Data mining-based analysis of the human activity in healthy subjects using smart phones

Human-activity-recognition (HAR) ascertains the nature of interaction with the surrounding environment. The present work quantifies the activities such as those of walking (WA), walking-upstairs (WU), walking-downstairs (WD), sitting (SI), standing (ST) and laying (LA) of 30 healthy subjects of age 19-48 years using accelerometer and gyroscope sensors embedded in Samsung Galaxy-S2 smartphone using data mining methods. Support vector machine (SVM), multiple layer perceptron (MLP), decision tree (DT), extra tree (ET), K-nearest neighbour (KNN), random forest (RF) and gradient boosting machine (GBM) techniques are used with and without linear discriminant analysis (LDA) for dimension reduction. The accuracy is seen to be higher with LDA. SVM (with C = 10, gamma = 0.001 with RBF kernel) provided the highest accuracy for both cases (SVM without LDA = SVM with LDA = 96%). However, the highest variation based on LDA was seen in DT (DT without LDA = 85% and DT with LDA = 95%). Such approaches can be extended in rehabilitative applications and virtual reality in the near future.