Importance of Individual Differences in Physiological-Based Stress Recognition Models.

Stress is well-researched. Still, despite the potential economic and health benefits of a system that continuously monitor people's stress, there exists no mainstream real-world stress monitoring system. The most reliable methods use a fusion of multi-modal signals. However, these methods are both obtrusive and privacy-invasive. On the contrary, the most practical ones are often based on physiological signals. Nevertheless, while these methods may perform exceptionally well in research field trials, their results are yet to be incorporated in any practical, real-world stress recognition system. This paper argues that many of the published physiological based machine learning stress recognition models may not be practical to be used in real-world settings. The analysis conducted using electrodermal activity (EDA) and heart rate variability (HRV) seems to indicate that physiological-based stress recognition machine learning models perform well when tested on known users. However, they exhibit a high generalization error when tested on unknown users; thus cannot be used in real-world settings without significant tuning. Furthermore, while our results are not conclusive, we showed that it could be possible to design stress recognition system that is based on generic stress recognition models and further tune these models by incorporating the physiological fingerprints of new unseen users.