Towards human activity recognition for ubiquitous health care using data from a waist-mounted smartphone

Abstract

Understanding human activities is a newly emerging paradigm that is greatly involved in developingubiquitous health care (u-Health) systems. The aim of these systems is to seamlessly gather knowledge about thepatient’s health and, after collecting knowledge, make suggestions to the patient according to his/her health profile.For this purpose, one of the most important ubiquitous communication trends is the smartphone, which has drawn theattention of both professionals and caregivers for monitoring the aging population, childcare, fall detection, and cognitiveimpairment. Recognizing human actions in a ubiquitous environment is very challenging and researchers have extensivelyinvestigated different methods to recognize human activities in the past decade. However, this field of research still needsfurther exploration in order to improve the accuracy and reduce the computational cost of these health care systems.Therefore, for expediting the existing system, this research work investigated a novel approach based on feature selectionand classification. In the proposed work, sparse Bayesian multinomial logistic regression (SBMLR) is used for featuresubset selection and a multiclass support vector machine (SVM) is adapted for the classification of six human dailyactivities (laying down, walking up stairs, walking down stairs, sitting, standing, and walking). For identifying the bestfeatures among the features returned by the SBMLR, a tuned threshold value is used for the selection of the features.Further, other classification algorithms including K-nearest neighbor, decision tree, and naive Bayes and different featureselection methods such as principal component analysis and random subset feature selection are also used for evaluationand comparison. The dataset used for testing is obtained from the UCI Machine Learning Repository. It is collectedby using a smartphone embedded with an accelerometer and gyroscope. The experimental results show that the highestaccuracy of 99.40% can be achieved by using the proposed method. Moreover, the paired sample two-tailed t-test overthe significance level of 0.05 reveals that the performance difference between the proposed technique and a competingtechnique is statistically significant.