fNIRS-Based Action Detection for Lower Limb Amputees in Sit-to-Stand Tasks

Abstract

Traditional transfemoral lower-limb prostheses often overlook the intuitive neuronal connections between the brain and prosthetic actuators. This study bridges this gap by integrating a functional near-infrared spectroscopy (fNIRS) into real-time lower-limb prosthesis control with preliminary clinical tests on the above-knee amputee, enabling a more reliable volitional control of the prosthesis. Cerebral hemodynamic responses were measured using a 56-channel fNIRS headset, and lower-limb kinematics were recorded with a optical motion capture system. Artifacts in fNIRS were mitigated using short-separation regression, and eight features of the fNIRS data were extracted. ANOVA revealed the means, slope, and entropy as top-performing features across all subjects. Among eight classifiers tested, k-nearest neighbor (KNN) emerged as the most accurate. In this study, we recruited eleven healthy subjects and one unilateral transfemoral amputee. Classification rates surpassed 97% for all classes, maintaining an average accuracy of $99.86\pm 0.01$ %. Notably, the amputee exhibited higher precision, sensitivity, and F1 scores than healthy subjects. Maximum temporal latencies for healthy subjects were $120.00\pm 49.40$ ms during sit-down and $119.09\pm 45.71$ ms during stand-up, while the amputee showed maximum temporal latencies of 90 ms and 190 ms, respectively. This study marks the first application of action detection in sit-to-stand tasks for transfemoral amputees via fNIRS, which underscores the potential of fNIRS in neuroprostheses control.