Find in Library

This study investigated the potential of recognising arousal in motor activity collected by wrist-worn accelerometers. We hypothesise that emotional arousal emerges from the generalised central nervous system which embeds affective states within motor activity. We formulate arousal detection as a statistical problem of separating two sets - motor activity under emotional arousal and motor activity without arousal. We propose a novel test regime based on machine learning assuming that the two sets can be distinguished if a machine learning classifier can separate the sets better than random guessing. To increase the statistical power of the testing regime, the performance of the classifiers is evaluated in a cross-validation framework, and to test if the classifiers perform better than random guessing, a repeated cross-validation corrected t-test is used. The classifiers were evaluated on the basis of accuracy and Matthew’s correlation coefficient. The suggested procedures were further compared against a traditional multivariate paired Hotelling’s T-squared test. The classifiers achieved an accuracy of about 60%, and according to the proposed t-test were significantly better than random guessing. The suggested test regime demonstrated higher statistical power than Hotelling’s T-squared test, and we conclude that we can distinguish between motor activity under emotional arousal and without it.