Introduction: Liver allocation changes have led to increased travel and expenditures, highlighting the need to efficiently identify marginal livers suitable for transplant. We evaluated the validity of existing non-invasive liver quality tests and a novel machine learning-based model at predicting deceased donor macrosteatosis >30%.
Methods: We compared previously-validated non-invasive tests and a novel machine learning-based model to biopsies in predicting macrosteatosis >30%. We also tested them in populations enriched for macrosteatosis.
Results: The Hepatic Steatosis Index area-under-the-curve (AUC) was 0.56. At the threshold identified by Youden's J statistic, sensitivity, specificity, positive, and negative predictive values were 49.6%, 58.9%, 14.0%, and 89.7%. Other tests demonstrated comparable results. Machine learning produced the highest AUC (0.71). Even in populations enriched for macrosteatosis, no test was sufficiently predictive.
Conclusion: Commonly used clinical scoring systems and a novel machine learning-based model were not clinically useful, highlighting the importance of pre-procurement biopsies to facilitate allocation.
Keywords: Clinical outcomes research; Decision-making; Liver transplantation; Observational database outcomes assessment; Transplant surgery.
Copyright © 2023. Published by Elsevier Inc.