سامانه همایش‌ها TSTA | ثبت‌نام و ارسال مقاله کنفرانس‌ها و کنگره‌های ملی

Home / سی و دومین کنفرانس ملی و دهمین کنفرانس بین المللی مهندسی زیست پزشکی ایران

Hierarchical Task-Structured GNN Meta-Learning for Few-Shot EEG Motor Imagery Decoding

Authors :

Mohammad Armin Dehghan¹ Mohammad Mohammadianbisheh² Mohammad Bagher Shamsollahi³

1- دانشگاه صنعتی شریف، تهران، ایران 2- دانشگاه صنعتی شریف، تهران، ایران 3- دانشگاه صنعتی شریف، تهران، ایران

Keywords :

EEG Signals،BCI Decoding،Meta-learning،Graph Neural Networks (GNNs)،Subject-level Adaptation،Hierarchical Task Structures

Abstract :

Motor imagery classification from electroencephalo- gram (EEG) signals is a core challenge in brain–computer in- terface (BCI) systems. Yet, strong inter-subject variability, where each subject follows a distinct distribution, renders conventional learning approaches poorly suited for generalization to unseen subjects. Few-shot meta-learning offers a promising alternative by enabling rapid adaptation to new subjects with only limited labeled data. At the same time, neuroscience evidence emphasizes that EEG decoding should leverage network-level interactions rather than treating electrodes as independent sources, moti- vating graph-based representations. In this work, to leverage network-level structure, We propose a principled graph construc- tion pipeline to represent EEG data. Also to enable subject- level adaptation in few-shot settings, we use a meta-learning framework that learns Hierarchical Task Structures, through which we exploit inter-subject correlations, and employ GNN architectures as the learner. Experiments on the PhysioNet motor imagery dataset show that our method achieves over 10% higher accuracy than baseline models, while reducing variance across subjects by roughly 10%. This demonstrates that combining graph-based representations with few-shot meta-learning yields more reliable and subject-adaptive BCI systems.