The diagnosis of complex conditions remains challenging whenbiomarkers are lacking and diagnostic criteria rely on subjective clinicaljudgment. We propose a novel contrastive clustering framework for phenotypediscovery, combining instance- and class-level learning with softpriorsto 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 electroencephalographyand physical activity data from patients with Central Disorders of Hypersomnolence,a clinically ambiguous spectrum that lacks biomarkers andexhibits overlapping symptoms. To validate generalizability, we also testthe framework on an open-source dermatological image dataset characterizedby distinctly defined diagnostic categories. Our results highlightthe potential of our methodology for data-driven discoveries across arange of clinical contexts, whilst incorporating expert clinical knowledge.