Identification of Gait Events in Healthy Subjects and With Parkinson's Disease Using Inertial Sensors: An Adaptive Unsupervised Learning Approach

• Published in: IEEE transactions on neural systems and rehabilitation engineering Vol. 28; no. 12; pp. 2933 - 2943
• Main Authors: Perez-Ibarra, Juan C., Siqueira, Adriano A. G., Krebs, Hermano I.
• Format: Journal Article
• Language: English
• Published: United States IEEE 01.12.2020; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Adaptive algorithms; Algorithms; Automatic control; Foot; Gait; Gait analysis; hidden Markov model; Hidden Markov models; human biomechanics; Inertial sensing devices; Learning algorithms; Legged locomotion; Machine learning; Mechanical engineering; Movement disorders; Neurodegenerative diseases; Parameter modification; Parkinson's disease; Pathology; Real time; Real-time systems; Resistors; robotic rehabilitation; Sensors; Training; Treadmills; Unsupervised learning; Walking; wearable sensors; Wearable technology
• ISSN: 1534-4320; 1558-0210; 1558-0210
• DOI: 10.1109/TNSRE.2020.3039999

Automatic identification of gait events is an essential component of the control scheme of assistive robotic devices. Many available techniques suffer limitations for real-time implementations and in guaranteeing high performances when identifying events in subjects with gait impairments. Machine learning algorithms offer a solution by enabling the training of different models to represent the gait patterns of different subjects. Here our aim is twofold: to remove the need for training stages using unsupervised learning, and to modify the parameters according to the changes within a walking trial using adaptive procedures. We developed two adaptive unsupervised algorithms for real-time detection of four gait events, using only signals from two single-IMU foot-mounted wearable devices. We evaluated the algorithms using data collected from five healthy adults and seven subjects with Parkinson's disease (PD) walking overground and on a treadmill. Both algorithms obtained high performance in terms of accuracy (F 1 -score ≥ 0.95 for both groups), and timing agreement using a force-sensitive resistors as reference (mean absolute differences of 66 ± 53 msec for the healthy group, and 58 ± 63 msec for the PD group). The proposed algorithms demonstrated the potential to learn optimal parameters for a particular participant and for detecting gait events without additional sensors, external labeling, or long training stages.