On-line anxiety level detection from biosignals: Machine learning based on a randomized controlled trial with spider-fearful individuals

• Published in: PloS one Vol. 15; no. 6; p. e0231517
• Main Authors: Ihmig, Frank R., H., Antonio Gogeascoechea, Neurohr-Parakenings, Frank, Schäfer, Sarah K., Lass-Hennemann, Johanna, Michael, Tanja
• Format: Journal Article
• Language: English
• Published: San Francisco Public Library of Science 23.06.2020; Public Library of Science (PLoS)
• Subjects: Accuracy; Algorithms; Anxieties; Anxiety; Augmented reality; Biofeedback; Biology and Life Sciences; Classification; Clinical psychology; Clinical trials; Closed loop systems; Computer and Information Sciences; Datasets; Diagnosis; EKG; Electrocardiography; Emotions; Engineering and Technology; Fear & phobias; Feasibility studies; Health aspects; Heart rate; Labeling; Labelling; Learning algorithms; Machine learning; Mathematical models; Medicine and Health Sciences; Physical Sciences; Physiology; Psychological aspects; Psychotherapy; R&D; Randomization; Ratings; Research & development; Research and Analysis Methods; Respiration; Skin; Social Sciences; Stress; Windows (intervals); Germany
• ISSN: 1932-6203; 1932-6203
• DOI: 10.1371/journal.pone.0231517

We present performance results concerning the validation for anxiety level detection based on trained mathematical models using supervised machine learning techniques. The model training is based on biosignals acquired in a randomized controlled trial. Wearable sensors were used to collect electrocardiogram, electrodermal activity, and respiration from spider-fearful individuals. We designed and applied ten approaches for data labeling considering individual biosignals as well as subjective ratings. Performance results revealed a selection of trained models adapted for two-level (low and high) and three-level (low, medium and high) classification of anxiety using a minimal set of six features. We obtained a remarkable accuracy of 89.8% for the two-level classification and of 74.4% for the three-level classification using a short time window length of ten seconds when applying the approach that uses subjective ratings for data labeling. Bagged Trees proved to be the most suitable classifier type among the classification models studied. The trained models will have a practical impact on the feasibility study of an augmented reality exposure therapy based on a therapeutic game for the treatment of arachnophobia.