A deep learning pathomic signature predicts survival and recurrence in pancreatic ductal adenocarcinoma independent of CA19-9.

The stratification of pancreatic ductal adenocarcinoma (PDAC) remains challenging due to the limitations of current staging systems and the ineffectiveness of the serum biomarker CA19-9 in a significant patient subgroup. We developed and validated a deep learning-based pathomic signature (PCPSDL) using routinely acquired haematoxylin and eosin (H&E)-stained whole-slide images from 252 patients at our institution and an external cohort of 183 patients from TCGA. PCPSDL robustly predicted overall and recurrence-free survival (C-indices: 0.64 and 0.61 internally, 0.62 and 0.61 externally). Crucially, it effectively stratified CA19-9-negative patients into high- and low-risk groups with significantly distinct outcomes (OS: HR = 9.26, P < 0.001; RFS: HR = 5.02, P < 0.001). Biological interpretation linked the high-risk score to an immunosuppressive microenvironment and a basal-like epithelial cell state. Our study establishes PCPSDL as a powerful, biologically interpretable, and clinically actionable prognostic biomarker that could guide personalized therapy for PDAC, particularly for patients lacking conventional serological markers.