Predicting vigilance in a shift-work context with under-mattress sensor data from the preceding sleep

Introduction
Performance impairments are risky in many industries, including healthcare and defence. Predicting when impairments are more likely to occur is critical to reduce costly workplace accidents and errors. This study created models for predicting vigilance during simulated night shift-work from under-mattress sleep sensor data.

Methods
The parent study compared two conditions for phase delaying individuals to adjust to night shift-work. Participants (N=11) attended the laboratory for eight days per condition, one month apart. After a baseline sleep (10PM-7AM), participants remained awake for 27hrs, then slept from 10AM-7PM for the next four days with sleep metrics recorded by an under-mattress sensor (the Withings Sleep Analyzer). At night, participants completed simulated night shifts, including six psychomotor vigilance tasks (PVTs) per shift (48 resulting datapoints per participant). The current study predicted PVT performance from the preceding daytime sleep based on 27 sleep variables using machine learning (Extra trees) models. Data were randomly split into a 67% subset for model training and variable reduction, and a 33% subset for testing model fit.

Results
12 variables were retained following feature reduction. The final models demonstrated good fit for reaction time <500ms (R² = 0.79, RMSE = 20.4ms), reciprocal reaction time (R² = 0.70), and number of lapses (R² = 0.69).

Discussion
These preliminary findings are comparable to current fatigue prediction models, supporting that vigilance can be predicted from unobtrusively collected sleep data. Further research will confirm whether these models may assist in the safer delegation of work tasks and self-management of shift-workers.