Abstract
This article presents a quality-aware multimodal methodology for real-time stress recognition using wearable physiological tracking and motion sensors. The main contribution is the QA-soft fusion framework, which combines photo plethysmography (PPG) signal quality assessment, IMU-based motion control, adaptive soft fusion of PPG/Heart Rate Variability (HRV) and Electro Dermal Activity (EDA)+Temp branches, uncertain-window handling, adaptive decision windows, and calibrated intersubject validation. The wearable prototype is used as an experimental data acquisition platform rather than the primary novelty of this study. The proposed approach is designed to reduce false-positive stress recognition during physical activity, when PPG/HRV signals are strongly affected by motion artifacts. The fusion coefficient is adaptively adjusted according to the PPG signal quality index and motion intensity, reducing the contribution of unreliable PPG/HRV evidence when SQI decreases or ACCRMS increases. Low-quality or strongly motion-contaminated windows can also be marked as uncertain and excluded from alarm generation. Validation using LOSO/Group K-Fold showed that QA-soft reduced false stress alarms (false positives-FP) in class 2 from 0.260 ± 0.035 to 0.150 ± 0.025, with p < 0.01 and a large effect size of r = 0.68, while maintaining macro-F1, macro-PR-AUC, macro-ROC-AUC, Brier score, and Expected Calibration Error (ECE) at a level comparable to the baseline model. WESAD external validation and sensitivity analyses confirmed that the adaptive quality-aware fusion mechanism is responsible for the observed reduction in FP. These results indicate that QA-soft improves the robustness of wearable stress recognition under motion conditions and reduced PPG signal quality without degrading overall classification performance.