Design, implementation and evaluation of a fitness application to aid in compliance and correctness of home exercises for back pain using PoseNet

Abstract

Background: Back pain has a high prevalence. It is costly for both the patient and the health system. Exercise therapy is the most prescribed treatment for back pain. Compliance is a serious issue with back pain exercises performed at home. This paper aims to design, evaluate and implement a web application to help with the compliance and correctness of back pain using the machine learning model PoseNet. Methods: A Google search was completed to identify which exercises are most prescribed by physiotherapists. 20 programs were studied and the most commonly used exercises were tabulated. PoseNet was tested for 4 outfits, 11 poses, 3 brightness levels and 6 different backgrounds. 105 images were taken and fed into PoseNet to evaluate its ability to correctly locate 17 body parts. The application was developed using web technologies: HTML, CSS, JavaScript, PHP, MySQL and several libraries. Results: The exercises found in more than 5 programs were: Bridge, Arm and Leg Lift, Back extension, Plank, Crunch, Side Plank and Squat. In the PoseNet evaluation it became clear that the model is extremely robust towards clothes, backgrounds and brightness. However, camera quality and body poses have a big effect on the pose estimation. A prototype web application using the squat as an exercise was implemented. The application provides the user with an index page, exercise program page, how to page, start exercise page, history page and logout. Once in the exercise, the application uses the coordinates of the body parts, calculates angles and displays guidelines and an avatar which change colour and position according to the correctness of the execution. All the critical information obtained during the exercise is saved automatically and displayed in a graph in a history page. Conclusion: PoseNet needs to be retrained or another model needs to be used to include exercises which are executed on the floor. Otherwise the application provides a promising step towards improving the correctness of exercise execution and solving the compliance problems encountered in back pain treatment, without the cost and hassle of some of the currently available solutions.