Abstract Actigraphy is a non-invasive and inexpensive method to monitor sleep/wake patterns in a natural environment via a wrist-worn activity sensor. Traditionally, detection of sleep/wake periods from actigraphic data relies on smoothing and thresholding the time series of recorded “activity counts”. The first step is implemented by convolution with empirically chosen coefficients, tailored separately for the data and hardware used in each study. We propose to implement this step via a universal low-pass filter, applicable to wide ranges of recording hardware and sampling rates. For verification of this approach, we used 1635 overnight coregistrations of actigraphic and polysomnographic (PSG) data from three different datasets, including one dataset recorded for this study. Optimizations of the filter for concordance of sleep/wake scoring with PSG for different subsets of these data converged to similar parameters, which we tentatively treat as fluctuations around the characteristics of a universal filter. We assess the performance of the proposed approach and five classic algorithms (Cole-Kripke, Sazonov, Scripps, UCSD and Webster) in the same cross-validation scheme. Concordance with PSG, achieved using the universal filter, is significantly higher (at p < 0. 001) than any of the classical algorithms for the most relevant metrics.
Biegański et al. (Fri,) studied this question.