Citation

Lim, Choon Chen (2021) Recurrent Neural Networks For Stroke Rehabilitation in Virtual Reality. Masters thesis, Multimedia University.

Full text not available from this repository.

Abstract

Current virtual stroke rehabilitation system lacks rehabilitating both the impaired fingers and upper limb at one time. Most of the virtual rehabilitation modules require the attachment of wires and sensors to the body. Besides, they also lack patient performance outcome prediction using temporal information. Hence, an attractive and immersive rehabilitation system that applies contactless sensors in detecting motion is developed. The system also incorporates a patient performance prediction system that applies artificial intelligence technologies to estimate the rehabilitation outcomes. The first game application, ‘Pick & Place’ is integrated with a Leap Motion sensor to offer finger motor rehabilitation training. Besides, the second game application of ‘Stone Breaker’ utilises Microsoft Kinect sensor to provide upper limb motor training. The visual-based rehabilitation system is equipped with a machine learning method, namely, recurrent neural network (RNN). It is employed to closely track the performance of stroke patients when conducting the rehabilitation games. The recurrent neural network is trained by using the Broyden-Fletcher-Goldfarb-Shanno (BFGS) algorithm and Levenberg-Marquardt (LM) algorithm. A total of 60 participants (30 stroke patients and 30 control participants) are invited to join the experimental study. The stroke patients have achieved tremendous improvement when the difference of result (in Pick & Place and Stone Breaker) with the control group becomes minimal at the end of the session. Besides, the stroke group shows an excellent improvement in muscle power. The average grip strength is upgraded by 68%, and the upper limb power is improved by 11%. Moreover, the RNN exhibits a high accuracy performance in predicting the performance of the patients. The BFGS algorithm has a better predictive capability than the LM algorithm. It achieves a high average accuracy of 90.17% in Pick & Place and 93.60% in Stone Breaker in the performance estimation. The rehabilitation results also illustrate the existence of a prediction system in the rehabilitation module helps to enhance the patients’ performance. The RNN in this project outperforms four existing techniques that apply RNN to predict medical events. All the results show the potential to integrate the visual-based rehabilitation with the patient performance prediction system in restoring stroke patients’ impaired limbs.

Item Type:	Thesis (Masters)
Additional Information:	Call No: QA76.87 .L56 2021
Uncontrolled Keywords:	Neural networks (Computer science)
Subjects:	Q Science > QA Mathematics > QA71-90 Instruments and machines
Divisions:	Faculty of Engineering and Technology (FET)
Depositing User:	Ms Nurul Iqtiani Ahmad
Date Deposited:	22 May 2023 08:17
Last Modified:	22 May 2023 08:17
URII:	http://shdl.mmu.edu.my/id/eprint/11430

Downloads

Edit (login required)