Advisor Information
Hesham Ali
Location
Room 112
Presentation Type
Oral Presentation
Start Date
1-3-2019 9:00 AM
End Date
1-3-2019 10:15 AM
Abstract
A Comparative Study for Feature Selection Algorithms to Analyze Gait Patterns for Healthcare Purposes
Donovan Orn
Abstract
With the explosive use of wearable devices, there is an urgent need to find ways to utilize the data collected by such devices. In many cases, wearable devices contain accelerometers that collected different types of data. It is not always clear which data points can be used to extract information that can be used to provide useful knowledge about the health of the individuals. In this study, we proposed to research, implement, and comparing various feature selection techniques used for analyzing on gait parameters obtained from accelerometers attached to individuals’ ankles. The goal of this project is to identify the best feature selection algorithm for processing mobility parameters used to assess health and aid healthcare providers in the diagnosis of conditions associated with mobility impairment.
In order to achieve the proposed goal, different feature selection techniques are selected and implemented in conjunction with machine learning algorithms. The process will be executed in a five-step plan. 1) Data Acquisition, 2) Removal and Segmentation of data 3) Extract Features 4) Testing and Applying Feature Selection Techniques 5) Building Machine Learning Models. Based on previously reported results in various applications, selected feature selection techniques for implementation are genetic algorithms, mutual information, principal component analysis, and maximum information gain minimum correlation.
In the earlier stages of this research domain, the problem was not having enough data, which resulted in few insufficient features. Having less than necessary features can lead to creating models that are not accurate because the features do not have enough discriminating power.
With advanced technologies, such as wearable devices, it has become easier to collect more data. Consequently, the main research question in this area has now shifted from not having enough data to having too much data. Considering too much data for developing any predictive model may result in overfitting, which in turn leads to poor performance. Feature selection techniques allow researchers to remove the data that are irrelevant or redundant, allowing researchers to avoid overfitting wile reaching acceptable performance levels.
A Comparative Study for Feature Selection Algorithms to Analyze Gait Patterns for Health Purposes
Room 112
A Comparative Study for Feature Selection Algorithms to Analyze Gait Patterns for Healthcare Purposes
Donovan Orn
Abstract
With the explosive use of wearable devices, there is an urgent need to find ways to utilize the data collected by such devices. In many cases, wearable devices contain accelerometers that collected different types of data. It is not always clear which data points can be used to extract information that can be used to provide useful knowledge about the health of the individuals. In this study, we proposed to research, implement, and comparing various feature selection techniques used for analyzing on gait parameters obtained from accelerometers attached to individuals’ ankles. The goal of this project is to identify the best feature selection algorithm for processing mobility parameters used to assess health and aid healthcare providers in the diagnosis of conditions associated with mobility impairment.
In order to achieve the proposed goal, different feature selection techniques are selected and implemented in conjunction with machine learning algorithms. The process will be executed in a five-step plan. 1) Data Acquisition, 2) Removal and Segmentation of data 3) Extract Features 4) Testing and Applying Feature Selection Techniques 5) Building Machine Learning Models. Based on previously reported results in various applications, selected feature selection techniques for implementation are genetic algorithms, mutual information, principal component analysis, and maximum information gain minimum correlation.
In the earlier stages of this research domain, the problem was not having enough data, which resulted in few insufficient features. Having less than necessary features can lead to creating models that are not accurate because the features do not have enough discriminating power.
With advanced technologies, such as wearable devices, it has become easier to collect more data. Consequently, the main research question in this area has now shifted from not having enough data to having too much data. Considering too much data for developing any predictive model may result in overfitting, which in turn leads to poor performance. Feature selection techniques allow researchers to remove the data that are irrelevant or redundant, allowing researchers to avoid overfitting wile reaching acceptable performance levels.