One Model for All: Universal Pre-training for EEG based Emotion Recognition across Heterogeneous Datasets and Paradigms

Abstract

Electroencephalogram (EEG)-based emotion recognition holds significant promise but is hampered by the profound heterogeneity across datasets, particularly in channel configurations and subject variability, which hinders the development of generalizable models. Existing deep learning approaches often require dataset-specific designs and struggle to transfer knowledge effectively. To address this, we propose 'One Model for All', a novel universal pre-training framework designed for EEG analysis across disparate datasets and paradigms. Our core paradigm decouples representation learning into two stages: (1) Univariate pre-training using self-supervised contrastive learning on individual channels, enabled by a proposed Unified Channel Schema (UCS) that leverages the union of channels to maximize data utilization across datasets with varying layouts (e.g., SEED-62ch, DEAP-32ch); (2) Multivariate fine-tuning using a novel spatio-temporal architecture, combining an 'ART' (Adaptive Resampling Transformer) encoder with a 'GAT'(Graph Attention Network), to capture complex spatio-temporal dependencies for downstream tasks. Comprehensive experiments demonstrate the remarkable effectiveness of our approach. Universal pretraining is proven to be an essential stabilizer and performance booster, preventing training collapse on SEED (vs. failure from scratch) and yielding substantial gains on DEAP (+7.65%) and DREAMER (+3.55%). Our framework achieves new state-of-theart (SOTA) performance across all within-subject benchmarks: 'SEED (99.27%)', 'DEAP (93.69%)', and 'DREAMER (93.93%)'. We further demonstrate SOTA performance on cross-dataset transfer tasks, with our model achieving '94.08%' (via channel intersection) and '93.05%' (via UCS) on the unseen DREAMER dataset, with the former surpassing the within-domain pretraining benchmark. In-depth ablation studies validate our architectural choices, revealing that the GAT module is critical for SOTA performance, yielding a '+22.19%' gain over GCN on the high-noise DEAP dataset, while the removal of the graph network causes a catastrophic '16.44%' performance drop. This work paves the way for building more universal, scalable, and effective pre-trained models for diverse EEG analysis tasks.