Contact-free actigraphy: Measuring sleep with a radar sensor

Abstract

Background Sleep is an essential biological process, and sleep deprivation and sleep-related disorders have devastating personal and societal consequences. The diagnosis and effective treatment of sleeprelated problems can be greatly aided by measuring the physiological phenomenon of sleep objectively, however, the current range of options for objective sleep monitoring in clinical and research settings is limited. Novel technologies can offer exciting new opportunities for expanding this range. Aims The aim of this thesis was to develop and evaluate the potential of a specific radar sensor as a clinically and scientifically useful tool for objective contact-free sleep measurement. Methods The work of this thesis began with Paper I, a simple proof-of-concept study, wherein 12 persons (age 19-60) underwent overnight monitoring with the radar sensor and polysomnography (PSG) recording simultaneously. Two participants contributed twice, giving a total of 14 nights of parallel data. A simple threshold-based model for classification of binary sleep/wake state was then applied to this radar data, and its performance evaluated in comparison to PSG-scored sleep/wake state. Two sets of data were then gathered. For Dataset 1 (DS1), 12 healthy participants (age 20-30) resided in a real state-of-the-art acute psychiatric hospital ward for ten days, before the unit had been opened for admission. Their bedrooms were continuously monitored with two radar sensors (on a nightstand and in the ceiling) for the duration of the study. Participant also wore wrist actigraphs for the duration, and each participant underwent a total of four nights of PSG. For Dataset 2 (DS2), a broad cross-sectional sample of 28 adults (age 20-71) referred to a sleep clinic were recruited to include simultaneous radar and wrist actigraphy monitoring together with their regularly ordered sleep exam. In Paper II, linear sliding sum models for sleep/wake classification inspired by models commonly used in wrist actigraphy were examined for use with contact-free radar data. To evaluate the potential for real-time monitoring, models wherein all reliance on future information was removed were also studied. Models were calibrated for both the radars, and for epoch-binned activity data exported from the wrist actigraphs. DS2 was used as an independent validation set, and to evaluate the performance of the models over a more heterogenous population. Paper III explored a tool known as Locomotor Inactivity During Sleep (LIDS), designed to emphasize and enhance the contrast between activity and inactivity as registered by a wrist actigraph during the night and by so doing reveal underlying ultradian patterns in the low-level residual activity. In Paper III, the LIDS method was adapted for use with the radar, and applied to both radar and wrist actigraphy data from DS1. Results The numerical results from the preliminary proof-of-concept study were fair, albeit with large interindividual variances in the classification results that were unsurprising given the small yet heterogenous dataset. However, the lack of an independent test or validation set in this study means that the results have limited translational value. The wrist actigraphy-inspired sleep/wake classification models applied to radar data achieved results on-par with, and sometimes exceeding, those often seen for wrist actigraphy. Real-time models performed slightly worse than non-real-time models, but in a situation where live monitoring were desired this trade-off would be small. Wrist actigraphy performed better than the radar over the heterogenous troubled sleepers of DS2, but both performed within the range of results reported from previous studies comparing wrist actigraphy to PSG over comparable populations. Radar-derived LIDS correlated strongly with wrist actigraphy-derived LIDS, and both were also highly correlated with reduced-resolution polysomnographic hypnograms. Similar slopes of decline per cycle were found for radar-LIDS, wrist actigraphy-LIDS, and hypnograms when aggregated over the dataset. Conclusions The work in the present thesis demonstrated that the radar is a promising tool for sleep assessment. The results have shown that data from the radar can be used with models taken from wrist actigraphy to provide valid estimates of sleep, wakefulness, and related parameters. Such estimates were also valid when all reliance on future information was removed. Furthermore, a promising novel tool was adapted to contact-free radar measurements of body movement. This LIDS technique could provide a simple and transparent way to study the ultradian dynamics of sleep in settings where such study has previously been difficult.