Reproducible and accurate subject-wise sleep posture detection by detecting and removing turns

Maintaining a good sleep hygiene is an important factor to avoid the symptoms of sleep disorders or worsen the symptoms of other diseases. Polysomnography is the study of sleep performed by professionals during a night at the hospital. On these studies they perform the diagnosis of diseases and patients are not monitored any more. A non-intrusive and low-cost ambulatory monitoring would allow a follow-up of the diagnosed patient. Such studies use numerous and uncomfortable sensors that disturb the patients’ rest. One of the sensors on the chest monitors 4 torso postures: prone, supine, left lateral and right lateral. In this work we analyze the reliability of performing posture monitoring during sleep with a wearable device on the wrist. In our methodology we develop classification models to prove that in order to make these models applicable on real data it is necessary to (i) perform a subject-wise training and (ii) detect and eliminate the monitoring periods corresponding to turns of torso or sudden movements. Our methodology improves the state-of-the-art results by more than 0.011 points with F-values on new subjects of 0.966 and 0.989 for Random Forest and k-Nearest Neighbors algorithms respectively.