BCIwiki (bciwiki.com) — A hybrid batch-streaming framework for EEG-based emotion recognition reached classification accuracies of 85%, 94% and 74% on the DEAP, AMIGOS and SEED benchmark datasets, researchers at the Faculty of Computer Engineering, K N Toosi University of Technology in Tehran reported in Neural Networks on May 20, 2026 (PMID 42155274; DOI 10.1016/j.neunet.2026.109072).
Existing brain-computer interface (BCI) emotion recognition models typically run either offline in batch mode or online in streaming mode, the authors wrote. Batch approaches tend to be more accurate but slow, while streaming approaches work in real time at the cost of accuracy.
The proposed architecture uses a probabilistic switching mechanism that estimates the reliability of the streaming module from its historical performance. That reliability score sets the probability of taking the output from either the batch or the streaming unit. The team also tested how the switch performed and how system components held up against concept drift, where data distributions shift over time.
The authors said the method balances classification accuracy and computational efficiency and expect such hybrid designs to be widely used in feedback-based systems. The evaluation relied on public datasets; the abstract does not report tests with live users.