Stress can be debilitating, affecting the mind and the body. The World Health Organization defines stress as a state of worry or mental tension caused by a difficult situation.

While stress affects everyone, construction workers are disproportionately impacted in society. Their work is physically demanding, sometimes hazardous, and frequently involves unrealistic deadlines. The resulting stress can lead to workplace accidents, diminished productivity, and even depression.

Sadly, the Centers for Disease Control and Prevention ranked the construction industry as the number one industry for deaths by suicide. Several manual surveys have been used by the industry to assess general stress levels in the field, but these can be cumbersome, lead to recall bias, and cannot show the entire picture of a construction worker’s daily experience.

Researchers Gaang Lee and SangHyun Lee wanted a means of detecting workers’ stress in a continuous, bias-free, and minimally invasive way. Monitoring subjects using wearable biosensors to continuously collect biosignals, such as cardiovascular function and sweat production, would be ideal; but research indicates that results can often be subjective and context dependent, which limits scalability. In this study, “Deep-Learning Domain Adaptation to Improve Generalizability across Subjects and Contexts in Detecting Construction Workers’ Stress from Biosignals,” the authors propose and test a stress detection technique that buffers individual and contextual variabilities in stress-related biosignal patterns. They employ transfer learning, a subset of machine learning, to train models by leveraging previously collected information from similar tasks.

Learn more about their domain adaptation–based stress detection technique in the Journal of Computing in Civil Engineering at The abstract is below.


Wearable biosensors, in conjunction with machine learning, have been employed to develop less invasive monitoring techniques for assessing stress among construction workers during fieldwork. However, existing techniques face limitations in terms of scalable field application due to their subject and context dependency; it is difficult to apply them to new people in new contexts without additional labeled data collection. Therefore, this study developed a stress detection technique that incorporates domain adaptation, simultaneously learning a classifier and a subject- and context-independent features, in this way advancing generalizability. The proposed technique consistently demonstrated superior accuracy compared with benchmarks in classifying stress levels within a testing data set whose subjects and contexts were different from those of training data sets. Thus, the technique can advance generalizability across subjects and contexts. This finding can help us to reliably detect stress for new people in new contexts without additional labeled data collection, thereby contributing to scalable field application of wearable-based stress monitoring at construction sites.

Learn more: