Design and Implementation of an Artificial Intelligence-Driven Gait Phase Recognition System for Orthotic Knee Control

Abstract

Microprocessor-controlled stance-control knee-ankle-foot orthoses (M-SCKAFO) can have multiple sensors at all lower-limb segments. This causes M-SCKAFO to be bulky and expensive, with complex control systems. Stance-control systems with sensors local to the knee-joint component would provide a modular orthosis component for easier orthotist-customization and personalization for users with knee-extensor weakness. A gait phase recognition model (GPR) is essential for a fast, accurate, and generalizable real-time orthosis-control. This thesis designed, developed, and evaluated a machine learning-based GPR model for intelligent M-SCKAFO control. The model used gait signals that mimicked thigh inertial sensor and knee angle. Machine learning was implemented to identify gait phases across multiple surface conditions and walking speeds. Thigh-segment angular velocity, thigh-segment acceleration, and knee angle were calculated from 30 able-bodied participants for level and up, down, right-cross, and left-cross slopes at 0.8, 0.6, 0.4 m/s, and self-paced speeds (1.33 m/s, SD = 0.04 m/s). A logistic model tree (LMT) was built with a set of 20 signal features extracted from 0.1s sliding windows. The GPR model determined the walking state and was fed through a “transition sequence verification and correction” (TSVC) algorithm to deal with continuous states. The GPR model was evaluated on a different data set from 12 able-bodied individuals that completed the same walking protocol (validation set). Gait phases were classified successfully regardless of surface-level, walking speed, and individual walking variability. The LMT had a tree size of 1643 nodes with 822 leaf nodes. The GPR model produced overall classification accuracy of 98.4% and increased to 98.7% when TSVC was applied. Results also demonstrated evidence of strong model-generalizability with GPR accuracy of 90.6% and increased to 98.6% when TSVC was applied, on the validation set. This research demonstrated that local sensor signals from thigh and knee, integrated with machine intelligence algorithms, provided viable GPR suitable for real-time orthosis-control. The logistic decision tree model and feature selection approach were computationally efficient for real-time GPR and gave reliable, robust, and generalizable results across multiple surfaces, walking speeds, and individual walking variability. GPR also benefitted from transition sequence verification and correction algorithms, providing enhanced gait phase classification performance.