New AI method boosts wearable stress detection by 48% with decision-level fusion

A new paper from researchers at the University of Southampton introduces a decision-level fusion framework for robust affect recognition from wearable physiology. Led by Lokesh Singh and Athina Georgara, the team addresses a critical challenge: real-world deployments of affective computing are plagued by non-stationary signals, motion artifacts, and missing sensors. Using the WESAD dataset (15 subjects, five sensor modalities—ECG, EDA, BVP, EMG, ACC—across baseline, stress, and amusement conditions), they propose a non-stationary pipeline that replaces traditional fixed-basis spectral features (e.g., FFT bandpower) with Fourier-Bessel Series Expansion (FBSE) combined with Empirical Wavelet Transform (EWT) for data-driven spectral segmentation. This captures short-lived, discriminative patterns that conventional methods oversmooth. For multimodal integration, they adopt decision-level aggregation: each modality has its own predictor, and outputs are weighted by predictive uncertainty and modality reliability. The results show decision-level aggregation is at least as good as feature-level 84% of the time and strictly better 48% of the time, indicating significantly improved robustness under heterogeneous and partially reliable sensing conditions.

This work has direct implications for public health, preventive care, and stress-aware interventions. By making affect recognition robust to real-world imperfections—like a worn sensor losing contact or a sudden movement—the method brings wearable emotional monitoring closer to practical deployment. The researchers emphasize that their approach does not sacrifice accuracy for robustness; instead, the uncertainty-based weighting actively filters unreliable modalities while amplifying reliable ones. The paper (arXiv:2605.14878) also highlights that the decision-level framework is modular, allowing easy integration of new sensor types without retraining the entire system. For tech professionals building health wearables, this provides a clear path to more dependable affective computing, potentially enabling applications like stress tracking, early mental health screening, and adaptive interventions that respond to a user's emotional state in real time.