Introduction and aims: the gold standard for sleep staging is the in-laboratory polysomnography (PSG), followed by manual scoring. A wide range of limitations of this approach has been reported, ranging from high costs to low compliance. The market of the so-called "quantified self" lists an increasing number of tracking devices, which also offer the opportunity to measure sleep parameters. Still, the validation of these devices is limited.
Methods: in 20 young healthy subjects, we recorded 24 hrs of portable EEG data combined with by a low-cost commercially available accelerometric recordings (Fitbit Ultra). A validation procedure based on a multi-layer perceptron artificial neural network (ANN) has been employed in order to optimize actigraphy-based versus EEG-based vigilance state scoring.
Results: The ANN approach extracted an algorithm leading to high accuracy (0.939+-0.03), sensitivity (0.936+-0.07) and specificity (0.944+-0.03) in the estimation of 5-minute sleep epochs, for the comparison of EEG-based and actigraphy-based scoring. The training phase reached saturation after 4 subjects. The estimation of standard sleep parameters (TST, WASO, Sleep Onset) showed no statistical difference between the automatic ANN-actigraphy-based scoring and the standard EEG-based one.
Conclusion: The high concordance between ANN-actigraphy-based scoring and the standard manual EEG-based one, as well as the estimation of sleep parameters, makes low-cost actigraphy a viable strategy for collecting objective sleep-wake data. Finally, we propose a validation procedure that could be employed for testing future devices as well as existing ones, requiring relative long (24 hrs) simultaneous portable EEG and actigraphic recordings, in a relative small sample (n=4).