Classification of walking conditions of persons after stroke using accelerometers and gyroscopes

Abstract

Kinematic sensors have been widely applied in motion analysis. Together with the control and detection algorithm, they form a movement analysis system that can detect the external changes and also interpret those changes to control output. Recently, a number of researches have investigated automated movement detection using miniature kinematic sensors and various detection algorithms. Machine learning techniques were applied in analyzing gait patterns and features of normal walking. These methods were based on supervised machine learning approaches such as artificial neural network (ANN) and fuzzy logic etc. On the other hand, gait classification has been used for the diagnosis, assessment, and evaluation of pathological gait. In view of subjects with dropped foot, an optimal sensor set and a simple algorithm for the identification of gait phases could be useful for clinical evaluation, controlled movement and activities monitoring. The purposes of this study were to develop a kinematic gait measurement system using accelerometers and gyroscopes; and validate the method for gait phases detection in both normal and dropped foot gait after stroke. Support vector machine (SVM) is a new type of machine learning technique, which was employed to classify different walking conditions in this study. Ten hemiparetic subjects with dropped foot gait (Functional Ambulatory Category 3 to 4) and three healthy subjects participated in the study. The subjects were required to walk in five different conditions, namely level ground, stair ascent, stair decent, upslope, and downslope walking. The measurement system comprised three sensor units, each contained a dual-axis accelerometer and a single-axis gyroscope. The sensor units were attached at the thigh, shank, and foot segments to measure the segmental linear acceleration and angular velocity. The kinematic data was analyzed by a threshold method for detecting the turning points from which their time difference was compared with the force sensitive resistors attached underneath the foot. The reliability of the system to detect the stance and swing phases of gait was evaluated by the performance index (PI), with the ideal value set equal to 1, which depended on the classification accuracy and timing variation of the gait event obtained from the detection method. The performance of SVM was also compared with other machine learning techniques such as ANN and radial basis function (RBF) network. In the reliability study of the gait phases detection, it was found that some specific turning points could effectively identify gait events with a high value in the PI. The results using SVM classification showed that stair ascent and stair descent could be distinguished from each other, and from other walking conditions with 100% accuracy by using a single sensor unit attached to the shank segment in healthy subjects. For the classification in the five walking conditions, performance improved from 78% using the kinematic inputs from the shank sensor unit to 84% with two additional inputs from foot sensor unit. In the same classification tasks for the subjects with dropped foot, the classification accuracy of stair ascent, stair decent and other walking conditions were 92.9% using the kinematic inputs from the shank sensor unit. It was further improved to 97.5% with two additional inputs from foot sensor unit. Stair ascent was also classified by the inputs from foot sensor unit with an accuracy of 96%. The results of different classification methods showed that the performance of SVM was superior to other machine learning techniques using ANN and RBF networks. The kinematic measurement system and SVM method could be applied to classify various walking characteristics in both healthy subjects and subjects with dropped foot. This study shows an alternative system and method for gait analysis, pattern recognition and activity monitoring during rehabilitation.