Abstract
Abstract
Liver transplant (LT) is a life-saving treatment for patients with cirrhosis and/or hepatocellular carcinoma (HCC), but organ shortages and suboptimal donor-recipient matching remain major challenges and existing donor-recipient matching risk models offer limited predictive accuracy. Our study aims to develop a machine-learning-based model to predict and rank long-term post-transplant survival. We included adult deceased donor liver transplant recipients from the Scientific Registry of Transplant Recipients (2009–2019). We developed a first of its kind XGBoost-based survival model to predict graft survival at important discrete timepoints and built a corresponding ranking score. The reliability of the score was compared to existing scores. We included 60,649 waitlist registrants, of which 41,058 (67.7%) were males. We found that our predictive model,
DisCScore
, was the strongest at predicting donor-recipient pair survival times, with a C-index of 0.858 (95%CI: 0.8525–0.8643). We achieved the highest mean time-dependent AUROC at 6-months, 1- and 3- years post-LT using
DisCScore
(mean AUC (95% CI): 0.797 (0.783–0.812), 0.821 (0.809–0.832) and 0.865 (0.857–0.874) respectively and
DisCScore
outperformed CoxPH and DeepSurv models in all three timepoints. We found that our model performed best in the high MELD group (MELD > 30) – where the C-index was the highest: 0.873 (95% CI 0.863–0.891), as compared to the low and medium MELD groups. Our proposed model can accurately predict donor-recipient pair survival probabilities, especially in those recipients with high MELD scores. This has the potential to be integrated into the complex decision-making pathway of organ allocation.