Upper limb motion capturing and classification for unsupervised stroke rehabilitation

In recent years, substantial amount of researches have been carried out on cost-effective remote/unsupervised stroke rehabilitation methods due to the increasing number of post-stroke hospitalisation and the healthcare expenditure associated. This leads to the need of a reliable remote monitoring scheme that can assists medical specialists in monitoring the condition of patients. The information required to provide this remote monitoring system include the general physiological signals and patient's movement during the exercise. This information can be recorded and analysed to assess the patient condition.. Most of the conventional motion-capturing methods are visual based where large and expensive equipments are required and the tracking is generally limited in a certain area by the vision of the optical sensor. The non-visual based methods using inertial sensor suffer from drifting problem even when high-precision accelerometers and gyroscopes are used. This paper provides an alternative solution for patient's motion capturing using a low-cost non-visual based wireless sensor tracking system. Each sensor node only contains a 3-axis accelerometer and a micro-controller embedded communication module. Unlike the conventional systems, the output of the sensor, which is a mixture of change in acceleration and gravity influence, is directly used without extracting the precise acceleration and rotation information. Although the output from the sensors does not provide a direct image of the patient's movement, the waveform is still unique for all different types of motion and since the rehabilitation exercises are generally repetitive and pre-defined, the patient's movement can be classified by comparing the sensors' output to a series of templates. A series of experiments involving multiple subjects have been conducted and 200 movement samples were recorded and used to obtain templates, train classifier, and test system performance. K-Nearest Neighbor algorithm has been adopted in the experiment and the validation test has shown a promising result of a 97.22% overall classification accuracy from 36 independent testing inputs.