Abstract B090: Integrated radiomic-molecular model enhances early prediction of progression in advanced pancreatic cancer patients treated with gemcitabine/nab-paclitaxel with or without tocilizumab

Purpose: Given the poor prognosis and rapid deterioration risk in patients with advanced pancreatic cancer (PDAC), selecting the appropriate treatment strategy is crucial. To aid clinical decision-making, we developed a machine learning framework integrating molecular blood biomarkers and imaging features at baseline to identify patients at risk of early progression. Methods: Early progression was defined as clinical or radiological progression within 3 months of treatment initiation. Baseline CT scans, demographic, clinical, and molecular data from 147 patients in the PACTO trial (PMID: 40354592) were analyzed. Variables included age, sex, BMI, cancer stage, diabetes, smoking status, alcohol use, treatment regimen, PS, and date of progression. Blood-based data included circulating biomarkers, routine blood parameters and 92 inflammation-related proteins measured using the Olink immune-oncology panel. Wilcoxon tests and Cox models applied to baseline variables identified ∼40 candidate features for model development. CT images were preprocessed using standardized pipelines, including intensity normalization and ComBat harmonization to correct for inter-center variability. Radiomic features were extracted using both PyRadiomics and SCOPE-MIA. For multimodal analysis, radiomics features were combined with molecular and clinical features to construct a unified feature matrix, followed by recursive feature elimination, which yielded 10 predictive features (8 radiomics + 2 molecular) for progression-free survival at 3 months. Random Forest classifiers were trained and optimized for sensitivity, with performance assessed using cross-validation, confusion matrices, and feature importance. Results: The cohort included 11 locally advanced and 136 metastatic disease patients, treated with gemcitabine/nab-paclitaxel (n=71) or plus tocilizumab (n=76). Early progression occurred in 43 patients, who were more often female, under 70, and had poorer PS. They also had elevated YKL40, CRP, IL6, IL8, CA19-9, neutrophils, and LDH, and lower eosinophils, lymphocytes, hemoglobin, and albumin. Univariate analysis identified 26 baseline olink proteins associated with progression. The biomarker-only and radiomics-only models, each trained on 80% of the patients, achieved an AUC of 0.80 [95% CI: 0.63 – 0.943] and 0.78 [95% CI: 0.75 – 0.81] respectively correctly identifying 82% (9/11) and 55% (6/11) of early progressors. The multi-modal model (biomarkers + SCOPE-MIA radiomics embeddings) achieved the best performance with an AUC of 0.90 (95% CI: 0.87–0.93) correctly identifying 90% (10/11) of early progressors. The multi-modal model included 8 imaging features as well as lymphocytes and PFG (Plasma Fibrinogen). Conclusion: The combined radiomics-molecular biomarker model outperformed individual models and may help identify PDAC patients at risk of early progression. Pending external validation, this approach could guide future trial design and support more personalized care. Citation Format: Kevin Zi Ming Lim, Indu Khatri, Kubra Karagoz, Morteza Rezanejad, Iris Kolder, Lauren Brady, Han Si, Teng Jin Ong, Deepti Nagarkar, Gina Al-Farra, Susann Theile, Julia Sidenius. Johansen, Sidsel Cristy. Lindgaard, Brandon Higgs, Inna Markovna. Chen. Integrated radiomic-molecular model enhances early prediction of progression in advanced pancreatic cancer patients treated with gemcitabine/nab-paclitaxel with or without tocilizumab [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Advances in Pancreatic Cancer Research—Emerging Science Driving Transformative Solutions; Boston, MA; 2025 Sep 28-Oct 1; Boston, MA. Philadelphia (PA): AACR; Cancer Res 2025;85(18_Suppl_3):Abstract nr B090.