Robust and Explainable Wearable Stress Recognition under Physical Activity Confounding

Abstract

Wearable stress-recognition models may learn movement-related shortcuts when stress and exercise are evaluated together. This study audits that risk in a subject-independent three-class task containing stress, aerobic exercise, and anaerobic exercise sessions from a public wrist-worn sensing dataset. Sixty-second windows with fifty percent overlap produced 6,269 samples and 151 features from accelerometry, blood volume pulse, electrodermal activity, and temperature. Seven classifiers were evaluated through modality ablation, calibration assessment, sensor dropout, Gaussian noise, random missingness, activity-intensity matching, and permutation importance. The best clean configuration was an accelerometer-only radial basis function support vector machine, achieving 0.742 accuracy, 0.715 balanced accuracy, a 0.709 macro-averaged harmonic mean of precision and recall, and a 0.883 macro-averaged area under the receiver operating characteristic curve. Its macro-averaged classification score exceeded the corresponding all-sensor model by 0.033 and the model without accelerometer features by 0.120. After fold-wise matching on accelerometer-magnitude energy, the maximum standardized mean difference decreased to between 0.012 and 0.073, while the macro-averaged classification score changed only from 0.709 to 0.701. Random missingness caused a clearer severity-dependent degradation than Gaussian noise. The results show that predictive performance, probability quality, and sensor dependence should be evaluated jointly and that accelerometer dependence cannot be explained solely by aggregate activity intensity.