An ensemble mixed effects model of sleep loss and performance.

ABSTRACT

Sleep loss causes decrements in cognitive performance, which increases risks to those in safety-sensitive fields, including medicine and aviation. Mathematical models can be formulated to predict performance decrement in response to sleep loss, with the goal of identifying when an individual may be at highest risk for an accident. This work produces an Ensemble Mixed Effects Model that combines a traditional Linear Mixed Effects (LME) model with a semi-parametric, nonlinear model called Mixed Effects Random Forest (MERF). Using this model, we predict performance on the Psychomotor Vigilance Task (PVT), a test of sustained attention, using biologically motivated features extracted from a dataset containing demographic, sleep, and cognitive test data from 44 healthy participants studied during inpatient sleep loss laboratory experiments. Our Ensemble Mixed Effects Model accurately predicts an individual's trend in PVT performance, and fits the data better than prior published models. The ensemble successfully combines MERF's high rate of peak identification with LME's conservative predictions. We investigate two questions relevant to this model's potential use in operational settings the tradeoff between additional model features versus ease of collecting these features in real-world settings, and how recent a cognitive task must have been administered to produce strong predictions. This work addresses limitations of previous approaches by developing a predictive model that accounts for interindividual differences and utilizes a nonlinear, semi-parametric method called MERF. We methodologically address the modeling decisions required for this prediction problem, including the choice of cross-validation method. This work is novel in its use of data from a highly-controlled inpatient study protocol that uncouples the influence of the sleep-wake cycle from the endogenous circadian rhythm on the cognitive task being modeled. This uncoupling provides a clearer picture of the model's real-world predictive ability for situations in which people work at different circadian times (e.g., night- or shift-work).