Assessment of temporal parameters of gait in hemiplegic children by machine learning approach

Hemiplegia is a pathology caused by a neurological disorder and is quite frequent in children with cerebral palsy. A lot of methods have been proposed to perform gait analysis in order to understand how muscles activate during walking activity, so that it is possible to understand how hemiplegia affects muscular recruitment. To this aim, there are techniques based on the use of classi sensors, but the most recent approaches consist in using machine/deep learning techniques. In the present work, artificial neural network was employed with the aim of classifying the two main gait phases (stance and swing periods) and to predict foot-floor contact signals from surface electromyographic (sEMG) signals in hemiplegic children. To this purpose, sEMG signals from the main leg muscles and foot-floor-contact signals were acquired during walking at self-selected pace of 20 hemiplegic children. These data were then fed as input to a feed-forward multi layer perceptron (MLP) neural network. Successively, prediction evaluation was performed to assess the goodness of the chosen model for the neural network. Sensitivity analysis of classification and prediction performances to the processing parameters was also performed. Results show as acceptable levels of accuracy in gait-phases classification and gait-event prediction were reached both for learned and unlearned subjects. Although there is still work to be done, deep learning approach was proved to be a reliable tool for gait analysis, also in this preliminary attempt. The fact that with deep learning approach gait events are extractable from sEMG signals is useful because in this way it is possible to avoid the use of foot-switch sensors and, consequently, to perform gait analysis with a smaller number of sensors and thus with reduced costs, time-consumption, and patient invasiveness.