CROCS: Clustering and Retrieval of Cardiac Signals Based on Patient Disease Class, Sex, and Age

The process of manually searching for relevant instances in, and extracting\ninformation from, clinical databases underpin a multitude of clinical tasks.\nSuch tasks include disease diagnosis, clinical trial recruitment, and\ncontinuing medical education. This manual search-and-extract process, however,\nhas been hampered by the growth of large-scale clinical databases and the\nincreased prevalence of unlabelled instances. To address this challenge, we\npropose a supervised contrastive learning framework, CROCS, where\nrepresentations of cardiac signals associated with a set of patient-specific\nattributes (e.g., disease class, sex, age) are attracted to learnable\nembeddings entitled clinical prototypes. We exploit such prototypes for both\nthe clustering and retrieval of unlabelled cardiac signals based on multiple\npatient attributes. We show that CROCS outperforms the state-of-the-art method,\nDTC, when clustering and also retrieves relevant cardiac signals from a large\ndatabase. We also show that clinical prototypes adopt a semantically meaningful\narrangement based on patient attributes and thus confer a high degree of\ninterpretability.\n