Abstract:

Objective Maintaining stable concentrations of anti-rejection drugs represents a critical facetof post-kidney transplantation patient care; however，achieving personalized and precise management for each patient remains challenging. This study leverages an artificial intelligence-based deep learning framework to develop a machine learning predictive model for tacrolimus concentration，with the objective of recommending optimal dosing regimens for individual kidney transplant recipients. Methods Fifty kidney transplant recipients who underwentsurgery at the Urology Department of Beijing Friendship Hospital，Capital Medical University，between January 2024 and April 2025，were enrolled in this study. Drawing on prior experience in tacrolimus dosing，we collected data on patients' gender，age，weight，comorbidities，CYP3A5 metabolic phenotypes，initial tacrolimus（Tac）doses，and FK506 levels measured on postoperative days 7，9，11，13，and 15，with subsequent dosage adjustments made according to each concentration measurement. A LightGBM regression model was employed to predict and optimize tacrolimus dosing regimens. Results Among the 50 kidney transplant recipients enrolled in this study，none developed severe complications，including delayed recovery of graft function，postoperative infections，or bleedingThe dataset was partitioned into training and validation sets using a five-fold cross-validation approach. The final model demonstrated robust predictive performance in the test set，with a mean absolute error（MAE）of 0.166，root mean square error（RMSE）of 0.227，mean absolute percentage error（MAPE）of 7.035%，P20 of 0.935，P30 of 0.97，and a coefficient of determination（R-squared）of 0.932. Conclusion The LightGBM regression model exhibited excellent performance，providing a novel and effective strategy for personalized tacrolimus dosage adjustment in kidney transplant recipients. 

Key words:

Kidney transplantation , Blood drug concentration , Deep learning , Tacrolimus , Individualized pharmacotherapy

摘要：

目的 维持稳定的抗排斥反应药物浓度，是肾移植术后患者管理最重要的环节之一。但对于每一位患者实行个体化和精细化的管理实施困难。本项研究基于人工智能的深度学习系统，拟利用机器学习开发预测他克莫司浓度模型并建议肾移植受者的最佳剂量。 方法 纳入 2024 年 1 月至 2025 年 4 月在首都医科大学附属北京友谊医院泌尿外科接受肾移植手术的受者共 50 例作为本项研究对象。依据既往给患者首次他克莫司（tacrolimus，Tac）剂量的经验，收集患者性别、年龄、体重、基础疾病、CYP3A5 代谢水平、他克莫司初始剂量、术后第 7、9、11、13、15 天的 Tac 水平和依据每次浓度调整后的给药剂量。使用 LightGBM 回归模型预测并优化模型。 结果 纳入研究的 50 例肾移植受者均未发生移植肾功能延迟恢复、术后感染、出血等严重并发症。采用 five fold 将数据分为训练集和验证集。最终模型在测试集中的平均绝对误差为 0.166，均方根误差为 0.227，绝对百分比误差为 7.035，P20 为 0.935，P30 为 0.97，决定系数（R-squared）达 0.932。 结论 LightGBM 回归模型表现良好，为个体化调整肾脏移植受者他克莫司剂量提出了新的方案。

关键词: 肾移植 , 血药浓度 , 深度学习 , 他克莫司 , 个体化药物治疗