Artificial intelligence for variant-provenance adjudication in gastrointestinal cancer liquid biopsy: Distinguishing tumor, germline, and clonal hematopoietic signals

Liquid biopsy can inform treatment and monitoring in gastrointestinal (GI) cancer, but a plasma alteration may originate from tumor, constitutional variation, clonal hematopoiesis, or assay error.This focused narrative review examines artificial intelligence for adjudicating individual plasma variant calls.It separates analytical detection from biological source assignment and distinguishes direct per-call provenance, proxy, feature-development, and adjacent detection-pipeline evidence.Models integrate variant allele fraction, sequence context, fragmentomics, methylation, copy number, matched comparators, serial kinetics, and clinical context.Comparatorvalidated models demonstrate feasibility, but source classes, cohorts, and performance measures differ.GI calibration, mixed-source handling, and prospective clinical utility remain limited.Figures 1 and 2, a five-stage workflow, and hypothetical examples are authorproposed, unvalidated conceptual tools, not clinical guidelines.They retain uncertainty, mixed-source support, and an indeterminate state, with confirmation determined by the intended decision.No prospective interventional GI study showing improved patient outcomes from computational provenance assignment was identified in the reviewed literature.Artificial intelligence may prioritize ambiguous calls for confirmation, but current evidence does not support replacing direct comparator testing for consequential treatment, molecular residual disease, or hereditary-cancer decisions.