Deep Learning of Sleep Quality based on Ballistocardiographic sensors, Stigmergic Perceptrons and LSTM Networks

Culcasi F. P.

Deep Learning  Sleep Quality  Ballistocardiography  Stigmergic Perceptron  LSTM network 

The negative effects due to inadequate sleep in human beings of any age are well known. Neuroscientists and sleep experts work every day to understand how and what makes sleep more effective in its physical and mental recovery action. Among the most adopted techniques, the standard for the measurement of vital parameters on sleeping subjects is certainly Polysomnography. It is a method that involves many sensors and a laboratory environment, factors that introduce inconveniences that could negatively influence the sleep of the individual himself. In this thesis we will base the experimentation on signals obtained through Ballistocardiography, a portable and less intrusive technique that deduces the heartbeat and respiratory acts based on the accelerations of the body lying on the bed due to forces imparted by the heart to the mass of blood that is pumped to the peripheral body systems. In order to solve the problem of the parametric complexity of explicit analytical models of sleep quality in terms of "sleep architecture", in this work we propose a Deep Learning architecture based on multiple levels of Stigmergic Perceptrons. This approach describes a soft classification technique on time series with respect to a collection of archetypes, each representing a different behavioral class. Finally, the Deep Learning architecture is integrated with recurrent Long Short-Term Memory networks, known in literature for the classification of time series, above all for their ability to solve the problem of long-term dependencies over time. This technology is appropriate in order to establish a classifier able to recognize the level of quality of sleep declared by the subjects of the measurements and confirmed by a result obtained through the application of quantitative heuristics.