Academic Awards 2025 booklet

89 Uncovering sleep structure through discrete representation learning The clinical standard for sleep diagnostics recommends monitoring someone's sleep by measuring a variety of physiological signals, including brain activity, eye movements and muscle tone. This measurement is called a polysomnography, and it results in large amounts of data that are challenging to interpret. The clinical standard to interpret these data categorizes each 30-second window into one out of five sleep stages, creating an extremely compressed data representation that may, therefore, omit clinically- relevant information. Richer and less compressed representations might reveal new patterns in polysomnography data. In this thesis we investigated such new representations, which we acquired through machine learning. We investigated three directions. First, we extended standard sleep staging with additional information, for example, a distribution across sleep stages that resembles disagreement in human scorers. Second, we developed a model to cluster hidden patterns in polysomnography data in a two-dimensional space that facilitates pattern recognition. This revealed more intricate sleep states compared to the five classical sleep stages, and in sleep walkers it showed differences in the brain measurements between waking up with or without a typical sleep walking behavior. Third, we developed a model that indicates the location of task-related information in a measurement. This teaches us about hidden patterns in data that neural networks use. Ultimately, this thesis can contribute to a paradigm shift in sleep medicine, where polysomnography data might be interpreted beyond standard sleep staging. This paves the way to a better understanding of both healthy sleep and sleep disorders. Figure 1: Visual overview of SOM-CPC, an unsupervised model that we developed for pattern recognition in multi-channel data. It reduces the data to a structured 2D grid of learned clusters/centroids. Annotations are not used during training, but only to interpret the learned clusters.