Artificial intelligence in pancreatic cancer surgery: evidence maturity, boundaries, and clinical translation

Pancreatic cancer surgery requires decisions about resectability, treatment sequencing, operative planning, complications, and prognosis. We mapped artificial intelligence (AI) studies to five surgical decision domains and evaluated study design, validation, performance, decision utility, workflow fit, and evidence limitations. This structured narrative review included 31 direct AI studies and three adjacent or protocol studies; only direct AI studies underwent evidence-maturity grading. Eighteen studies met Level 1, ten Level 2, and three Level 3; none reached Levels 4 or 5. Three-dimensional assessment of vascular involvement has undergone external and prospective performance validation, but its effect on surgical decisions or patient outcomes remains unproven. Automated segmentation can aid anatomical visualization, although tumor delineation and workflow demands limit use. Transcriptomic analyses of trial samples suggest differential adjuvant treatment benefit but do not establish prospective biomarker-guided treatment allocation. Models for complications, recurrence, and survival offer risk-stratification signals, yet calibration, actionable thresholds, class imbalance, and clinical utility remain concerns. AI should support radiologists, surgeons, and multidisciplinary teams rather than replace their judgment. Clinical adoption requires robust external validation, prospective workflow assessment, and evidence that AI-assisted decisions improve patient outcomes.