Skip to main content

Classifying Sleep Episodes, Recovery and Stress in Extreme Work Environments

Team: Michael David Wilson, Alex Boeing, Belinda Cham, Karina Jorritsma, and Mark Griffin.

CIC specialists: Dan Marrable

For teams working in safety-critical work environments (e.g., military, health care, and space operations), it is imperative that physiological and psychological measurements do not disrupt operators during field operations. In industrial settings, there is a growing interest in using wearable technologies to measure physiological states and psychological states. A collaborative project between the Future of Work Institute (FOWI) and Curtin Institute of Computation (CIC) has investigated how heart rate sensors could be used for sleep detection in the context of operational naval teams. While sleep detection isn’t a new problem, existing ambulatory methods (e.g., motion sensors) can misclassify passive activities as sleep (e.g., watching TV). Furthermore, unlike motion sensors, heart rate sensors offers the promise of also providing information about stress and recovery (not just sleep status). 

In a longitudinal field study, the data from ECG sensors were linked with daily diary observations made by participants in order to perform ‘activity modelling’ — the process of classifying activities using passive sensors. “If we are able to reliably detect when sleep is occurring with non-invasive sensors — we can then send this data to other bio-mathematical models in order to predict mental and physical fatigue, and make decisions around rostering. Long term, we would be hoping to do this on the fly” Michael Wilson (FOWI) explained.

In this project, a series of exploratory analyses were conducted in order to identify the feasibility of using machine learning to classify activities as sleep or wake using heart rate data alone. Early results show that using a combination of heart rate and heart rate variability it was possible classify sleep with an average accuracy of 86.9%, with the lowest accuracy being 79.9% and the highest being 89.3% when the model was tuned for individual participants.

An example of the sleep prediction model compared to the sleep diary of one of the study participants.

In future research, the team is hoping to also examine how to characterise stress, recovery, and sleep quality from the heart rate sensors. If possible, this would open a new avenue for reliable non-invasive estimates of psychological endurance, recovery and mental fatigue over operations without the need for invasive assessments. Importantly, monitoring of stress could be performed continually through work shifts. While the work here is preliminary, an end goal would be to examine how the information from these predictions could be embedded with system platforms to inform decision-making processes and improve the endurance of the workforce. 

An example of very early work using machine learning to predict stress and recovery patterns from heart rate data when compared to the proprietary software predicted heart rate.