WEARABLE PSYCHOPHYSIOLOGICAL STRESS MONITORING SYSTEM: PROTOCOL, DATASET AND REPRODUCIBLE ML‑CONVEYOR BELT

The paper considers the task of monitoring psychophysiological stress under conditions close to real life, using a wrist-mounted wearable device and a reproducible machine-learning pipeline. An experimental protocol with three classes of states — rest, induced psychosocial stress and standardized physical activity — implemented on a sample of 12 subjects is proposed. The multimodal device registers PPG/HRV, EDA/GSR, skin temperature, and IMU signals. The data is segmented into 60-s windows in 15-s increments, and preprocessing and feature extraction are performed. An ML pipeline has been built using an intersubject LOSO validation scheme and a nested selection of hyperparameters for basic models (logistic regression, SVM, Random Forest, gradient boosting), accounting for class imbalance. The best configuration reaches macro-F1 ≈ 0.68, ROC-AUC ≈ 0.90 and PR-AUC ≈ 0.71, noticeably surpassing the majority basic classifier; disabling IMU leads to a drop in macro-F1 to ≈ 0.60, which emphasizes the importance of movement to separate stress and physical activity. Additionally, the calibration of probabilistic predictions was carried out, reducing the Brier score from 0.19 to 0.14 and bringing the reliability diagrams closer to the ideal bisector. The main contributions of the work are: (a) a protocol and an open dataset format for the three-class task "rest/stress/physical load", (b) a reproducible ML pipeline with correct intersubject assessment and calibration report, (c) a quantitative analysis of the contribution of sensory modalities and the resilience of models to motion artifacts.