Identification of Phenotypes in Cardiac Arrest Patient Cohorts

Sammanfattning: In this thesis, it is analysed if cardiac arrest patients can be grouped into similar clusters based on different underlying conditions and clinical variables and if there is a difference between clusters in either mortality or neurological outcome as measured by the Cerebral Performance Categories (CPC) scale. The two data sets both contain a targeted temperature management variable which indicates whether or not patients are cooled down upon arrival as well as a variety of continuous and categorical variables. Thus, the clustering methods need to be able to handle mixed data. The four methods that are presented in this thesis are Latent Class Analysis, KAMILA, which stands for KAy-means for MIxed LArge data, $k$-prototypes, and Partitioning Around Medoids with Gower's distance. These methods are then applied to the two data sets of cardiac arrest patients in order to find underlying phenotypes. For both data sets, when only using the cluster assignment and targeted temperature management variables to predict the binary CPC score, the KAMILA algorithm leads to the best results. Furthermore, there is also a significant difference in CPC score and mortality across the obtained clusters. The evidence suggests that it is not only possible to cluster cardiac arrest patients into different groups based on variables obtained upon admission and the patients' medical history but also that the cooling might be more useful to some clusters than others.