BCIwiki (bciwiki.com) — An AI model combining EEG and visual gaze achieved 91.8% accuracy in attention recognition and 89.4% in cognitive load assessment. The study by Yabo Yang was published in Discover Artificial Intelligence on August 15, 2026.
The proposed multimodal framework uses Relational Graph Convolutional Networks (R-GCN) to model spatial topological relationships among EEG electrodes, and an AI-Arbitration-based Cross-Modal Attention Network (ACM-Net) that encodes visual gaze dynamics in real time to weight EEG spatial features. A multi-task learning architecture with a joint loss function enables independent quantitative outputs for attention level and cognitive load.
On public datasets MAHNOB-HCI, DEAP, and SEED, attention recognition accuracy reached 87.6%, 85.2%, and 88.1%, respectively, while cognitive load assessment accuracy was 84.3%, 82.7%, and 85.9%, with RMSE values of 0.251, 0.273, and 0.244. Compared with the best single-modal model, RMSE dropped to 0.218, an approximate 30.1% improvement. Ablation studies showed R-GCN spatial modeling and ACM-Net contributed most, and cross-subject validation indicated good generalization and robustness.