[home] [Personal Program] [Help]

---

Session: Neurophysiology & Sleep
Session starts: Thursday 26 January, 10:30
Presentation starts: 11:30
Room: Room 530

---

Iris Huijben (Eindhoven University of Technology)
Ruud van Sloun (Eindhoven University of Technology)
Sebastiaan Overeem (Sleep Medicine Center Kempenhaeghe)
Merel van Gilst (Eindhoven University of Technology)

Abstract:
Introduction The expressiveness of a hypnogram in sleep medicine is limited due to the assignment of one AASM sleep stage label per 30-sec sleep epoch. We explored a new pattern-recognition method to further study sleep structure, possibly yielding new insights in disordered sleep. Cluster analysis is a common approach for unsupervised pattern recognition, but classical methods are typically applied on handcrafted features (e.g. spectral power bands), selection of which relies on expert knowledge and often limits an approach to specific applications. We propose SOM-CPC, a method that learns features using Contrastive Predictive Coding (CPC) [1], and subsequently clusters, as well as visualizes, them using a Self-Organizing Map (SOM) [2]. SOM-CPC is sensor-agnostic and takes temporal information into account. Methods Video-polysomnography recordings of 96 healthy subjects were studied (60 F, age: 33±13.6 years) from which we sub-selected F3/F4, C3/C4, O1/O2, Chin1/Chin2 and E1/E2 derivations. We created a hold-out test set of the even channels from n=11 recordings on which conclusions were drawn. The rest of the data were used for training and validating the model. SOM-CPC resulted in a 2-dimensional grid of 100 clusters. For interpretability, each cluster was labelled with the most-frequent sleep stage label and the distribution over time-in-night. Cluster statistics were compared using the non-parametric Mann Whitney U test. Results Labelling each cluster with the most-frequent AASM label of the training set, yielded a Cohen’s kappa of 0.6±0.16 with respect to the expert annotations of the test set. Assigning a distribution over AASM labels to each cluster, revealed non-transitional clusters and transitional clusters that received varying labels. Interestingly, these transitional clusters were positioned at the boundaries of non-transitional clusters on the grid. Adding time-in-night labelling, we found a cluster of early-night (n=147, median epoch: 19) and late-night Wake epochs (n=262, median epoch: 498; U=6.4e3, p=2.15e-29). Conclusions We propose a new approach to cluster raw sleep recordings. Visualizing the grid of clusters labelled with different variables per cluster, allows for recognition of patterns beyond those that can be deducted from the hypnogram. Next, training on data from patients with different sleep disorders may cluster certain patients with specific demographics. [1] Aaron van den Oord, Yazhe Li, and Vinyals Oriol. Representation Learning with Contrastive Predictive Coding. arXiv preprint arXiv:1807.03748, 2019. [2] Teuvo Kohonen. The self-organizing map. Proceedings of the IEEE, 78(9):1464–1480, 1990.