Improved motor imagery BCI performance via task-unaware compression in the BELT Bayesian Edge-Cloud architecture

Description

Abolfazl Danayi and Hamid Soltanian-Zadeh present a new architecture for brain-computer interfaces in this 2026 paper. Such systems let people control devices by imagining movements, but many designs are tied to one task, need heavy retraining for each user. They also raise privacy worries when raw brain signals leave the device. Their framework, called BELT, splits the work between a small device at the user and a cloud server.

A stripped-down version built only from simple signal-processing steps was tested on two standard datasets of imagined hand and foot movements. After light personal tuning it classified movements accurately and ran faster on modest hardware than a well-known neural network, at a small cost in accuracy. A compression step that shrank the data more than threefold, without being trained for the task, barely hurt accuracy, which helps keep raw brain data private.

Catalog record

Source: https://doi.org/10.1371/journal.pone.0354976