TRANSIT-EEG-A Framework for Cross-Subject Classification With Subject Specific Adaptation

• Published in: IEEE transactions on cognitive and developmental systems Vol. 17; no. 4; pp. 923 - 937
• Main Authors: Ahuja, Chirag, Sethia, Divyashikha
• Format: Journal Article
• Language: English
• Published: IEEE 01.08.2025
• Subjects: Accuracy; Adaptation models; Analytical models; Brain modeling; Cross-subject transfer; Data augmentation; Data models; diffusion model; Electroencephalography; few-shot learning; graph neural network (GNN); Graph neural networks; LoRA; Pattern classification; Training; transfer learning
• ISSN: 2379-8920; 2379-8939
• DOI: 10.1109/TCDS.2025.3529669

Electroencephalography (EEG) is pivotal in monitoring and analyzing cerebral activity across diverse domains, including medical diagnostics, cognitive neuroscience, and brain-computer interfaces. However, the inherent intricacy of EEG signals and their subject-specific characteristics pose formidable challenges in devising robust and generalizable classification models. Traditional EEG signal classification paradigms rely on extensive subject-specific datasets. Also, the domain adaption for new subjects often leads to "catastrophic forgetting," thereby diminishing the performance of model trained on prior subjects. This article proposes a novel framework, transfer, and robust adaptation of new subjects in EEG technology (TRANSIT-EEG), designed to adapt adeptly to new subjects. TRANSIT-EEG demonstrates resilience to subject-specific artifacts by integrating synthetic data generation using the proposed subject-specific augmentation model - individualized diffusion probabilistic model (IDPM). Also, it employs a robust self organising graph attention transformer (SOGAT) that dynamically constructs a graph for each subject, fostering a more accurate classification. Moreover, TRANSIT-EEG introduces adapter-based finetuning using low-rank adaptation (LoRA) for new subjects, enriching the adaptation process. The TRANSIT-EEG framework presents a promising avenue for advancing the realm of EEG signal classification. Evaluation of widely studied datasets, specifically focusing on two significant tasks, SEED for emotion recognition and PhyAat for auditory activity recognition, substantiates the efficacy and versatility of TRANSIT-EEG. This validation indicates a substantial stride toward achieving more generalizable and accurate EEG signal classification.