Transforming Liver Transplant Care with Artificial Intelligence: A Narrative Review
TL;DR
Artificial intelligence is increasingly applied across the liver transplant continuum and shows particular promise for HCC recurrence prediction and waitlist risk stratification, but the field is limited by a predominance of single-center exploratory studies with internal validation only, inadequate attention to algorithmic fairness and bias, and an absence of prospective evaluation or regulatory engagement.
Abstract
Outcomes in liver transplantation remain constrained by challenges in donor selection, organ allocation, and long-term graft survival, and traditional scoring systems such as MELD, SOFT, and the Donor Risk Index have well-recognized limitations in capturing the nonlinear complexity of transplant data. Artificial intelligence (AI), including machine learning (ML) and deep learning (DL), offers analytical methods capable of processing high-dimensional data beyond conventional regression, but the maturity and clinical readiness of this evidence base across the liver transplant continuum have not been comprehensively characterized. A narrative review was conducted via systematic search of PubMed, MEDLINE, and the Cochrane Library (January 1991-June 2025) using combinations of terms including “artificial intelligence,” “machine learning,” “deep learning,” “liver transplant,” “organ allocation,” and “graft failure.” Of 754 records identified, 45 studies met inclusion criteria following independent dual-reviewer screening. Each study was classified by evidence level as exploratory (internal validation only), externally validated, or implementation-oriented, and organized according to the pre-transplant and post-transplant continuum. AI applications span transplant candidacy assessment, organ allocation and waitlist prioritization, donor-recipient matching, graft quality assessment, and post-transplant prediction of acute kidney injury, sepsis, cardiovascular events, graft dysfunction, biliary complications, hepatocellular carcinoma (HCC) recurrence, immunosuppression dosing, and mortality. Performance varied widely (AUC/C-index 0.63-0.99), with the most methodologically mature evidence found for HCC recurrence prediction (C-index 0.75-0.839, including one internationally validated model) and waitlist mortality modeling. Notably, logistic regression outperformed several ML algorithms in the largest donor-recipient matching registry analyzed (n > 39,000), and cross-national validation of mortality models revealed substantial performance degradation (AUROC 0.70-0.74) when applied across healthcare systems. Fewer than half of included studies used explainability methods such as SHAP, and only two models explicitly incorporated equity considerations into their design. AI is increasingly applied across the liver transplant continuum and shows particular promise for HCC recurrence prediction and waitlist risk stratification. However, the field is limited by a predominance of single-center exploratory studies with internal validation only, inadequate attention to algorithmic fairness and bias, and an absence of prospective evaluation or regulatory engagement-no AI model in liver transplantation has yet received regulatory approval for clinical use. Future progress will require standardized data frameworks, multicenter prospective validation, formal reporting guidelines, and equity auditing before AI tools can be considered genuinely practice-changing.