Statistical methods for sleep stage detection using raw accelerometry data

We develop methods to detect sleep stages using raw accelerometry. The methods are assessed by comparison to the gold standard of polysomnography. As the latter is intrusive and costly, it is of great interest to investigate easier and cheaper alternatives based on accelerometry. Our best models detect sleep better than existing algorithms, and detection extends to many individual sleep stages, something not previously achieved without some type of additional data. Our algorithms may be more transferrable to other accelerometer devices than existing algorithms as the raw data does not undergo any proprietary filtering.