Towards Personalized Tacrolimus Dosing Using an Algorithm-Driven Prediction Pipeline for Kidney Transplant

Abstract

Background: Tacrolimus (TAC) dosing presents a persistent challenge in postoperative care owing to its narrow therapeutic window and high inter‑patient variability, which often leads to suboptimal exposure with increasing risks of nephrotoxicity or graft rejection. Algorithm‑based personalized dosing strategies offer a promising approach to support clinical decision and improve long‑term outcomes. Methods: Unlike approaches relying on a wide range of variables and local clinical scopes, this study proposed a novel and versatile algorithm-driven strategy to predict TAC doses. A hybrid optimization method was first employed to identify a minimal set of key clinical factors. These factors were then used to construct a cascaded deep forest model capable of predicting both follow‑up and initial TAC doses in adult kidney transplant recipients. Results: When validated on 615 patients using leave-one-subject-out cross-validation, it achieved predictions within ±20% of actual values, with an accuracy of 89.8% for follow-up doses and 83.2% for initial doses. Independent external validation confirmed its robustness. A Shapley additive explanation analysis revealed significant correlations between input features and predictive doses. To support real-time clinical use, an open‑access web platform was provided (http://www.jcu-qiulab.com/tacp/). Conclusion: This approach offers a practical, effective, and algorithm-driven pipeline for automated drug dose analysis and prediction in clinical practice.