AI-based Clinical Phenotyping for Pan-Cancer Stratification Using Unsupervised Learning

Objective: The clinical organization of cancer has traditionally been constrained by tumor-centric classifications that inadequately capture the shared phenotypic structure across malignancies. A comprehensive understanding of cancer heterogeneity requires a framework capable of reconstructing the continuous clinical landscape defined by disease burden, demographic factors, and stage at presentation without directly incorporating survival outcomes as model inputs. Methods: This study is a retrospective computational analysis utilizing the MLOmics harmonized benchmark dataset derived from The Cancer Genome Atlas (TCGA). A nonlinear artificial intelligence framework was implemented, integrating Uniform Manifold Approximation and Projection (UMAP) for manifold learning and Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN) for unsupervised phenotype discovery. Clinical variables, including age, gender, and pathological stage, were encoded and standardized prior to embedding. Results: The analysis revealed that pan-cancer clinical data are organized within a continuous nonlinear manifold rather than into discrete tumor-specific compartments. Density-based clustering identified multiple stable clinical phenotypes, alongside a subset of noise points representing atypical or heterogeneous cases. The observed clustering structure demonstrated moderate silhouette separation, consistent with the overlapping clinical realities of oncology. Importantly, phenotype formation was not driven by tumor origin but by shared clinical attributes, particularly disease stage and patient demographics. Conclusion: In radiation oncology, where treatment decisions are often driven by anatomical extent and disease burden, such phenotype-centric models enable cross-disease insights that support precision treatment planning and resource optimization.