Cognitive load-aware VR rehabilitation: Real-time EEG–EMG fusion for adaptive task difficulty adjustment

Real-time fusion of electroencephalography (EEG) and electromyography (EMG) is essential for cognitively aware virtual reality (VR) rehabilitation systems that adapt task difficulty based on the user state. This work presents an enhanced NeuroFusion-based architecture incorporating Self-Attention Multi-Temporal Convolutional Networks (SAMTCNs) and a Task-Aware Difficulty Adjustment (TADA) module to address key challenges in existing EEG–EMG fusion approaches, including temporal misalignment, limited cross-modal interaction, and poor adaptability. Extending the NeuroFusion-Trans framework, which employed transformer-based cross-modality attention and online learning for assistive robotics, the proposed system enables adaptive VR rehabilitation by dynamically modifying task complexity according to real-time cognitive load estimated from multimodal physiological signals. While NeuroFusion-Trans achieved high classification accuracy (up to 97%), it lacked real-time feedback loops and user-specific adaptation within immersive environments. To overcome these limitations, an SAMTCN is employed for fine-grained temporal feature extraction across EEG and EMG signals, while TADA continuously regulates exercise intensity based on decoded cognitive and motor states. The system synchronizes brain and muscle activity in real time, decodes user intent, and adjusts rehabilitation tasks to deliver a personalized and responsive training experience. Experimental evaluation on EEG–EMG gesture datasets augmented for VR-based motor rehabilitation demonstrates that the proposed model outperforms baseline methods and the original NeuroFusion-Trans architecture, achieving a classification accuracy of 98.2%, an F1-score of 0.981, and a synchronization score of 0.79. Ablation studies reveal performance degradations of 6.3% without TADA and 5.8% without the SAMTCN, confirming the critical contribution of each component and highlighting the model’s robustness, generalizability across users, and stability under physiological variability.