Cross-task, explainable, and real-time decoding of human emotion states by integrating grey and white matter intracranial neural activity

Decoding human emotion states from neural activity has significant applications in human-computer interfaces, psychiatric diagnostics, and neuromodulation therapies. Intracranial electroencephalogram (iEEG) provides a promising modality for decoding by balancing temporal, spatial, and noise resolutions. However, real-world application of decoding requires high performance that integrates neural activity from both grey and white matter, stable generalization across different contexts, sufficient neural encoding explainability, and robust real-time implementation, all of which remain elusive. Here, we simultaneously recorded iEEG and abundant self-rated valence and arousal scores--measuring the two primitive dimension of emotion--across two emotion-eliciting tasks in eighteen epilepsy subjects. We developed self-supervised deep learning models that achieved high-performance decoding by integrating grey and white matter signals, generalizing across tasks. The models provided strong explainability by revealing shared and preferred mesolimbic-thalamo-cortical subnetworks encoding valence and arousal, as well as the structural connectivity basis underlying grey and white matter integration. Finally, the models were implemented online and realized robust real-time decoding in four new subjects. Our results have implications for advancing emotion decoding neurotechnology and understanding emotion encoding mechanisms.