Background & aims: We aimed to develop and validate an artificial intelligence score (GEMA-AI) to predict liver transplant (LT) waiting list outcomes using the same input variables contained in existing models.
Methods: Cohort study including adult LT candidates enlisted in the United Kingdom (2010-2020) for model training and internal validation, and in Australia (1998-2020) for external validation. GEMA-AI combined international normalized ratio, bilirubin, sodium, and the Royal Free Glomerular Filtration Rate in an explainable Artificial Neural Network. GEMA-AI was compared with GEMA-Na, MELD 3.0, and MELD-Na for waiting list prioritization.
Results: The study included 9,320 patients: training cohort n=5,762, internal validation cohort n=1,920, and external validation cohort n=1,638. The prevalence of 90-days mortality or delisting for sickness ranged 5.3%-6% across different cohorts. GEMA-AI showed better discrimination than GEMA-Na, MELD-Na and MELD 3.0 in the internal and external validation cohorts, with a more pronounced benefit in women and in patients showing at least one extreme analytical value. Accounting for identical input variables, the transition from a linear to a non-linear score (from GEMA-Na to GEMA-AI) resulted in a differential prioritization of 6.4% of patients within the first 90 days and would potentially save one in 59 deaths overall, and one in 13 deaths among women. Results did not substantially change when ascites was not included in the models.
Conclusions: The use of explainable machine learning models may be preferred over conventional regression-based models for waiting list prioritization in LT. GEMA-AI made more accurate predictions of waiting list outcomes, particularly for the sickest patients.
Keywords: Artificial Neural Networks; Disparities; Gender; Liver Allocation; Machine Learning; eXplainable Artificial Intelligence.
Copyright © 2025. Published by Elsevier Inc.