黑料福利网

At 黑料福利网鈥檚 Advanced Sleep Monitoring Group, Professor Sebastiaan Overeem and Assistant Professor Merel van Gilst are quietly rewriting the rules of sleep science. Instead of adding ever more sensors to patient鈥檚 heads, they鈥檙e questioning how data should be interpreted to better diagnose patients suffering from one of the more than eighty currently recognized sleep disorders.

鈥淗ospital equipment is impressive,鈥� says Overeem, who is also a clinical somnologist (a clinician specializing in sleeping disorders). 鈥淏ut it still only measures surface activity and places complex brain processes in simple boxes. In this process of data interpretation, you lose a lot of what鈥檚 actually interesting.鈥�

Learning from ambiguous signals

Their solution sounds almost counterintuitive: don鈥檛 hide uncertainty, but instead show it. 鈥淲e want monitoring that鈥檚 less obtrusive and more practical,鈥� Van Gilst explains. 鈥淎nd we also want it to be more truthful.鈥�

Van Gilst is a somnologist as well as a neuroscientist, whose PhD research focused on sleep disorders in patients with Parkinson鈥檚 disease. Currently she is Assistant Professor in the Signal Processing Systems group (department of Electrical Engineering) and leads the Advanced Sleep Monitoring Team together with Sebastiaan Overeem.

鈥淏y admitting what we don鈥檛 know, we open ourselves up to learning more from the signals our sleeping brains produce,鈥� Van Gilst notes. It鈥檚 an approach that forms the basis of the work of several PhD students in her research group. All focus on one or more parts of the measuring technology involved in sleep monitoring, from smarter algorithms to new sensor designs that are less obtrusive yet remain accurate.

The illusion of certainty

For decades, sleep research has relied on a standard approach: recording the brain鈥檚 electrical activity, dividing the night into thirty-second intervals and assigning each interval to a specific sleep stage. The current system distinguishes five stages: Wake, then N1, N2 and N3 (which together comprise non-dreaming sleep, ranging from light to deep) and the generally better-known dream state REM sleep, or Rapid Eye Movement sleep. The resulting graph of electrical activity is known as a hypnogram. It looks neat and decisive. But according to Overeem and Van Gilst, that confidence is an illusion.

鈥淓ven with our most advanced recordings, we鈥檙e not capturing what sleep really is,鈥� Overeem says. 鈥淲e鈥檙e seeing a representation of something deeper, not sleep itself.鈥�

Probing the depths of the brain

The reason for the lack of a robust visualization of sleep is that sleep is a notoriously hard thing to accurately measure. Tracking sleeping patterns relies on decisions and compromises that throws away a lot of 鈥榝uzzy鈥� data and presents what鈥檚 left as certainties about sleep. However, collecting the data behind the actual sleep state is easier said than done.

First, there鈥檚 the challenge of accessing the parts of the brain that regulate sleep. They lie deep within some of the more 鈥檖rimitive鈥� regions of the brain, such as the hypothalamus, and cannot be directly monitored. EEG sensors on the scalp, which measure electrical activity in the brain, pick up only faint, heavily filtered signals that have traveled through the cortex (the outer layers of the brain) and the skull before reaching the sensors.

Another layer of simplification is technicians (or algorithms) compressing hours of complex data from multiple sensors into one label per 30-second window. It鈥檚 a necessary coarse graining of the data, but a misleading one in that it averages out so much brain-activity data.

鈥淚f ten trained technicians score the same night鈥檚 sleep of a patient, they won鈥檛 fully agree on all the different sleep stages that the patient went through,鈥� Overeem says. 鈥淭hat uncertainty is hidden behind the clean lines of the hypnogram, that only allows for one sleep stage per window. That鈥檚 false certainty.鈥�

Hypnodensity as an alternative

To deal with the uncertainty inherent in sleep data, the research team uses the hypnodensity graph. This is a way of analysing, processing and representing sleep data which replaces definitive labels with probability landscapes. Instead of assigning a single sleep-stage every 30 seconds, a hypnodensity graph displays the likelihood of each stage over time.

It was originally developed by an in a study of narcolepsy utilizing machine learning. The group at 黑料福利网 further builds upon that initial research to find out what it really represents and applies it to several new contexts.

A hypnodensity graph can spotlight underlying instability, which might point us in the direction of the cause of symptoms.

Sebastiaan Overeem

鈥淚n traditional graphs there is always a hard choice for one of the five sleep stages, even if the underlying data is not clear or ambiguous,鈥� says Van Gilst when responding to the question why this novel approach might be preferable. 鈥淲hen the picture seems solid but doesn鈥檛 match the patient鈥檚 symptoms, it鈥檚 possible that crucial information got lost in translation from the raw signals to the hypnogram. Showing uncertainty helps us connect the data to real symptoms, because it paints a more nuanced picture.鈥�

鈥淢achine-learning models already think in probabilities,鈥� Overeem adds. 鈥淪o, why not show that instead of hiding it? Maybe it鈥檚 80 percent N1 sleep, 10 percent N3 sleep, and so on.鈥�

To REM or not to REM

This approach reveals observations that might have stayed otherwise hidden. Two people may both be labeled as being in REM sleep, but one might show 90 percent certainty (meaning that person is most probably dreaming), while the other鈥檚 signal fluctuates between stages, making it much harder to reach definitive conclusions.

鈥淭raditional hypnograms might look normal even when a patient feels unrested,鈥� says Overeem. 鈥淎 hypnodensity graph can spotlight underlying instability. This ambiguity might point in the direction of the cause of the symptoms. For us, it鈥檚 exactly the uncertainty about sleeping patterns that might hold useful information in the long run.鈥�

Article continues below the highlighted box.

Collaboration with Kempenhaeghe

The Advanced Sleep Monitoring Team is in close collaboration with Kempenhaeghe, an Eindhoven-based expertise center for sleep medicine, where Overeem regularly sees patients in his role as clinician. Kempenhaeghe provides one of their biggest assets: the Somnia dataset, a rich collection of complex sleep recordings used to train and validate new algorithms.

Access to this dataset revealed that sleep patterns are far more complex than they appear -especially in disordered sleep- highlighting the importance of a more nuanced method. Van Gilst: 鈥淲e are not only interested in building high performing classifiers, but also in understanding why they sometimes do not work.鈥�

Four PhD鈥檚, four paths toward smarter monitoring

In addition to studying the usefulness of the hypnodensity graph, an important goal for the Advanced Sleep Monitoring Group is to get more information out of less invasive monitoring. Four PhD students involved with the group may provide even more ways to refine measurements, train smarter algorithms and develop sensors that are both more informative and less intrusive.

Jaap van der Aar and Hans van Gorp focus on developing new ways of interpreting sleep data, extracting more information from signals while explicitly accounting for uncertainty. Fons Schipper and Luca Cerina, on the other hand, work on new measurement methods and sensor designs, aiming to reach the same clinical conclusions using smarter, more efficient tools even though these gather less data than in a hospital setting.