Re-Evaluating Donor-Recipient Risk Scores in the Machine Perfusion Era Using a Large Multicenter Cohort: Do Traditional Predictive Models Still Apply?

BACKGROUND: With increasing use of donation after circulatory death (DCD) donor grafts, the need for accurate risk-stratification tools in liver transplantation has become critical. Machine perfusion has been shown to improve both graft and patient survival. However, existing prediction models were developed in the static cold storage (SCS) era and their accuracy when using machine perfusion remains unknown. Our study aims to study if current scoring systems can predict graft and patient survival for DCD livers preserved using normothermic machine perfusion (NMP). METHODS: This was a retrospective study of patient and graft survival in a two-center U.S. DCD NMP cohort (n = 532) between January 2022 and December 2024. Predicted graft and patient survival rates were obtained from established risk scoring systems, including the United Kingdom DCD risk score (UK-DCD), Early Allograft Dysfunction (EAD) Olthoff score, Liver Graft Assessment Following Transplantation (L-GrAFT7) score, Donor Risk Index (DRI), Balance of Risk (BAR) score, and Survival Outcomes Following Liver Transplant (SOFT) score. Model discrimination was assessed using receiver operating characteristic (ROC) curves (C-statistics). Survival outcomes across risk strata were estimated and compared using Kaplan-Meier (KM) analysis. RESULTS: Across scoring systems, predicted survival consistently underestimated actual graft and patient survival in the NMP DCD cohort. The DRI, SOFT, EAD and L-GrAFT7 did not demonstrate predictive value. UK-DCD predicted ≤85% and <40% 1-year graft survival for its high-risk and "futile" groups, whereas observed rates were 93.7% and 86.8%. BAR low-risk predicted >60% 5-year survival but 94.2% was observed; the high-risk group predicted <60% and observed was 53.3%. Despite underestimations and generally poor C-statistics, in KM stratification, high-risk groups still showed statistical differences in the UK-DCD categories (p = 0.007), and in SOFT score groups (p < 0.001). No statistical difference was seen in BAR (p = 0.36), DRI (p = 0.36), EAD (p = 0.12), or L-GrAFT7 (p = 0.75). CONCLUSION: Existing liver transplant risk scores show limited predictive value for donor-related risk factors in the NMP era, whereas recipient-related factors remain somewhat predictive, despite overestimating risk. Early allograft dysfunction metrics were not reliable predictors of short or long-term outcomes. Future risk models should integrate machine-perfusion parameters and refined donor variables to reflect outcomes more accurately in contemporary DCD liver transplantation.