Rationale:
Robust seizure detection from scalp EEG remains a major challenge in clinical neurology due to the complex spatiotemporal nature of the signals and significant inter-subject variability. Traditional deep learning models often fail to generalize across patients, limiting their clinical applicability. To address these limitations, we developed a foundation model for EEG that learns patient-invariant and anatomically informed representations, with the explicit goal of enabling accurate and robust seizure detection across diverse patient populations.
Methods:
The foundation model is designed with three key components: (1) inverse gradient training with patient identity labels to achieve domain-invariant representation, (2) coordinate-based positional encoding to reflect anatomical structure in the attention mechanism, and (3) joint reconstruction of time- and frequency-domain EEG signals to preserve electrophysiological relevance. The model was pretrained on scalp EEG recordings from 24 participants provided by the publicly available PhysioNet database. For downstream seizure classification, we used a separate neonatal EEG dataset comprising 78 subjects, with expert annotations identifying seizure and non-seizure epochs. The encoder output was passed to an attention-based classifier composed of residual multi-head attention layers and learnable channel-wise attention pooling, followed by a multi-layer perceptron for binary classification. We compared our model against a randomly initialized transformer model trained from scratch and three recent EEG foundation models (BFM, CBraMod, NeuroGPT).
Results:
The proposed model outperformed all baselines in cross-subject seizure detection. The baseline transformer trained from scratch achieved an area under the curve (AUC) of 0.69, accuracy of 68.5%, sensitivity of 66.1%, and specificity of 70.8%, highlighting the difficulty of training robust models without inductive biases. Among the three recent EEG foundation models, BFM achieved an AUC of 0.75; CBraMod achieved an AUC of 0.79; and NeuroGPT achieved an AUC of 0.81. These models demonstrated improved performance over the scratch model. In contrast, our proposed model achieved the highest overall performance, with an AUC of 0.83.
Conclusions:
This study presents a transformer-based EEG foundation model that integrates anatomical priors, domain-invariant learning, and frequency-aware reconstruction to enable robust and generalizable seizure detection. By pretraining on publicly available EEG data and fine-tuning on a neonatal seizure dataset, the model achieved superior performance compared to existing models. These findings highlight the potential of spatially structured and patient-invariant EEG representation learning for advancing clinical seizure detection in heterogeneous populations.
Funding: This research was supported by the Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education (2025R1), South Korea.