Feature selection for stress level classification into a physiologycal signals set

This paper describes the methodology and results obtained when classifying two states of stress, low and high using a data base from Physionet that contains the recordings of physiological signals under several stress conditions. The signals were first denoised and then, features were extracted for segments of 5 minutes. Four out of 6 signals were chosen: heart rate variability, respiration, galvanic skin response from the hand, and galvanic skin response from the foot. Two non-linear features were extracted: approximate entropy and correlation dimension, both with m=2 and m=3. Besides, three linear features were extracted: energy, mean and standard deviation. Five machine learning classifiers were compared: K-nearest neighbours, Support vector machines with a linear kernel, support vector machines with a Gaussian kernel, Naïve Bayes classifier, Random forest classifier and logistic regression. It was found that approximate entropy and correlation dimension with m=3 provide the greater differences between the two stress states. It was also found that choosing only three physiological signals and correlation dimension with m=3 the logistic regression classifier achieved and accuracy of 81.38%, the best performance compared to other combinations of signals and classifiers. The three physiological signals that provided the best features were heart rate variability, respiration and galvanic skin response on the foot.