Impact of Subject-specific Training Data in Anxiety Level Classification from Physiologic Data

Abstract

The autonomic nervous system is known for the fight or flight response. Anxiety affects the autonomic nervous system, causing heightened heart rate and electrodermal activity. This paper explores machine learning methods to predict two- and three-level anxiety in spider fearful individuals watching spider video clips in a controlled trial. Features are extracted from electrocardiogram and electrodermal time-series signals. Specifically, this paper explores the performance of such models as the amount of data pertaining to the test subject increases in the training set. Standard K-fold cross-validation is here compared to leaky group-fold cross-validation with sample imputation, where we systematically vary the the number of samples from the test subject that are included in the training set. While it is possible to reach 78% and 60% k-fold accuracy for a two- and three-level anxiety prediction, respectively, excluding all test subject data from the training set causes the accuracy to drop to 73% and 45%. The results demonstrate that the features and models used here do not generalize for inter-subject classification tasks and that care should be taken when splitting subject data between training and test data. Furthermore, our results address the "cold start problem" by providing an indication of how much data would be required from a new subject before accurate prediction of anxiety is possible from physiologic data.