Machine learning and artificial intelligence in liquid biopsy-based early detection of pancreatic cancer: a scoping review.

Abstract

Pancreatic ductal adenocarcinoma (PDAC) presents as a cancer with an especially poor prognosis, largely due to the challenges surrounding its early diagnosis. Liquid biopsy has emerged as a promising, noninvasive method for screening across a variety of cancers. This approach is limited, however, by the extensive heterogeneity of biological samples, a challenge that teams are looking to address using artificial intelligence (AI) and machine learning (ML) strategies. By harnessing the ability of ML algorithms to extract the most salient features from complex datasets, researchers can identify biomarkers with high predictive value for PDAC. This review explores the current landscape of AI-powered liquid biopsy for early PDAC diagnosis, focusing on specific techniques and their respective degrees of success. Following PRISMA-ScR guidelines, 85 studies were extracted from PubMed and Scopus with a final 18 studies included. The majority of papers utilized blood (n = 15) as the source of liquid biopsy, with the remainder analyzing urine, bile, or cyst fluid. Random forests (n = 9) and support vector machines (n = 7) were the most frequently implemented ML models, while two papers focused on deep learning methods. Limitations include the lack of standardized reporting for model performance metrics and small cohort sizes with non-granular labels.