Application of neural network to assess wheelchair driving abilities by power mobility road test

Abstract

The standard procedure for prescribing electric-powered wheelchairs for people with physical disabilities involves several steps, including the assessment of wheelchair driving skills by a healthcare professional who uses clinical tools developed by a researcher to assist in the evaluation process. However, one of the main problems found in the clinical setting is that these tools are generally dependent on the professional's judgment and experience, and can therefore be subjective. There are studies that evaluate the possibility of using objective metrics to determine whether an individual has the necessary skills to use a wheelchair. The use of virtual reality technologies allows obtaining such parameters; collecting this data could otherwise be too complex while using a real wheelchair, or may even present risks to the user's safety. The data obtained can then be used to aid the process of decision-making, and although such analyses may be performed by the health professional, it demands knowledge and an understanding of their meaning. Machine learning algorithms are presented as an alternative to automate the evaluation process by using neural networks with supervised training. The objective of the present thesis is the development and assessment of a system created using neural networks, using four objective metrics obtained through the EWATS wheelchair simulator, which was modeled according to tasks provided by the Power Mobility Road Test to evaluate the driving skills of the user. The selected metrics were: time elapsed during the task execution, number of commands sent to the joystick that controls the motorized wheelchair, number of collisions that occurred during the task, and the value of the root-mean-square error (RMSE, which evaluates the distance of the trajectory of a given object relative to the shortest or optimized trajectory). Experiments were carried out using two different groups, the first to provide data to train the neural network in the task evaluation process and the second to test the neural network post-training. Both groups were supervised by a health professional. Three classifier models were compared using a Wilcoxon Signed-Rank test: Multi-layer Perceptron (MLP), SVM (Support Vector Machine) and KNN (k-Nearest Neighbors). It was verified with statistical significance that the under the testing conditions the SVM obtained better prediction accuracy than both other models (80%), but several attributes could be explored in the design of the MLP to improve its accuracy. Further tests with a larger sample size and with greater representativity of the data are also necessary to obtain better classification results.