Design of an Asynchronous BMI with Interpretable Neural Networks for Exoskeleton Control: A Proof of Concept on Data Evolution and Scalability Over One Week.

This paper presents a concept study of a week-long experimental protocol for controlling a lower-limb exoskeleton via a brain-machine interface. The system employed a neural network adapted from EEGNet that distinguishes motor imagery and resting states in a two-dimensional space under both static and movement conditions. Each day, the model was fine-tuned with that day's training data as well as data from previous days. Daily closed-loop asynchronous evaluations were carried out to assess real-time exoskeleton control performance. The results indicate steady improvements in system accuracy over the week, likely due to the cumulative integration of additional data, which enhanced the neural network-based approach to cognitive state classification in a multi-day setting.Clinical relevance-Incorporating repetitive robotic therapies in which the patient can actively engage in rehabilitation is a core goal of neurorehabilitation. Developing non-invasive brain-machine interfaces that enable an increasingly effective mind-robot connection is of great importance. This work outlines a protocol for creating a brain-machine interface controlled by motor imagery.