Tacrolimus variability and creatinine predict readmission after liver transplantation

Unplanned readmissions after liver transplantation occur in over 30% of recipients, yet no validated prediction models exist, and prior observational studies suffer from immortal time bias. The optimal readmission window for outcome prediction and the feasibility of early risk stratification remain undefined. This study is a retrospective analysis of 922 adult liver transplant recipients (August 2018-August 2025) at a single center. Time-varying Cox regression evaluated 14-, 30-, and 90-day readmission windows as predictors of 1-year mortality, correcting for immortal time bias. Gradient-boosted machine learning models leveraging 528,400 laboratory measurements (28 analytes) predicted 90-day readmission using either complete hospitalization data or data restricted to postoperative day 7. Feature importance was quantified by gain, and clinical utility was assessed through risk stratification. Among 902 hospital survivors, 342 (37.9%) experienced an unplanned readmission within 90 days of initial discharge. Only the 90-day readmission window predicted 1-year mortality in time-varying analysis (HR 1.73, 95% CI 1.17-2.57, p=0.006). The model for readmission using complete data achieved AUC 0.614 (95% CI 0.576-0.652); the postoperative day 7 restricted model achieved AUC 0.615 (95% CI 0.577-0.652), with no meaningful performance difference. The tacrolimus coefficient of variation x peak creatinine interaction was the dominant predictor in both the complete model (17.3% importance, rank 1) and the day 7 restricted model (20.4% importance, rank 2). This interaction stratified patients into high-risk (tacrolimus CV >0.3 and creatinine >2.0 mg/dL; 49.8% readmission) versus low-risk (24.8% readmission) groups (risk ratio 2.01, p<0.001). These results identify a modifiable biological determinant of readmission and establish a framework for targeted interventions to reduce unplanned readmission and improve post-transplant outcomes.