Kidney Week

Abstract: TH-PO747

Using Artificial Intelligence to Improve Tacrolimus Dosing in Kidney Transplant Patients

Session Information

• Transplantation: Clinical - 1
  October 24, 2024 | Location: Exhibit Hall, Convention Center
  Abstract Time: 10:00 AM - 12:00 PM

Category: Transplantation

• 2102 Transplantation: Clinical

Authors

• Perez, Sean A., University of California San Diego Department of Surgery, La Jolla, California, United States

Sean A. Perez, MD

• Huo, Mingjia, University of California San Diego Department of Electrical and Computer Engineering, La Jolla, California, United States

Mingjia Huo

• Awdishu, Linda, University of California San Diego, Skaggs School of Pharmacy and Pharmaceutical Sciences, San Diego, California, United States

Linda Awdishu, PharmD, FASN

• Pour, Hayden H., University of California San Diego Department of Medicine, La Jolla, California, United States

Hayden H. Pour, MS

• Kerr, Janice, University of California San Diego, Skaggs School of Pharmacy and Pharmaceutical Sciences, San Diego, California, United States

Janice Kerr, PharmD

• Xie, Pengtao, University of California San Diego Department of Electrical and Computer Engineering, La Jolla, California, United States

Pengtao Xie

• Khan, Adnan A., University of California San Diego Department of Medicine, La Jolla, California, United States

Adnan A. Khan, MD, FASN

• Mekeel, Kristin, University of California San Diego Department of Surgery, La Jolla, California, United States

Kristin Mekeel, MD

• Nemati, Shamim, University of California San Diego Department of Medicine, La Jolla, California, United States

Shamim Nemati, PhD

Background

Tacrolimus (FK) is a mainstay of post-transplant immunosuppression with a narrow therapeutic window. Achieving therapeutic FK trough concentrations in the immediate post-operative period is challenging due to changing perioperative physiology and pharmacokinetics resulting in variable day to day exposures. We developed a machine learning (ML) model to predict the next day FK concentration and guide dosing to prevent persistent over- or under-dosing.

Methods

Retrospective data was extracted for adult kidney and/or liver transplant recipients from January 2016 to December 2023. Patient demographics, vital signs, laboratory values, diet, and medications were utilized to build the model. Data was randomly split 70%/15%/15% for training, validation, and test sets. XGBoost, traditional Recurrent Neural Network, and Long Short-Term Memory models were evaluated for best fit given the complexity of relationships and time dependency.

Results

1126 (706 kidney, 420 liver) transplant recipients were included in the study with a median (IQR) age of 58 (47-65) years, 51% white race, with FK daily dose of 4.0 (1.5 – 8.0) mg and concentration of 8.38 (5.9-10.3) ng/mL. The XGBoost model achieved a Mean Absolute Error (MAE) of 1.916 when predicting the next day’s FK concentrations. Figure 1 depicts the model's ability to predict underdosage or overdosage—using a target concentration of 10-13 ng/mL, via 3-class classification task. Overall, the model achieved accuracy of 0.73 with average F1-score of 0.58, F1-score 0.85 for predicting underdosing, and area under the curve of 0.67.

Conclusion

Machine learning accurately predicts next day FK concentrations which may improve precision dosing. Future directions include clinical implementation and evaluation and development of decision support for dosing recommendations.

Figure 1. XGBoost 3-class classification task.