Objective:Modeling normal tissue complication probability (NTCP) from dose–volume histogram (DVH) data, a 3D planning graph used in radiation therapy to summarize the radiation dose distribution within specific anatomical structures, remains challenging due to the high dimensionality of DVHs and the complex, multi-organ interactions underlying radiation-induced toxicities. Conventional approaches often rely on simplified DVH summaries or single-organ features, limiting their ability to capture full dose–volume relationships and inter-organ effects. This study proposes a self-supervised framework to learn discriminative representations from multi-organ DVH data for improved toxicity prediction. Methods:We developed a self-supervised contrastive clustering framework that learns compact, discriminative representations directly from full multi-organ DVH data. Structure-specific convolutional encoders independently process each organ’s DVH, and the resulting embeddings are jointly refined through a deep embedded clustering objective that promotes well-separated latent representations. The framework was applied to a cohort of head and neck cancer (HNC) patients treated at The University of Texas MD Anderson Cancer Center to predict patient-reported xerostomia severity at six months post-radiotherapy. Results:The framework identified three distinct dosimetric subgroups with a silhouette score of 0.60. Gradient-based attribution analysis revealed that the parotid glands and hard palate, particularly within low-to-intermediate dose ranges, were the primary contributors to cluster differentiation, consistent with established dose–response relationships for xerostomia. Incorporating cluster assignments into a logistic regression model improved prediction performance with an AUC of 0.74 compared to 0.66 using clinical variables alone, with cluster membership emerging as the strongest predictor.Conclusion:Self-supervised dosimetric representation learning can uncover clinically meaningful subgroups and enhance toxicity prediction. The proposed framework provides a principled and scalable alternative to handcrafted feature engineering for NTCP modeling.