Loading...

Robustness Improvement of the PD Patients' Activity Recognition Algorithm in Presence of Variations in Patients' Motion Patterns (Inter-Class Variations)

Tariverdi, Amir Hossein | 2021

376 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 53728 (08)
  4. University: Sharif University of Technology
  5. Department: Mechanical Engineering
  6. Advisor(s): Behzadipour, Saeed
  7. Abstract:
  8. Parkinson’s disease is considered as a progressive neurodegenerative disease that hasn’t any certain treatment. In Iran until 1390, there were about 150 thousand patient struggling with this disease. Rehabilitation is known as an effective treatment to decrease destructive progress of the disease. Because of motional problems of PD patients, it is hard to come to the clinics. So developing remote rehabilitation would be interested by researchers and occupational therapists. Therefore in the recent years, an activity recognition system has been developed in Mowafaghian research center. This system is based on IMU sensors and a NM classifier.These systems are challenging with some problems, one of the most important ones is inter-class variations. These variations is due to differences of styles in performance of activities between subjects. These variations would decrease performance metrics (e.g. precision and recall) in recognition of activities. The purpose of this research is to make a robust algorithm against existence of variations. Process of problem solving will be categorized in three stage of identification, solving method study and evaluating methods. Motion data of 9 subjects (young men with average age of 22.65) in 34 activities (containing functional and LSVT-BIG activities) is used.In identification phase, a process of quantification of these variations in each activity is performed. In the result of this quantification, is observed that functional activities data has more of these variations in compare with other types of activities. In the method study phase, after checking out of the methods in literature, data augmentation with producing virtual data by amplitude warping is used to enrich database of the system. After changing control variables of the method, performance of the system is determined by use of leave-one-subject-out cross validation.After running cross validation for the original system, accuracy is to be about 68.55 percent. Similarity of some of activities, insufficiency of features for recognition of a group of activities and errors in data sorting are some reasons other than inter class variations, leads to deficiency in recall value. Data augmentation results in 2 percent accuracy increase, but changes in activities’ recall and precision values are sensible. Dynamic functional activities (except from standing/sitting and putting/picking up) have about 9 percent of increment in recall and precision. So can be concluded that these activities that have higher value of variation in variation quantification phase, after data augmentation would have an increment in recall and precision in compare with LSVT-Big and static functional activities
  9. Keywords:
  10. Parkinson Disease ; Data Augmentation ; Activity Recognition ; Tele-rehabilitation ; Warp in Time and Amplitude ; Iinter-Class Variation

 Digital Object List

 Bookmark

No TOC