Multi-sensor sleep modeling based on contextual data fusion

Humans spend about one third of their lives sleeping. A good way to understand the role of sleep is to look at what would happen if we did not sleep. Lack of sleep has serious effects on the functionality of the brain. With continued sleep deprivation, the brain regions that control language, memory, planning and sense of time are severely affected. In fact, 17 hours of sustained wakefulness leads to a decrease in performance equivalent to a blood alcohol level of 0.5 (two glasses of wine). Research also shows that sleep-deprived individuals often have difficulty in responding to rapidly changing situations and making rational judgments. In real life situations, the consequences are grave and lack of sleep is said to have been be a contributory factor to a number of international disasters such as Exxon Valdez, Chernobyl, Three Mile Island and the Challenger shuttle explosion. Sleep deprivation not only has a major impact on cognitive functioning but also on emotional and physical health. Disorders such as sleep apnea, which results in excessive daytime sleepiness, have been linked to stress and high blood pressure. Research has also suggested that sleep loss may increase the risk of obesity because chemicals and hormones that play a key role in controlling appetite and weight gain are released during sleep. To better understand the structure, regulation mechanisms and affects of sleep on human behavior, the sleep process modeling approach became increasingly important in sleep science. The currently widely accepted model of sleep is based on the assumption that several distinct stages of sleep are involved in the cycle between sleep and wakefulness. Almost 40 years ago, the guidelines of Rechtschaffen and Kales (R&K) were introduced as a reference method to unify the definition of individual sleep stages, based on EEG and other biosignals. Despite its drawbacks, the R&K manual became a worldwide accepted standard for sleep stages classification and contributed significantly to a better understanding of the sleep process and its disturbances in the past. However, the thorough validation conducted over the last years revealed many significant limits of the R&K rules. Because of these scientific observations, new models of sleep were proposed. Especially models based on solid probabilistic principles have shown promising results. Nevertheless, until now, a systematic comparison and validation of these alternative sleep models has not been accomplished. Moreover, there exist several fundamental theoretical and conceptual limitations associated with these models. This might be the reason that a wider community of sleep researchers did not accept the models so far. The aim of this project is the systematic extension of the current probabilistic sleep models, with biosignals like EEG as input, by addressing previously observed limitations. This will be accomplished by using novel advanced model architectures, learning algorithms and data representation together with extensive previous experimental and theoretical expertise. Availability of a wide collection of sleep recordings collected during the previous projects provides a unique opportunity to methodically support the correct validation process inevitably associated with the planned extension of the current sleep models. The project involves the novel Dynamic Bayesian Networks based approach for systematic probabilistic multi-sensor and contextual data fusion, which has not been considered in sleep science so far. The project aims to study and model effects of circadian, homeostatic and sleep environment factors in a principled probabilistic way as a prior influencing the final profiles of the sleep process. This scientifically challenging novelty of the project provides the opportunity to not only better understand the sleep process itself but also to provide new insights into understanding the sleep regulating and influencing factors. In summary, it is expected that the results of the project will represent another important scientific step towards a novel widely acceptable approach of sleep modeling - a long time expected achievement in the sleep research community.

For decades, the diagnosis of sleep disorders in sleep medicine world-wide has relied on the concept of a `sleep architecture` derived from physiological measurements like electroencephalography (EEG) by dividing them into 4 or 5 discrete sleep stages for each 30 seconds of recording. For the past 20 years more and more researchers have argued that the EEG (and other signals) contains more information about sleep than those crude sleep stages, demanding for alternative models of sleep architecture. This project has successfully developed such a model and, for the first time in the light of previous attempts, has unambiguously proven that such a model can indeed extract more information in terms of higher correlation between variables characterizing the new type of sleep architecture and outside morning and evening measures about sleep quality, including questionnaires about a person`s well- being and psychometric tests. The model is based on so-called Gaussian mixture models that cluster spectral characteristics of the signals into a number of sleep states with a high temporal resolution and without a prior definition of how many and which states are reached during a night of sleep. Thus, the model frees itself from some of the limits of classical sleep signal analysis, namely that stages are defined by what an expert can identify visually in the signal and by the arbitrary rough division into 30 second pieces, historically still rooted in the use of paper EEG. The proof that this new way of describing sleep correlates better with how a patient feels and performs in the morning points to the clinical validity of the approach which could lead to new ways of analyzing sleep for diagnosis in medicine and beyond. At the same time in this project it could also be shown that high-resolution sleep profiles show decisive individual differences, probably more than could be expected from known differences in EEG between patients. In a sense, a sleep profile, or a summary of one`s EEG spectrum during sleep, almost proves to be a certain individual `fingerprint` that allows matching corresponding nights of the same person in a surprising large number of cases. This points to important limits of using sleep profiles for diagnosing sleep disorders from a single night measurement. Further results from research performed, such as a feasible factor model to condense a large number of sleep quality variables into a few factors, as well as a model for dealing with spatially distributed EEG information, complement the results from this project.