249 Clustering in Central Disorders of Hypersomnolence Results EU-NN database The database included 1,078 individuals, of whom 108 were adolescents (between 13 and 18 years of age) and 970 were adults. There were 489 female and 589 male participants. Cataplexy was present in 724 and absent in 354. In line with ICSD-3 criteria, 752 people were diagnosed as having narcolepsy type 1 (646 definite, 51 probable, 33 possible, and 33 unknown diagnostic certainty), 200 as having narcolepsy type 2 (132 definite, 49 probable, 10 possible, and nine unknown certainty), and 126 as having idiopathic hypersomnia (83 definite, 32 probable, six possible, and five unknown certainty). Appendix D provides an overview of the number of inclusions per centre and ethical approval. The clustering algorithm and the researchers were blinded to the current diagnosis and inclusion centre until the analyses were finished. Clustering algorithm and number of clusters The clustering evaluation metrics did not clearly favour any single number of clusters below 15 (Appendix E). This suggests that individuals in the EU-NN database are not organized according to a single number of archetypes. This means that subdivision can still result in distinct clusters but in the presence of individuals closely bordering different clusters. The number of clusters was therefore based on subgrouping of people without cataplexy because datadriven subdivision of these individuals was our main aim. A simple model with a small total number of clusters was preferred. Visual inspection of the clustering steps from 13 to seven clusters revealed that changes occurred only in clusters with people with cataplexy and that people without cataplexy were consistently divided into the same two clusters. Thus, seven was chosen as the final number of clusters. The dendrogram showing the clustering steps from seven to two clusters is included in Appendix E. People without cataplexy were generally grouped as one large cluster when six and five clusters were selected. This large cluster of people without cataplexy was subsequently combined with people with cataplexy at four clusters. This resulted in a steep worsening of clustering evaluation metrics, indicating poorer performance below five clusters. Clustering outcome The means barcodes (Figure 1) show that people with cataplexy were grouped into four clusters (1–4) with 231, 298, 92, and 99 individuals, respectively. Those without cataplexy were grouped into clusters 5 (157 people) and 6 (158 people), 9
RkJQdWJsaXNoZXIy MjY0ODMw