EEG-based Spatial-Channel Interaction Attention Neural Networks for Detecting Empathy in Motor Collaboration.

Abstract

Embodied intelligence and humanoid robots aim to mimic interpersonal interactions to achieve affective human-robot interaction (HRI). A major challenge in advancing HRI lies in effectively emulating interpersonal affective interactions and evaluating the resulting artificial empathy. To address these challenges, we propose SpatialChannel Interaction Attention Neural Networks (SCIANN)-a novel EEG-based architecture that combines topological brain activation and connectivity patterns to decode empathy in motor collaboration. A private EEG dataset from a collaborative brain-computer interface motor control experiment and a public EEG dataset from a dyadic perceptual crossing experiment were used for evaluating SCIANN's performance with comparisons with five baseline models. Results showed that SCIANN outperformed the state-of-the-art baseline models. In the private dataset, SCIANN reached an accuracy of 100% both in inter-subject and cross-subject tests for detecting whether empathy is induced or not. For classifying 4-class empathy, it achieved an accuracy of 98.3% in the inter-subject test, and 48.1% in the cross-subject test. In the public dataset, SCIANN reached a classification accuracy of 92.2% in inter-subject and 91.7% in cross-subject tests for detecting whether empathy is induced or not. Feature visualization results revealed that contributing EEG channel importance features and channel interaction features aligned with established neurophysiological findings. These results collectively demonstrate SCIANN's potential as a robust, generalizable framework for artificial empathy assessment in HRI applications.