Evaluating feature selection for stress identification

In modern society, more and more people are suffering from stress. The accumulation of stress will result in poor health condition to people. Effectively detecting the stress of human being in time provides a helpful way for people to better manage their stress. Much work has been done on recognizing the stress level of people by extracting features from the bio-signals acquired by physiological sensors. However, little work has been focused on the feature selection. In this paper, we propose a feature selection method based on Principal Component Analysis (PCA). After the features are selected, their effectiveness in terms of correct rate and computational time are evaluated using five classification algorithms, Linear Discriminant Function, C4.5 induction tree, Support Vector Machine (SVM), Naïve Bayes and K Nearest Neighbor (KNN). We use the driver stress database contributed by MIT Media lab for our experiments. Leaving one out as well as 10-fold data preparation approach is implemented as the cross validation method for our evaluation. Paired t-test is then performed to analyze and compare the experimental results, as well as for their statistical significance. Our study demonstrates the importance of feature selection and the effectiveness of the methods used in accurately classifying stress levels.