A Classification Approach Based on Directed Acyclic Graph to Predict Locomotion Activities With One Inertial Sensor on the Thigh

• Published in: IEEE transactions on medical robotics and bionics Vol. 3; no. 2; pp. 436 - 445
• Main Authors: Papapicco, V., Chen, B., Munih, M., Davalli, A., Sacchetti, R., Gruppioni, E., Crea, S., Vitiello, N.
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01.05.2021; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Active control; Algorithms; Body parts; Classification; Classifiers; directed acyclic graph; Feature extraction; Graph theory; inertial measurement unit; Inertial platforms; Inertial sensing devices; Legged locomotion; Locomotion; Locomotion mode classification; machine learning; Prostheses; Robot sensing systems; Sensors; Stairs; Steady state; support vector machines; Thigh
• DOI: 10.1109/TMRB.2021.3075096
• ISSN: 2576-3202

Current state-of-the-art locomotion mode classifiers for controlling robotic lower-limb prostheses rely on multiple sensors to achieve high accuracy, prediction performance, and robustness to both speed changes and subject-specific gait patterns. However, multiple sensors placed on different body parts usually entail discomfort and poor usability for the user. This paper presents an intention detection method that relies on the features extracted from an inertial measurement unit worn on the thigh and an online phase estimator. The algorithm classifies the locomotion mode of the upcoming stride among the three modes of ground-level walking, stair ascent, and stair descent. A two-stage classification process first distinguishes between transient and steady-state strides and then classifies the locomotion mode of the impending stride based on directed acyclic graphs of binary classifiers. The classification is performed at 75% or 85% of the previous stride phase, respectively for steady-state and transient strides. Data were gathered from 10 healthy subjects and processed offline. Feature design and selection were based on the data of all subjects, while the classification performance was assessed by leave-one-subject-out cross-validation. Results presented a median recognition accuracy of 98.7% for steady-state strides and 95.6% for transitions, suggesting that the method was inherently robust to variations in gait cadence, since all of the features were phase-based and not dependent on fixed time intervals. These results inform the design of control strategies for active transfemoral prostheses able to predict the user's locomotion intention during the next stride, using minimum sensors.