Biologically Guided Hierarchical Fusion Framework for Early Pancreatic Cancer Detection Using Urinary Proteomic Biomarkers

Abstract

Early detection of pancreatic cancer is still a major clinical challenge as the disease progresses silently without any apparent symptoms and the diagnostic tools available today are often not sensitive enough for early diagnosis. Urinary proteomic biomarkers have emerged as a promising non-invasive alternative, with enhanced specificity through the application of appropriate diagnostic cut-offs. But these biomarkers have complex, non-linear relationships, making it difficult for traditional computational models to identify clinically relevant patterns. Most current computational approaches follow a single learning paradigm, either deep feature representation or threshold-based decision models, but seldom combine the merits of both in a biologically informed and clinically interpretable framework. To overcome the limitation, we propose a Biologically Guided Hierarchical Fusion Framework (HBFF) for the early detection of pancreatic cancer based on urinary proteomic biomarkers. The proposed framework consists of three complementary stages: (i) an attention module based on a transformer architecture to derive context-aware biomarker representations, (ii) a tree-ensemble module to model nonlinear biological threshold effects, and (iii) a ridge-regularized fusion layer for generating calibrated clinical risk estimates.The predictive performance and clinical utility of the proposed framework were evaluated via stratified cross-validation with calibration analysis, learning curve analysis, SHAP (SHapley Additive exPlanations)-based interpretation, and decision curve analysis. HBFF achieved a cross-validated ROC-AUC of 0.9756, significantly outperforming the TabTransformer and Tabular ResNet models. The dataset consisted of just over one thousand samples with 432 features per patient. The learning curve analysis did not show evidence that models were learning data-starved. SHAP analysis consistently identified biologically meaningful biomarkers that contributed to the hybrid framework predictions. Decision curve analysis also showed that the combination of the complementary information of the individual models was associated with a greater clinical benefit than either of the individual models or traditional decision-making strategies especially at higher threshold probabilities. These findings indicate that the proposed HBFF provides a reliable, interpretable, and less invasive decision-support framework for identifying individuals at higher risk of developing early-stage pancreatic cancer.