Assessment of anxiety detection in students using machine learning and wearable sensors during virtual and in-person laboratory sessions

Abstract

Anxiety is an increasingly prevalent concern among graduate students, highlighting the need for objective approaches that complement conventional self-report assessments. Wearable sensing technologies combined with machine learning offer opportunities to characterize anxiety-related physiological patterns during authentic educational activities. This pilot study investigated the feasibility of multimodal physiological monitoring for participant-level anxiety screening in graduate engineering education. Eleven graduate students completed virtual and in-person Control Engineering laboratory activities while electroencephalographic (EEG) and peripheral physiological signals, including electrodermal activity (EDA), skin temperature (TEMP), heart rate (HR), inter-beat interval (IBI), and heart rate variability (HRV), were recorded using wearable devices. Participant-level anxiety reference labels were derived from a context-adapted administration of the Generalized Anxiety Disorder 7-item questionnaire (GAD-7). Six supervised machine learning classifiers were evaluated using subject-independent Leave-One-Subject-Out (LOSO) cross-validation. Model performance was assessed using accuracy, F1-score, ROC-AUC, and confusion matrices, complemented by feature importance analyses. Physiological differences were observed between participant anxiety groups and across virtual and in-person laboratory conditions, with EDA and temperature showing comparatively larger effect sizes than cardiovascular measures. Logistic Regression achieved the highest classification performance, although this result requires cautious interpretation given the limited sample size. All evaluated models identified anxiety-labeled participants without false negatives. Feature importance analyses further indicated that EEG-derived variables, particularly frontal and temporal Gamma and Beta activity, contributed substantially to discrimination. The findings provide preliminary evidence that integrating central and peripheral physiological signals with subject-independent machine learning evaluation may support anxiety screening during authentic educational activities. Rather than replacing established psychological assessment methods, wearable sensing may provide complementary objective evidence for wearable learning analytics and student well-being research. Given the small sample, single institutional context, and context-adapted GAD-7 reference labels, the results should be interpreted as evidence of feasibility rather than clinical validation.