Artificial intelligence for early detection of pancreatic cancer: pre-diagnostic detection across imaging, biomarkers, and EHRs: a systematic review

Abstract Pancreatic Ductal Adenocarcinoma (PDAC) is highly lethal, yet outcomes improve markedly when the disease is detected at its earliest stages. This systematic review synthesizes recent applications of Artificial Intelligence (AI) and Machine Learning (ML) for clinically actionable early detection across three data streams: medical imaging, Electronic Health Records (EHRs), and liquid-biopsy biomarkers. We surveyed studies published between 2020 and mid-2025, emphasizing task design, validation strategies, and the authors’ description of model reliability and deployment. Across modalities, retrospective evidence indicates that AI can surface weak, distributed signals, preceding overt radiologic findings, support challenging differential diagnoses, and aid procedure-level verification. EHR-based population pre-filtering (triage via routine data to select patients for imaging) shows promise, and biomarker-driven approaches typically outperform single-analyte baselines. Taken together, the current literature supports the possibility of a pragmatic, tiered workflow, in which EHR/biomarker triage directs confirmatory Computed Tomography (CT) and Magnetic Resonance Imaging (MRI) and where indicated, Endoscopic Ultrasound (EUS)-guided verification. Clinical implementation remains constrained by limited external validation, variability in imaging protocols and biomarker assays, lack of clinical-utility analysis (decision-curve net benefit), and sparse reporting of subgroup outcomes. Across modalities, studies frequently report good Area Under the Receiver Operating Characteristic Curve (AUROC)-based discrimination, high negative predictive value at conservative thresholds in EHR- and biomarker-based triage, and, in some studies, AI models detected PDAC on pre-diagnostic CT scans, providing a meaningful lead time before clinical diagnosis. These findings underscore the promise, but also the need for prospective, pathway-aware evaluation.