Politecnico di Torino (logo)

Development of Machine Learning-Based Algorithms for Assessing Tai Chi Exercise Proficiency in Older Adults

Andrea Valerio

Development of Machine Learning-Based Algorithms for Assessing Tai Chi Exercise Proficiency in Older Adults.

Rel. Danilo Demarchi. Politecnico di Torino, Corso di laurea magistrale in Ingegneria Biomedica, 2020

PDF (Tesi_di_laurea) - Tesi
Licenza: Creative Commons Attribution Non-commercial No Derivatives.

Download (6MB) | Preview

Development of Machine Learning-Based Algorithms for Assessing Tai Chi Exercise Proficiency in Older Adults In the past decade Tai Chi health benefits such as preventing fall risk, fighting social isolation, depression and improving cognitive capacity in older adults were widely investigated in order to give to physicians the possibility to give evidence-based recommendations to their patients regarding its use in rehabilitation purposes. The goal of the project is to develop a wearable system in order to generate meaningful feedback, to improve exercise performance and continuously monitor safety in individuals participating in a home-based program. This study, in particular, will focus on developing a machine learning algorithm to assess exercise proficiency in individuals performing Tai Chi. To do that thirty two Tai Chi practitioners were enrolled and they were asked to take part in data collection sessions in which, first, their physical eligibility had to be evaluated through a quantitative assessment tool known as BESTest and, finally, they were asked to perform different Tai Chi exercises. To collect data thirteen Shimmer units were put on each body segment of the the body of the subject. Each unit is composed by a 3-axes accelerometer, gyroscope and magnetometer. The beginning and the end of each task performed during the data collection was marked with an event marker so it was possible to keep only the part of the signal related to the tasks. Before starting the analysis, data needed to be processed. This was made in two main phases. Firstly, data needed to be checked, in order to be sure that each Shimmer unit had recorded properly the entire data collection and, then, check that the applied event markers were in the correct position. Once the controls were over, data were re-sampled and both band pass and low pass filtered. The next processing step was the segmentation of each repetition performed by the subject in each Tai Chi exercise. Depending on the task the number of repetitions could change from six to nine. To achieve this purpose, a semi-automatic algorithm based on accelerometer data was developed and then validated through a graphic user interface realized to check the results of the segmentation. From processed data, features were extracted and then selected through minimum redundancy maximum relevancy (mRMR) features selection algorithm which ranked them according to their importance and relevancy. Features selected were used to train a Random Forest model able to predict the proficiency level of the analyzed subject. The analysis was carried out using the recorded data and scores assigned, according to a specific visual-scoring system, by a Tai Chi master. Based on five criteria, designed to evaluate a specific aspect of the execution, ranging from 1 to 5, these scores tried to sum up the proficiency level of the subject according to the master point of view. The achieved results allowed us to draw conclusions concerning the position and the minimization of the number of wearable sensors needed to catch the most significant information reducing the noise added to the dataset by the redundant units. For the exercises present in the protocol the obtained predicted performances, according to the Leave-One-Subject-Out validation process, are generally satisfactory; indeed the averaged accuracy calculated among the exercises is around the 70%.

Relators: Danilo Demarchi
Academic year: 2019/20
Publication type: Electronic
Number of Pages: 87
Corso di laurea: Corso di laurea magistrale in Ingegneria Biomedica
Classe di laurea: New organization > Master science > LM-21 - BIOMEDICAL ENGINEERING
Aziende collaboratrici: Harvard Medical School
URI: http://webthesis.biblio.polito.it/id/eprint/13796
Modify record (reserved for operators) Modify record (reserved for operators)