Artificial Intelligence in Endohepatology: A Roadmap Toward an Intelligent One-Stop Shop for Liver-Directed Endoscopy

The review organizes the intersection of AI and endohepatology into four functional pillars: intelligent hemodynamic assessment, virtual histology, precision tissue acquisition, and integrated risk stratification with therapeutic and decision support. Previously, AI progress in endoscopy and hepatology was scattered across luminal endoscopy, hepatology imaging, digital pathology, and outcome prediction and had never been synthesized into a coherent domain; this work integrates them around the liver-directed endoscopic workflow. A framework-level synthesis within a narrative review, based on organizing and generalizing evidence from adjacent fields rather than new primary experimental data.

Because direct EUS-specific AI evidence in the liver remains limited, the review draws on adjacent proof-of-concept work—AI-assisted EUS in nonhepatic indications, transabdominal AI elastography, and AI histopathology—to sketch possible near-term integration. It separates transferable evidence from still-extrapolated applications, letting readers see which capabilities have adjacent supporting evidence and which remain conceptual. Evidence comes from proof-of-concept work in adjacent indications and modalities, i.e., indirect extrapolation, and the text explicitly notes that direct EUS liver evidence is limited.

It proposes a three-tier readiness framework, summarized in a domain table and a clinical-pathway figure, to distinguish demonstrated capabilities from extrapolated and conceptual applications. It offers an operational maturity scale for a very early-stage field, making it possible to view applications at different levels of maturity in a layered way. The framework is a conceptual tool proposed by the authors, grounded in their inductive judgment of the maturity of existing evidence.

The review concludes that although the field is still very early-stage, AI has strong potential to transform endohepatology into a single, machine-driven platform for diagnostic and therapeutic success, provided standardized datasets, prospective validation, and clear regulatory and governance standards are in place; the most immediate value will be in applications with transferable evidence, with hemodynamic and risk-prediction applications following as device-level data accrue. It gives a temporal ordering of where value lands first, prioritizing applications with transferable evidence rather than broadly claiming across-the-board breakthroughs. A forward-looking judgment and roadmap-style conclusion that depends on conditional prerequisites (standardized data, prospective validation, regulatory governance) rather than validated clinical results.

This work is positioned as a forward-looking roadmap, intended for researchers focused on integrating the liver-directed endoscopic workflow, endoscopy and hepatology clinical teams, and those engaged in regulatory and governance discussions. It aims to show that, given standardized datasets, prospective validation, and clear regulatory and governance standards, AI may move endohepatology toward a single machine-driven platform; its most immediate value is expected in applications with transferable evidence, while hemodynamic and risk-prediction applications are placed in a phase that follows as device-level data accrue.

Readers should still watch that the integration paths described rest largely on adjacent evidence—AI-assisted EUS in nonhepatic indications, transabdominal AI elastography, and AI histopathology—and how well that evidence transfers to the liver EUS workflow remains to be tested; which applications in the three-tier readiness framework will hold up under prospective validation, and what form standardized datasets and regulatory and governance standards will take, are open questions. In addition, this reading is a fast parse at summary scope and does not include the details of the domain table and clinical-pathway figure in the full text, so the specific tiering information those table and figure carry cannot be expanded here, which is a limitation on what can be summarized.