Training stress models on open-access data for a continuous human state monitoring platform

Abstract: Stress can be an indicator of discomfort with a task, which is of relevance for training in safety-critical fields. Knowing a trainee’s stress level could be especially useful when objective performance outcomes are unclear or when success in training tasks alone is insufficient to predict proficiency in real-life safety-critical scenarios. In this study, stress classification models trained on open-access physiological data and integrated in Sensor Hub, a multi-sensor system for near real-time monitoring, were developed. To obtain ground-truth neurophysiological data recorded under high-stress conditions, raw electrocardiogram (ECG) and respiration data in an open-access database sourced from PhysioNet, consisting of 57 participants with arachnophobia watching spider videos, was used. Machine learning algorithms were trained on features extracted from these raw signals. A first set of algorithms focused on heart rate, respiratory rate, and heart rate variability (HRV) features. The second set included feature normalization according to an individual’s baseline. Models based on individually normalized features reached balanced prediction accuracy >80%. A pilot data collection was conducted with a different sensing device than the device used to obtain these measures. Qualitative analysis revealed that real-time R-R intervals from the new sensors were sensitive to artifacts, suggesting that the model relying on HRV features may not be reliable. The model that used only the baseline normalized heart and respiratory rate was selected as the final choice, exported in the Open Neural Network Exchange format and integrated into the Sensor Hub platform, providing predictions every second. This research demonstrates the potential of open-access data for providing a solid starting point for training cognitive models, while also highlighting the necessity of real-time testing to confirm that models can generalize across different sensors and processing pipelines.

Keywords: Human State Monitoring, Electrocardiography, Respiration, Stress, Machine Learning

DOI: 10.54941/ahfe1004736

Cite this paper:

​