Identification of Gait Events in Healthy and Parkinson's Disease Subjects Using Inertial Sensors: A Supervised Learning Approach

• Published in: IEEE sensors journal 2020-12, Vol.20 (24), p.14984-14993
• Main Authors: Perez-Ibarra, Juan C., Siqueira, Adriano A. G., Krebs, Hermano I.
• Format: Article
• Language: English
• Subjects: Algorithms; Classifiers; Event detection; Foot; Gait; Gait analysis; gait phases; Heuristic methods; Inertial sensing devices; Legged locomotion; Loading; machine learning; meta-heuristics; Optimization; Parkinson's disease; Phases; Real-time systems; Sensors; Simulated annealing; Supervised learning; Support vector machines; Treadmills; Walking; wearable sensors; Wearable technology
• ISSN: 1530-437X; 1558-1748
• DOI: 10.1109/JSEN.2020.3011627

Automatic detection of gait phases through supervised learning is a feasible approach that takes advantage of the consistency of the gait cycle among healthy subjects. However, gait patterns among subjects with impairments are not as consistent and most of the existing algorithms have limited performance detecting phases during impaired gait. In this paper, we proposed one algorithm that used linear classifiers to detect in real-time the transition between consecutive gait phases. Our approach is a generalization of the rule- and threshold-based algorithms for event detection. Linear classifiers are parametric models that require appropriate values in order to perform correct classification of the gait phases. We introduced a modified Support Vector Machine (SVM) to compute such sub-optimal combinations of those values, and a further optimization with a hybrid meta-heuristic approach that integrates a Genetic and a Simulated Annealing Algorithm. We tested our approach on data collected by a single-IMU foot-mounted wearable device during overground and treadmill walking for two groups: one with healthy and one with Parkinson's Disease subjects. The {F}_1 -scores were 0.987 and 0.953 for the two groups, which were comparable with our previously developed threshold-based method, which obtained 0.988 and 0.974, respectively. Our proposed approach achieved similar performance as the threshold-based scheme, with the advantage of not relying on any prior knowledge of specific features for any particular inertial signal.