黑料福利网

To diagnose many sleep disorders, clinicians rely on a test known as polysomnography, or PSG. During a PSG, patients spend an entire night connected to multiple sensors that record brain activity, eye movements, muscle tone, breathing patterns, and other physiological signals. Specialists then divide the recording into 30-second segments and visually classify each segment into a specific sleep stage. Although this approach is considered the gold standard, it comes with significant drawbacks. The procedure is costly and labor-intensive, the equipment and sensors are intrusive and uncomfortable, and the manual scoring process can take hours per patient. In addition, even highly trained experts often disagree with one another when assigning sleep stages.

The promise鈥攁nd limitations鈥攐f deep learning

In recent years, deep learning has emerged as a promising tool for automating sleep staging. These systems learn from large datasets of human-scored sleep recordings and can analyze a full night of data in a fraction of the time required by human scorers. Some models can even operate using fewer sensors, raising the possibility of simpler and more accessible sleep studies. However, traditional deep learning approaches still face important challenges. They inherit the variability and ambiguity present in human annotations, their performance often drops when alternative or reduced sensor setups are used, and their ability to scale across different sensor types remains uncertain. Moreover, many models are validated primarily on healthy individuals rather than on real clinical populations.

Rethinking sleep staging through uncertainty

This research proposes a fundamentally different perspective by treating sleep staging as a probabilistic problem rather than a task with a single correct answer. Instead of assigning one definitive sleep stage to every moment of the night, the approach estimates a range of possible sleep stages and how likely each one is, given the measured signals and prior knowledge about how sleep normally evolves. This is achieved using deep generative models, which naturally represent uncertainty. The work distinguishes between aleatoric uncertainty, which reflects the inherent ambiguity in sleep signals that even human experts cannot fully resolve, and epistemic uncertainty, which arises from limitations in data or model knowledge. Explicitly modeling both types of uncertainty leads to more transparent and clinically meaningful results.

Matching and explaining human disagreement

Using data scored independently by six human experts, the research demonstrates that deep generative models provide better-calibrated estimates of uncertainty than conventional discriminative models. When the model is uncertain, that uncertainty closely mirrors the level of disagreement observed among human scorers. This ability to quantify and explain uncertainty is crucial in clinical settings, where knowing when an automated system is unsure can be just as important as knowing when it is confident.

High-quality sleep staging with fewer sensors

One of the most striking findings is that high-quality sleep staging can be achieved using only a single eye-movement sensor, known as an EOG. Even in clinical populations, performance reaches the level of agreement typically observed between human experts. This is possible because the EOG signal naturally contains useful information from brain activity, making it particularly informative for sleep staging. These results suggest that future sleep studies could be far simpler, more comfortable, and potentially conducted outside of traditional sleep laboratories.

Combining sensors in a flexible and scalable way

The research also introduces a new modeling approach called the Factorized Score-based Diffusion Model. This method is designed to combine information from any set of sensors in a flexible manner. It can be trained on new sensor types independently and then immediately applied to new combinations of sensors it has never encountered before. The model achieves state-of-the-art performance on brain and eye signals and shows strong results on cardio-respiratory measurements and even unconventional signals such as single-channel muscle recordings.

A breakthrough for REM sleep behavior disorder

An especially important application of this work concerns REM sleep behavior disorder, a condition in which patients physically act out their dreams because normal muscle paralysis during REM sleep is absent. Diagnosing this disorder typically requires a full polysomnography setup. The research shows that REM sleep can be accurately identified using only a single chin muscle sensor, producing clinical measures that are statistically indistinguishable from those obtained through full PSG and expert human scoring. This approach could dramatically improve access to diagnosis, particularly in home-based or ambulatory settings.

Toward more accessible and trustworthy sleep diagnostics

Taken together, this PhD research of Hans van Gorp demonstrates how deep generative artificial intelligence can overcome many of the longstanding challenges in sleep diagnostics. By explicitly modeling uncertainty, accounting for human disagreement, enabling flexible fusion of sensor data, and validating performance in real clinical populations, these methods point toward a future in which sleep diagnostics are more accessible, scalable, and transparent. Understanding sleep may no longer require a night in a laboratory filled with wires, but could instead be achieved through intelligent systems that combine rigor, flexibility, and clinical trust.

Read more about the research of Sebastiaan Overeem, Merel van Gilst and their team on new ways to monitor sleep disorders. Besides that there was the PhD defense on the combination of unobtrusive wearables with specialized AI to diagnose sleep disorders of Jaap van der Aar in December 2025. 

Title of PhD thesis: Deep generative modelling in sleep diagnostics. Supervisors: Dr. Ruud van Sloun, Dr. Merel van Gilst, and Prof. Sebastiaan Overeem.