End-to-End Decoding of Motion Intention from fNIRS Signals Using Deep Learning Framework

Motion intention decoding is an indispensable control link for active rehabilitation robots. In order to overcome the problem of low accuracy caused by limited time resolution of functional near-infrared spectroscopy (fNIRS) signal in intention recognition, this study proposed a hemodynamic gating module named HemoGate based on fNIRS signals for lower limb mobility action recognition. This module models fNIRS through time response modulation and spatial dependence reconstruction mechanism, and realized lower limb motion intention recognition through an end-to-end decoding method. Subject-independent and subject-dependent experiments were conducted, and the performance of the model was measured by accuracy and area under the curve (AUC). The results show that the recognition accuracy of the proposed model is higher than that of the traditional method. Among them, the best-performing subject can achieve an accuracy of 94.38% and an AUC of 0.9822. The proposed method is able to be deployed in wearable assisted walking robots, improving the performance and utility of brain-computer interfaces for lower limb intention recognition.