Laying the Foundation for Correlating Daytime Behaviour with Sleep Architecture Using Wearable Sensors.

The paper presents results from the SmartSleep project which aims at developing a smartphone app that gives users individual advice on how to change their behaviour to improve their sleep. The advice is generated by identifying correlations between behaviour during the day and sleep architecture. To this end, the project addressed two sub-tasks: detecting a user's daytime behaviour and recognising sleep stages in an everyday setting. In the case of daytime activity detection the best results were achieved using an accelerometer at the wrist and another one at the ankle (87%). A subsequent smoothing step increased the accuracy to over 90%. For recognising sleep architecture we experimented with various consumer wearables that we used in addition to the usual PSG sensors in a sleep lab. Several sleep stage classifiers were learned from the resulting sensor data streams segmented into labelled sleep stages of 30 s each. Apart from handcrafted features we experimented with unsupervised feature learning based on the deep learning paradigm. Our best results for correctly classified sleep stages are between 86 and 90% for Wake, REM, N2 and N3, while the best recognition rate for N1 is 37%. Finally, we discuss a preliminary design of the algorithm for determining correlations between daytime behaviour and sleep architecture.