Clinical and gene-expression models are both used to predict breast cancer outcomes, but they are often combined using simple linear rules that do not account for how their risk scores relate. Using the METABRIC breast cancer cohort, we examined whether directly modeling the joint relationship between clinical and gene-expression machine-learning risk scores could improve risk stratification for 5-year cancer-specific mortality. We created a binary 5-year cancer-death outcome and defined two predictor views: a clinical view based on demographic, tumor, and treatment-related variables, and a gene-expression view based on mRNA expression features. We trained several supervised classifiers and used 5-fold cross-validated predicted probabilities as out-of-fold risk scores. These scores were converted to pseudo-observations on \((0,1)^2\) and used to fit Gaussian, Clayton, Gumbel, and Frank copulas. In the primary METABRIC analysis, the clinical model showed stronger discrimination (AUC 0.783) than the gene-expression model (AUC 0.721). Among the copula families considered, the Frank copula yielded the smallest goodness-of-fit statistic, while the Gaussian copula showed nearly identical performance. Although the copula-fused score did not improve ROC-AUC relative to the clinical model, joint risk stratification based on the clinical and gene-expression scores identified clear differences in long-term survival, with patients in the high-both group experiencing the least favorable outcomes. Competing-risks analysis showed the same overall pattern for the cumulative incidence of cancer death. In an independent TCGA cohort, we conducted an external evaluation under a reduced harmonized specification using shared predictors and a common 5-year overall-mortality endpoint. The copula-fused score had discrimination comparable to the individual and simple-fusion scores, with substantially overlapping confidence intervals. All three scores underwent the same METABRIC-based recalibration procedure before calibration was evaluated using the Brier score, calibration slope, calibration intercept, integrated calibration index, Hosmer-Lemeshow diagnostic, and reliability curves. No gene met the prespecified stability criterion under repeated cross-validated permutation importance; gene-level findings were therefore treated as exploratory biological context rather than evidence of stable predictive drivers. We frame this work as a methodological and exploratory study of interpretable, dependence-aware score fusion: The results show that copula modeling provides an explicit description of dependence between clinical and gene-expression risk scores and supports descriptive joint score-group analyses. The study does not establish superior prediction, validated clinical risk categories, or clinical utility.
