The diagnosis of complex conditions remains challenging when biomarkers are lacking and diagnostic criteria rely on subjective clinical judgment. We propose a novel contrastive clustering framework for phenotype discovery, combining instance- and class-level learning with softpriors to guide representation learning. Paired with consensus clustering, our method guides the identification of subgroups in heterogeneous populations. We apply this approach to a dataset of electroencephalography and physical activity data from patients with Central Disorders of Hypersomnolence, a clinically ambiguous spectrum that lacks biomarkers and exhibits overlapping symptoms. To validate generalizability, we also test the framework on an open-source dermatological image dataset characterized by distinctly defined diagnostic categories. Our results highlight the potential of our methodology for data-driven discoveries across a range of clinical contexts, whilst incorporating expert clinical knowledge.
