Comparison of Machine Learning Methods for the Construction of a Standalone Gait Diagnosis Device

Abstract

In this research, the authors investigate the feasibility of selecting three-dimensional thigh and shank angles as the features of machine learning methods. Four common machine learning techniques, i.e. random forest, k-nearest neighbour, support vector machine and perceptron, were compared in terms of accuracy and memory usage so that a real-time standalone gait diagnosis device can be constructed using low-end inertial measurement units (IMUs). With proper re-sampling and normalisation, they discovered that the support vector machine and perceptron resulted in the top two highest accuracies (96-99%) among the four machine learning methods. The memory requirement of the perceptron is the lowest among the machine learning methods. Therefore, perceptron was selected as the classification algorithm for the standalone gait diagnosis device. The trained perceptron was transferred to the thigh and shank's IMUs to process the data locally in real-time. The constructed standalone gait diagnosis device lit up green or red light emitting diodes when normal or abnormal gaits were detected, respectively. This standalone device was further tested in real-life and achieved a mean classification accuracy of 96.50%.