Harnessing the Power of Multimodal Physiological Signals and Deep Learning for Stress Monitoring

Physiological and behavioral information obtained from wearable or mobile sensors has been utilized to estimate self-reported stress levels. However, the reliance on self-reports for stress annotation during studies often leads to a scarcity of labeled data, impeding the development of accurate and broadly applicable stress prediction models. In contrast, sensors can continuously capture signals without requiring annotations. Researchers have endeavored to identify stress using conventional machine learning techniques for analyzing physiological signals. The achieved results, which vary between 50% and 90% accuracy, have produced mixed outcomes. A drawback of traditional machine learning algorithms is their dependency on manually crafted features. Accuracy tends to decline if these features are misidentified. To overcome this limitation, we introduced a deep neural network approach. Deep neural networks eliminate the need for hand-crafted features and instead extract pertinent features directly from raw data through the network's layers. In our study, deep neural networks were employed to analyze physiological data obtained from the DRYAD dataset. The objective encompassed multi class classification for stress detection, where the networks differentiated among low-stressed, medium-stressed, and high-stressed states, effectively distinguishing baseline, stressed, and amused states.