Classification Accuracy of Support Vector Machine, Decision Tree and Random Forest Modules when Applied to a Health Monitoring with Flexible Sensors

In this study, distinct machine learning-based methods of analysis were used to evaluate the most effective feature combination for a previously proposed system which detects changes in the physical conditions of human beings through eight sensors placed on various locations of a bed and floor near...

Full description

Saved in:
Bibliographic Details
Published inSensors & transducers Vol. 245; no. 6; pp. 83 - 89
Main Authors Uchida, Yasutaka, Funayama, Tomoko, Kogure, Yoshiaki
Format Journal Article
LanguageEnglish
Published Toronto IFSA Publishing, S.L 01.10.2020
Subjects
Accuracy
Datasets
Decision trees
Evaluation
Feature extraction
Flexible components
Machine learning
Methods
Monitoring systems
Multilayer perceptrons
Older people
People with disabilities
Regression analysis
Sensors
Support vector machines
Values
Online AccessGet full text

Cover

More Information
Summary:In this study, distinct machine learning-based methods of analysis were used to evaluate the most effective feature combination for a previously proposed system which detects changes in the physical conditions of human beings through eight sensors placed on various locations of a bed and floor near the bed. First, the proposed system was applied to evaluate the presence or absence of restrictions in the movement of the knee joint in two subjects of different age. A set of measurements were obtained from four classes, under distinct scenarios, with 10 case in each class. Then, 30 % of the measured feature values were randomly extracted for testing. Scikit-learn was used as the machine learning library, and the k-nearest neighbor method, logistic regression, perceptron, and multilayer perceptron were used as the four learning models in the first step. The k-nearest neighbor method assigned the highest accuracies to all three feature value combinations that were investigated in this study. The accuracy and feature value of the distribution, which could be high or low, were investigated. It was found that the analysis methods corresponding to the feature value at specific locations were effective. The decision tree method, random forest and support vector machine methods were found to meet our proposed system. The accuracies ranged from 0.6 to over 0.8 when t2-t7 and t4-t7 feature values were combined.
ISSN:2306-8515
1726-5479