2649 - Improving Prediction of Surgical Resection after Chemoradiation in Borderline and Locally Advanced Pancreatic Cancer - a Novel Hierarchical Residual AI Framework

Purpose/Objective(s): Outcomes after chemoradiation for borderline resectable and locally advanced pancreatic ductal adenocarcinoma (PDAC)—including progression to surgical resection—are primarily modeled using clinical variables such as vascular involvement, stage, CA19-9, and performance status. However, this paradigm lacks incorporation of underlying tumor biology due to sparsity and subsequent lack of utility of somatic mutation features. To address this gap, we developed a hierarchical residual artificial intelligence (AI) framework designed to use patient genomic features to adjust outcome prediction beyond standard clinical risk. In addition to improved accuracy, this novel approach should also preserve biological interpretability through stable pathway-level representations.

Materials/Methods: Public PDAC clinical and somatic mutation data (n=580) were curated from the Genomic Data Commons. Patients had borderline resectable or locally advanced disease treated with neoadjuvant chemotherapy and radiation therapy and were classified by ultimate resection status. A clinical ensemble model generated leakage-safe risk predictions (p_clin). Genomic modeling used gene-level mutation features and single-sample pathway alteration scores from curated public gene-set libraries. We then trained genomic models orthogonal to clinical features using a residual learning algorithm to explain variance not captured by p_clin. Integration strategies were: simple feature concatenation, residual stacking, a meta-learner combining clinical and pathway risk predictions, and a multi-agent fusion framework that adaptively weighted clinical and genomic models. Performance was evaluated on a fixed held-out test set (n=174). Primary endpoints were ROC-AUC and PR-AUC.

Results: In the held-out cohort, the clinical-only model achieved ROC-AUC 0.630 and PR-AUC 0.486. Mutation-only modeling was inferior (PR-AUC 0.432), and simple stacking yielded minimal improvement (PR-AUC 0.489). Hierarchical residual fusion improved discrimination: a meta-learner achieved ROC-AUC 0.711 and PR-AUC 0.570, and a residual pathway plus clinical ensemble achieved ROC-AUC 0.713 and PR-AUC 0.562. Gains were concentrated in the highest clinical-risk tertile (n=58), where multi-agent genomic fusion increased PR-AUC from 0.533 to 0.70. Residual feature attribution prioritized cell-cycle/E2F-associated biology, with immune-related pathway representations (‘Allograft Rejection’).

Conclusion: When constraining somatic mutation features within our hierarchical residual AI framework, pathway-level representations meaningfully improve prediction—particularly among clinically high-risk PDAC patients. Our biology-aware residual modeling approach identified immune and cell cycle dynamics as orthogonal predictive features, and provided a new paradigm integrating sparse genomic data into clinically dominant risk models in radiation oncology.