A Basic Study on Detection of Movement State in Stride by Artificial Neural Network for Estimating Stride Length of Hemiplegic Gait Using IMU

In order to perform rehabilitation training for improving motor function, measurement of movements and evaluation of motor function become effective. In our research group, the method of estimating stride length during walking by using an inertial sensor attached to the foot was developed. However, since the method used thresholds to detect movement state in each stride for calculation of stride length, there was a difficulty in determination of threshold values for each subject and each stride with hemiplegic subjects in some cases. This study aimed at developing an automatic detection method of movement state in stride by artificial neural network (ANN) for hemiplegic gait. In this paper, three-layer ANN and four-layer ANN with feature extraction layers by autoencoder were tested. Teacher signals were obtained from measured sensor signals by the threshold-based method. The ANN with feature extraction layers was shown to be effective for detecting the movement state of healthy subjects and a hemiplegic subject. The movement state detected by ANN was also suggested to be effective in stride length estimation. It is expected to evaluate the ANN-based method using data measured with more hemiplegic subjects.