Machine learning-driven prognostication in liver transplantation: A tacrolimus intrapatient variability enhanced predictive model

Background: Accurately predicting long-term survival after liver transplantation (LT) remains a major clinical challenge. Tacrolimus intrapatient variability (Tac-IPV) has emerged as a potential prognostic marker, yet its integration into clinical decision-making remains limited. Methods: This retrospective study analyzed 381 adult LT recipients from two centers: 288 from the First Affiliated Hospital of Xi’an Jiaotong University (2015–2019) and 93 from the First Hospital of Shanxi Medical University (2020–2023). We developed a novel composite index, the Liver Transplantation Prognosis Predictor (LTPP), which integrates Tac-IPV (coefficient of variation), total bilirubin, and donor age using a Euclidean norm-based fusion method. The prognostic value of LTPP was evaluated through survival analysis and compared with conventional scores (model for end-stage liver disease and Child-Pugh scores). In addition, a random forest model was developed and externally validated to predict post-transplant survival. Results: In three-fold cross-validation, the random forest model demonstrated robust predictive performance for 1-, 2-, and 3-year survival (area under the receiver operating characteristic curve: 0.76, 0.73, and 0.68 in internal validation; 0.71, 0.74, and 0.70 in external validation). Survival analysis showed that LTPP significantly outperformed model for end-stage liver disease and Child-Pugh scores in risk stratification (log-rank P < 0.0001), establishing it as a reliable prognostic indicator. Conclusions: This study introduces LTPP, a clinically accessible composite index incorporating Tac-IPV. Compared to existing scoring systems, LTPP provides superior prognostic accuracy for long-term survival following LT, offering a promising tool for individualized post-transplant management.