Unsupervised Machine Learning for Multi-Phenotype Anxiety Discovery: A Large-Scale Longitudinal Study of Workplace Physiology

Anxiety disorders affect more than 301 million people worldwide and remain the most prevalent class of mental health conditions. Detection methods have traditionally relied on subjective self-reports, which fail to capture the dynamic physiological signatures of anxiety in daily life. In this longitudinal study, continuous wearable monitoring was conducted with 67 working professionals over nine months, yielding 44,443 h of multimodal data segmented into 890,494 five-minute windows. An unsupervised approach was applied, and ten distinct autonomic states were identified through spectral clustering without the use of emotion labels during clustering. Among these, a rare dysregulated state was discovered, characterized by extreme peripheral cooling (d=−6.77), tachycardia (d=+4.41), and heightened electrodermal activity, showing a 31-fold enrichment for self-reported anxiety (p=0.0022). Beyond this extreme signature, anxiety was expressed heterogeneously across four physiological phenotypes ranging from acute dysregulation to mild perturbation. External validation with the WESAD dataset confirmed generalizability, while the absence of the rare state in laboratory stress paradigms highlighted its specificity to ecological anxiety. By integrating unsupervised discovery with post hoc self-report validation, this work advances our theoretical understanding of autonomic affect and provides a methodological foundation for scalable workplace mental health monitoring.