EEG-based Multimodal Sentiment Analysis plays a key role in building robust brain–computer interface (BCI) systems. However, current methods still face three main limitations: they often treat EEG signals as uniform, overlooking the region-specific mechanisms of emotion processing; they lack effective ways to capture both local and global spatiotemporal features of EEG; and they struggle to fully integrate EEG with complementary peripheral physiological signals. To address these issues, we propose BiMoE—a Brain-Inspired Mixture of Experts framework. BiMoE incorporates brain-topology-aware partitioning to model region-specific EEG dynamics, where each expert uses a dual-stream encoder to extract local and global contextual features. A separate expert processes PPS with multi-scale large-kernel convolutions. All experts are dynamically integrated through adaptive routing and optimized with a joint loss that encourages balanced, diverse, and accurate collaboration.

[2025/12] Submit and Open source

[2026/03] Accept by ICME 2026.

We conducted extensive experiments on the DEAP and DREAMER.

Due to the privacy policy, you need to apply for the dataset through the link, which is very simple.

DEAP: http://www.eecs.qmul.ac.uk/mmv/datasets/deap/

DREAMER: https://zenodo.org/records/546113

After download datastes, replace the data_path and save_path, run "save_deap.py" and "save_dreamer.py".

Run "load_deap.py" and "load_dreamer.py" to perform data set preprocessing.

Run "BiMoE_deap.py" and "BiMoE_dreamer.py".