Objective anxiety level classification using unsupervised learning and multimodal physiological signals

Abstract

Anxiety disorders are prevalent worldwide and can negatively impact physical and mental health. Thus, the timely detection of changes in anxiety levels is crucial for mental health management. This study used multimodal physiological features from wearable devices to classify anxiety levels across various conditions, normalized by individual baseline responses for personalized analysis. Gaussian Mixture Models clustered data into binary or ternary anxiety levels, interpreted by statistics of self-reported scores and physiological features. Clus ters showed modest alignment with State and Trait Inventory scores and physiological markers and demonstrated task-specific variability. Silhouette scores indicated moderate separation (0.40 for two clusters, 0.14 for three clusters). Binary and three-class classifications using unsupervised learning and leave-one-participant-out validation demonstrated effectiveness, with Support Vector Machine achieving highest accuracies (90.9% and 73.3%). This approach enables objective, personalized anxiety monitoring without relying on subjective labeling.