Machine Learning Driven Individualized Gait Rehabilitation: Classification, Prediction, and Mechanism Design

Abstract

Design of mechanisms for human-machine interaction involves numerous subjective criteria and constraints in addition to the kinematic task. This is particularly important for the rehabilitation devices, where the size, complexity, weight, cost, and ease of use are critical factors. A large majority of the approaches towards the design of such devices, which are based on limited degree-of-freedom mechanisms start with finding numerically optimal solutions to the task path followed by pruning for feasible design concepts. Given the highly nonlinear nature of the problem, this approach discards a large proportion of numerically sub-optimal solutions, which could potentially be pragmatically optimal solutions if the subject criteria were applied from the start. To overcome this limitation, in this paper, we present an end-to-end computational approach for developing a device for individualized gait rehabilitation using machine learning techniques focusing on gait classification, prediction, and specialized device design. These models generate a distribution of linkage mechanisms, which strongly correlate to the distribution of target path variations. This way of formulating the problem results in a large variety of solutions to which subjective criteria can be applied to yield practically useful design concepts that would otherwise not be possible using traditional synthesis methods.