BCIwiki (bciwiki.com) — A model that fuses EEG signals with multilingual natural language models identified emotions with 89.7% accuracy, compared with 75% for EEG alone and 79% for text alone, according to a paper by Disha Sushant Wankhede published in Natural Resources for Human Health on October 1, 2026.
The author argues that brain-computer interface (BCI) data are highly sensitive to small differences and noise, while multilingual language models infer sentiment from words but cannot detect implicit brain signals. EEG was recorded during emotional stimulation and preprocessed with Electromagnetic Decomposition Continuous Wavelet Transform (ECD-CT) to reduce noise and extract features; text was tokenized and converted into embeddings. A Dynamic Attention Joint Embedding (DAJE) model merges the two modalities, and a Robust Cross Modal Optimization (RCMO) technique refines the fused model for accuracy and generalization.
The resulting Adaptive Real time Emotion Recognition System (ARERS) achieved an F1-score of 0.88 with an average inference latency of 180 ms, which the author says supports low-latency, language-independent real-time emotion detection. The author suggests extending the work to mental health monitoring and to cross-cultural and cross-linguistic settings. The abstract does not report the number of participants, the datasets used or the evaluation protocol.