Frontiers in Oncology · Nov 19, 2026
Machine Learning-Assisted Integration of Liquid Biopsy Biomarkers for Early Detection of Pancreatic Cancer: Opportunities and Challenges
By Ruichao Li, Jiaqi Chen
Pancreatic ductal adenocarcinoma is usually found too late for resection, and carbohydrate antigen 19-9, the established serum marker, rises steeply only in the final two years before diagnosis, so machine learning applied to multianalyte liquid biopsy panels (circulating tumour DNA, extracellular vesicle cargo, microRNA, proteins and metabolites) has become a principal route pursued towards earlier detection. Reported discrimination is high, yet no such test is in routine use. A single-database search identified 60 eligible studies developing or validating such models, of which 18 reported performance in a cohort external to the training data and of different provenance; six further studies selected as the field's closest approximations to such a comparison were examined separately, and report discrimination between 0.89 and 0.97 under non-comparable validation regimesunder validation designs that cannot be directly compared. The specific claim on which integration rests isThe key assumption underlying multianalyte integration is that a combination detects disease its best single component would miss. It is rarely tested in a form that would establish it. Few of the studies examined construct such a comparator, and quoted improvements are frequently measured against something other than the best available alternative. Calibration was reported in three of sixteen assessable studies, leaving model outputs usable as ranks but not as risks, and nominally independent validation cohorts sometimes overlap, so external evidence is scarcer than counts suggestnominally independent validation cohorts sometimes overlap (among the six, three externally validated studies rest on two cohorts), so external evidence is scarcer than counts suggest. Tracing model inputs back to their primary literature showed that many selected features lack independent biological support at the resolution the models measure them. Multianalyte integration is therefore untested rather than refuted, and the principal remedies require analysis rather than new specimens. The most realistic near-term role for these models is triage within a defined high-risk population, deciding who enters imaging surveillance, rather than standalone population screening.