BDR-GCL: Toward imagined speech decoding in naturalistic BCI systems via brain dynamics representation enhanced graph contrastive learning

• A new approach of natural brain-computer interface (BCI) through graph contrastive learning-based imagined speech decoding. • An innovative brain dynamics representation (BDR) module integrates multi-level brain dynamics to enhance negative sample representations. • A novel sample attention adjustment (SAA) strategy regulates hard and false negative samples considering the differences in brain states among subjects. • Superior performance and clear visualization analysis highlight the significance of BDR and SAA. Naturalistic brain-computer interface (BCI) aims to enable thought-driven interaction between brain and peripherals to enhance usability and promote adoption. Electroencephalogram (EEG)-based imagined speech decoding can directly translate mental intent into semantic commands for naturalistic BCI, and current decoding methods primarily leverage deep learning models. However, the small size of imagined speech datasets, coupled with intricate neural interactions among related brain regions, poses challenges to robust performance. Graph contrastive learning (GCL) addresses these challenges by constructing positive/negative contrastive samples, providing implicit topological supervision. Despite its potential, GCL has not yet been applied to imagined speech decoding, as dynamic, complex brain activity complicates the construction of discriminative negative representations. Moreover, current GCL methods treat all negatives equally, ignoring false negatives and the effect of imagined speech individual variability on negatives’ hardness. To tackle these issues, we propose a novel GCL method enhanced by brain dynamics representation (BDR-GCL), pioneering GCL’s application in EEG-based imagined speech decoding. Specifically, Brain Dynamics Representation (BDR) module generates more comprehensive imagined speech representations of negative samples by simultaneously learning EEG dynamics in brain region, connectivity, and network levels under the guidance of neuroscience prior knowledge. Considering the impact of hard/false negatives, we further design a Sample Attention Adjustment (SAA) strategy to evaluate negatives’ importance and introduce an attention contrastive loss via a weighting term. Experiments on two public datasets demonstrate that BDR-GCL achieves state-of-the-art performance. Visual analysis validates the crucial role of BDR and SAA in improving decoding and interpretability.