gap, we propose a new pipeline for continuous EEG analysis combining a deep learning model with post-processing to reconstruct overlapping seizure events in long recordings. Our approach demonstrates strong generalization, including cross-species transfer from murine to human EEG.

Second, we explore the prediction of anti-seizure medication efficacy by analyzing seizure-free (interictal) EEG activity. By focusing on interictal signals, we reduce reliance on seizure occurrence, which is often sparse and can be affected by environmental factors. This approach helps mitigate biases in preclinical evaluation and opens new perspectives for faster and more reliable assessments of drug efficacy.

To improve robustness in EEG decoding, we introduce CoSupFormer, a deep learning architecture designed to handle noisy multi-channel signals. It combines multi-scale convolutional encoding with a gated global attention mechanism to model both intra- and inter-channel interactions while reducing the impact of corrupted channels. A hybrid loss function integrating cross-entropy and supervised contrastive learning improves class separability, particularly for pharmacological effect prediction. We also explore conditional EEG generation using diffusion models to address class imbalance and enhance data representativeness.

Finally, we propose SpecMoE, a cross-species EEG foundation model based on spectral-aware self-supervised learning. Existing masking strategies often introduce artificial high-frequency discontinuities and allow low-frequency information leakage, leading to suboptimal representation learning. To overcome this limitation, we introduce a Gaussian-smoothed masking strategy applied to time-frequency representations, encouraging the model to learn meaningful neural dynamics across scales instead of relying on reconstruction shortcuts. To effectively recover signals under this challenging masking regime, we design SpecHi-Net, a hierarchical encoder–decoder architecture for robust reconstruction, and integrate a spectral-guided mixture-of-experts framework that dynamically adapts to signal characteristics. This approach improves generalization across tasks, datasets, and species.

Overall, this work contributes to the development of more reliable, robust, and scalable EEG decoding methods, bridging the gap between controlled experimental settings and real-world clinical and preclinical applications, and paving the way toward more general and reusable neural representation learning frameworks.

Président·e:

Sophie Achard (CNRS, Université Grenoble Alpes)

Directeurs:

• Sergei Grudinin (CNRS, Université Grenoble Alpes )
• Julien Volle (SynapCell - SAS )

Rapporteur·e·s:

• Sylvain Chevallier (INRIA, Université Paris-Saclay )
• Yaël Frégier (Université d’Artois )

Examinateur·trice·s:

• Giulia Lioi (CNRS, IMT Atlantique )
• Vincent Magloire (INSERM, CRNL )