A framework for sleep staging based on unobtrusive measurements
Summary
For diagnosis and research purposes, a person's sleep should be objectively quantified. Traditionally,
sleep staging is performed by human experts via visual inspection of polysomnographic (PSG) data
measured during the sleep period. There are several drawbacks to this golden standard: the data are
acquired in a rather obtrusive manner and not in the natural sleep environment of the sleeper, hereby
affecting the sleep. Furthermore, this manual scoring is time consuming and prone to human errors.
Automatic sleep staging models exist but typically perform binary classifications, e.g. sleep/wake,
REM/NREM or light/deep sleep classification and on the whole use obtrusively obtained measurements.
In this research we aim to develop a framework for classifying sleep stages from unobtrusive
measurements, being measurements during which the participant is not disturbed and no sensors are
attached. Our current implementation is presented that constructs 3- and 4-step hypnograms (containing
wake, REM and NREM in the 3-step hypnogram and wake, REM, light sleep and deep sleep in the 4-step
hypnogram) utilizing body movement and respiration rate biosignals obtained during sleep.
This current implementation of the framework consists of three modules that each use a unique way of
interpreting one or more of these biosignals and a fusion process. The modules partly employ adapted
versions of existing models for sleep stage classification and partly newly developed models, grounded in
sleep physiology. The modules classify epochs of a person's sleep based on one or more biosignals and
relevant subject parameters such as age and gender. The framework ensures that each epoch is classified
by at least two different modules. A specialized fusion process analyses the output from the different
modules, merges the output into a final hypnogram and calculates a confidence level for each epoch
classification.
To compare the sleep stages as derived from the current implementation of the framework with the
golden standard, a secondary analysis on an existing data set of 23 participants is performed. The results
are promising and show potential: the three modules perform binary classifications with 90%, 85% and
72% agreement to the golden standard, the fusion process corrects 61% of the classifications on which
the modules disagree and fuses the binary classifications in definitive 3- and 4-step hypnograms. During
this process the overall agreement drops significantly because all erroneous classifications of the
modules are added up, but due to the correction of erroneous classifications by the fusion process these
hypnograms maintain 69% and 51% agreement with the golden standard, respectively.
This degree is expected to increase significantly when the existing modules are improved, additional
modules are added to allow the fusion process to correct more erroneous classifications and the
parameterization of the modules is done simultaneously.
Also, the framework contains various technical advancements not found in other contemporary sleep
staging methods. Exemplary here is the great flexibility of the framework, enabling the researcher or
clinician to easily add, remove or alter modules. Also complete parameterization, ensuring that at all
times the clinically relevant parameters of the participant are taken into account, can be mentioned here.
The current implementation of the framework is expected to be a basis for unobtrusive sleep staging that
can be performed in the natural sleep environment of the sleeper and without influencing or disturbing
sleep, hereby potentially lowering the threshold for (preliminary) sleep diagnostics, research and therapy.