Levente Kander
This thesis attempts to more accurately map the patterns underlying tabular data from patients with aortic valve stenosis by applying a self-supervised contrastive learning technique, thereby facilitating the creation of a more robust patient segmentation. First, we present the theoretical background of contrastive learning, followed by the structure and operation of the TabContrast method. The clusterability of the embeddings generated by the encoder was evaluated both visually and based on the silhouette metric. Based on the results, it was found that, compared to the original raw data, TabContrast provides better clusterability, alongside a structure that reflects general clinical and cardiovascular risk differences. However, there is no significant difference in accuracy between the random forest models trained on the embedded data and those trained on the original vector space when classifying calcium scores measured on the aortic valve.