Graph Adapter of EEG Foundation Models for Parameter Efficient Fine Tuning

TL;DR

This work addresses the high cost of fine-tuning large EEG foundation models while leveraging spatial sensor information. It introduces EEG-GraphAdapter (EGA), a parameter-efficient GNN-based adapter placed before a frozen temporal backbone (BENDR) to inject spatial context via a geodesic-distance EEG graph. Across MDD and TUAB tasks, EGA achieves notable gains in F1-score and AUROC with substantially reduced trainable parameters, demonstrating data-efficient, scalable EEG predictions. The approach highlights the potential of combining PEFT with graph-based spatial modeling to extend EEG foundation models to clinically relevant tasks with limited labeled data. This suggests practical impact for resource-limited clinical EEG applications and sets the stage for broader applicability to other pre-trained EEG backbones and disorders.

Abstract

In diagnosing neurological disorders from electroencephalography (EEG) data, foundation models such as Transformers have been employed to capture temporal dynamics. Additionally, Graph Neural Networks (GNNs) are critical for representing the spatial relationships among EEG sensors. However, fine-tuning these large-scale models for both temporal and spatial features can be prohibitively large in computational cost, especially under the limited availability of labeled EEG datasets. We propose EEG-GraphAdapter (EGA), a parameter-efficient fine-tuning (PEFT) approach designed to address these challenges. EGA is integrated into a pre-trained temporal backbone model as a GNN-based module, freezing the backbone and allowing only the adapter to be fine-tuned. This enables the effective acquisition of EEG spatial representations, significantly reducing computational overhead and data requirements. Experimental evaluations on two healthcare-related downstream tasks-Major Depressive Disorder (MDD) and Abnormality Detection (TUAB)-show that EGA improves performance by up to 16.1% in F1-score compared with the backbone BENDR model, highlighting its potential for scalable and accurate EEG-based predictions.

Figures (1)

• Figure 1: Model Architectures in Pre-training, and Downstream Tasks with Fully Fine-tuned BENDR and Proposed EGA with BENDR Frozen